(De Gruyter Textbook) Karl Kraus, Ian A. Harley, Stephen Kyle - Photogrammetry - Geometry From Images and Lase PDF

de Gruyter Textbook
Karl Kraus
Photogrammetry
Karl Kraus
Photogrammetry
Geometry from Images
and Laser Scans
Second Edition
Translated by
Ian Harley
Stephen Kyle
w Walter de Gruyter
DE
G Berlin · New York

Author
ο. Univ.-Prof. Dipl.-Ing. Dr. techn. Karl Kraus
formerly
Institute of Photogrammetry and Remote Sensing
Vienna University of Technology
Vienna, Austria
Translators
Prof. Ian Harley
Dr. Stephen Kyle
University College London
London, Great Britain
This second English edition is a translation and revision of the seventh German
edition:
Kraus, Karl: Photogrammetrie, Band 1, Geometrische Informationen aus Photogra-
phien und Laserscanneraufnahmen. Walter de Gruyter, Berlin · New York, 2004
First English edition:

Kraus, Karl: Photogrammetry, Volume 1, Fundamentals and Standard Processes.
Dümmler, Köln, 2000.
© Printed on acid-free paper which falls within the guidelines

of the ANSI to ensure permanence and durability.
Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.
ISBN 978-3-11-019007-6
© Copyright 2007 by Walter de Gruyter GmbH & Co. KG, 10785 Berlin, Germany.
All rights reserved, including those of translation into foreign languages. No part of this book
may be reproduced in any form or by any means, electronic or mechanical, including photocopy,
recording, or any information storage and retrieval system, without permission in writing from
the publisher.
Printed in Germany.
Coverdesign: +malsy, kommunikation und gestaltung, Willich.
Printing and binding: Hubert & Co. GmbH & Co. KG, Göttingen.
Foreword to the second English edition
The first edition of Volume 1 of the series of textbooks "Photogrammetry" was pub-
lished in German in 1982. It filled a large void and the second and third editions were
printed soon afterwards, in 1985 and 1990. The fourth edition was published in English
in 1992, translated by Peter Stewardson. The following three editions were published
in German in the years 1995, 1997, and 2003, making seven editions in all. The English
edition was re-printed in 2000.
Volume 1 was additionally translated into several languages, including Serbocroatian,

by Prof. Joksics, Technical University Belgrade; Norwegian, by Prof. Oefsti, Univer-
sity of Trondheim; Greek, by Dr. Vozikis and Prof. Georgopoulos, National Technical
University of Athens; Japanese, by Prof. Oshima and Mr. Horie, Hosei University;
Italian, by Prof. Dequal, Politecnico Torino; French, by Prof. Grussenmeyer and O.
Reis, Ecole Nationale Superieure des Arts et Industries de Strasbourg; Hungarian by
Prof. Detreköi, Dr. Melykuti, S. Mihäli, and P. Winkler, TU Budapest; Ukrainian by
S. Kusyk, Lvivska Politechnika; and Turkish, by Prof. Altan, Technical University of
Istanbul.
This second English edition is a translation of the seventh, German, edition by Dr. Ian
Harley, Professor Emeritus, and Dr. Stephen Kyle, both of University College London.
They not only translated the text, but they also made valuable contributions to it; their
comments and suggestions led to a clearly improved edition. Compared to the first
English edition there are major changes. Analogue and analytical photogrammetry are
reduced significantly, most importance is given to digital photogrammetry, and, finally,
laser scanning is included. Terrestrial as well as airborne laser scanning have gained
great importance in photogrammetry. Photogrammetric methods are, with small adap-
tations, applicable to data acquired by laser scanning. Therefore, only minor additions
to photogrammetry were necessary to cover the chapter on laser scanning. Compared
to the previous German edition there are, especially, updates on digital cameras and
laser scanners.
The original German version arose out of practical research and teaching at the Vi-
enna University of Technology. Volume 1 first introduces the necessary basics from
mathematics and digital image processing. It continues with photogrammetric acquisi-
tion technology with special consideration of photo-electrical imaging (CCD cameras).
Particular attention is paid to the use of the Global Positioning System (GPS) and Iner-
tial Measurement Units (IMU) for flight missions. The discussion on photogrammetric
processing begins with orientation methods including those based on projective geom-
etry. The orientation methods which are discussed for two images are extended to
image blocks in the form of photogrammetric triangulation.
vi Foreword
In the discussion of stereo-plotting instruments most attention is given to digital soft-

copy stations. In addition to automatic processing methods, semiautomatic methods,
which are widely used in practice, are also explained. This textbook first treats dig-
ital orthophoto production, and then includes three-dimensional virtual worlds with
photographic texture.
This selection and arrangement of material offers students a straightforward introduc-
tion to complex photogrammetry as practised today and as it will be practised in the
near future. It also offers practising photogrammetrists the possibility of bringing them-
selves up-to-date with the modern approach to photogrammetry and saves them at least
a part of the tedious study of technical journals which are often difficult to understand.
For technically oriented neighbouring disciplines it provides a condensed description
of the fundamentals and standard processes of photogrammetry. It lays the basis for
that interdisciplinary collaboration which gains ever greater importance in photogram-
metry. Related, non-technical disciplines will also find valuable information on a wide
range of topics.
For the benefit of its readers, the textbook follows certain principles: didactics are put
before scientific detail; lengthy derivations of formulae are put aside; theory is split into
small sections alternating with practically-oriented passages; the theoretical basics are
made clear by means of examples; and exercises are provided with solutions in order
to allow self-checking.
This series of textbooks is a major contribution to photogrammetry. It is very sad that
Prof. Kraus, who died unexpectedly in April 2006, cannot see it published. At that time
the translation was already in progress. Final editing was performed by Dr. Josef Jansa
and Mr. Andreas Roncat from the Vienna Institute of Photogrammetry and Remote
Sensing. Thanks are also due to the many people at the Institute of Photogrammetry
and Remote Sensing who did major and minor work behind the scenes, such as drawing
and editing figures, calculating examples and exercises, making smaller contributions,
proofreading, composing the I4TgX text, etc. This book, however, is truly a book by
Prof. Kraus.
Karl Kraus was born in 1939 in Germany and became Professor of Photogrammetry
in Vienna in 1974. Within these 32 years of teaching, counting all translations and
editions, more than twenty textbooks on photogrammetry and remote sensing bearing
the name Karl Kraus were published. Many examples and drawings in this textbook
were supplied by the students and collaborators of Prof. Kraus in Vienna. With deep
gratitude the entire Institute of Photogrammetry and Remote Sensing looks back at the
time spent with Karl Kraus and forward to continuing the success story of this textbook.
Norbert Pfeifer Vienna, Summer 2007

Professor in Photogrammetry
Institute of Photogrammetry and Remote Sensing
Vienna University of Technology
Notes for readers
This textbook provides an introduction to the basics of photogrammetry and laser scan-
ning. References to Volume 2, Chapters B, C, D, and E, refer to
Kraus, Karl: Photogrammetry, Volume 2, Advanced Methods and Appli-

cations, with contributions by J. Jansa and H. Kager. 4 th edition, Diimmler,
Bonn, 1997, ISBN 3-427-78694-3.
Volume 2 is a completely separate textbook and is currently out of print. It covers

advanced topics for readers who require a deeper theoretical knowledge and details of
specialized applications.
Contents
Foreword ν
Notes for readers vii
1 Introduction 1
1.1 Definitions 1
1.2 Applications 2
1.3 Some remarks on historical development 3
2 Preparatory remarks on mathematics and digital image processing . . . 10

2.1 Preparatory mathematical remarks 10
2.1.1 Rotation in a plane, similarity and affine transformations . . . . 10
2.1.2 Rotation, affine and similarity transformations in three-
dimensional space 14
2.1.3 Central projection in three-dimensional space 21
2.1.4 Central projection and projective transformation of a plane . . . 24
2.1.5 Central projection and projective transformation of the straight
line 29
2.1.6 Processing a stereopair in the "normal case" 31
2.1.7 Error theory for the "normal case" 33
2.2 Preliminary remarks on the digital processing of images 35
2.2.1 The digital image 36
2.2.2 A digital metric picture 38
2.2.3 Digital processing in the "normal case" and digital projective
rectification 40
3 Photogrammetric recording systems and their application 47

3.1 The basics of metric cameras 47
3.1.1 The interior orientation of a metric camera 47
3.1.2 Calibration of metric cameras 55
3.1.3 Correction of distortion 56
3.1.4 Depth of field and circle of confusion 58
3.1.5 Resolving power and contrast transfer 63
3.1.5.1 Diffraction blurring 63
3.1.5.2 Optical resolving power 64
3.1.5.3 Definition of contrast 68
3.1.5.4 Contrast transfer function 68
χ Contents
3.1.6 Light fall-off from centre to edge of image 70

3.2 Photochemical image recording 71
3.2.1 Analogue metric image 71
3.2.1.1 Glass versus film as emulsion carrier 72
3.2.1.2 Correcting film deformation 73
3.2.2 Physical and photochemical aspects 77
3.2.2.1 Colours and filters 77
3.2.2.2 The photochemical process of black-and-white
photography 79
3.2.2.3 Gradation 81
3.2.2.4 Film sensitivity (speed) 82
3.2.2.5 The colour photographic process 84
3.2.2.6 Spectral sensitivity 87
3.2.2.7 Resolution of photographic emulsions 89
3.2.2.8 Copying with contrast control 91
3.2.3 Films for aerial photography 91
3.3 Photoelectronic image recording 93
3.3.1 Principle of opto-electronic sensors 93
3.3.2 Resolution and modulation transfer 97
3.3.3 Detector spacing (sampling theory) 100
3.3.4 Geometric aspects of CCD cameras 102
3.3.5 Radiometric aspects of CCD cameras 103
3.3.5.1 Linearity and spectral sensitivity 103
3.3.5.2 Colour imaging 104
3.3.5.3 Signal-to-noise ratio 105
3.4 Digitizing analogue images 106
3.4.1 Sampling interval 107
3.4.2 Grey values and colour values 107
3.4.3 Technical solutions 109
3.5 Digital image enhancement 110
3.5.1 Contrast and brightness enhancement Ill
3.5.1.1 Histogram equalization 114
3.5.1.2 Histogram normalization 114
3.5.1.3 Compensation for light fall-off from centre to edge
of image 119
3.5.1.4 Histogram normalization with additional contrast
enhancement 120
3.5.2 Filtering 122
3.5.2.1 Filtering in the spatial domain 122
3.5.2.2 Filtering in the frequency domain 125
3.6 Image pyramids/data compression 128
3.6.1 Image pyramids 128
Contents xi
3.6.2 Image compression 129

3.7 Aerial cameras and their use in practice 131
3.7.1 Flight planning 131
3.7.2 Metric aerial cameras 137
3.7.2.1 Large format, metric film cameras 137
3.7.2.2 Digital cameras with CCD area sensors 144
3.7.2.3 Digital 3-line cameras 146
3.7.3 Satellite positioning and inertial systems 147
3.7.3.1 Use of GPS during photogrammetric flying missions
and image exposure 147
3.7.3.2 Accurate determination of exterior orientation
elements by GPS and IMU 148
3.7.3.3 Gyro-stabilized platforms and particular features of
line cameras and laser scanners 153
3.7.4 Image motion and its compensation 155
3.7.4.1 Compensation of image motion in aerial film cameras 157
3.7.4.2 Image motion compensation for digital cameras with
CCD area arrays 158
3.7.4.3 Image motion compensation for digital line cameras . 159
3.7.5 Effective illumination in aerial photography 159
3.7.6 Survey aircraft 161
3.8 Terrestrial metric cameras and their application 163
3.8.1 "Normal case" of terrestrial photogrammetry 164
3.8.2 Stereometric cameras 165
3.8.3 Independent metric cameras 166
3.8.4 Semi-metric cameras 169
3.8.5 Amateur cameras 170
3.8.6 Terminology and classification 170
3.8.7 CCD cameras 172
3.8.8 Planning and execution of terrestrial photogrammetry 173
4 Orientation procedures and some methods of stereoprocessing 180

4.1 With known exterior orientation 181
4.1.1 Two overlapping metric photographs 181
4.1.2 Metric images with a three-line sensor camera 183
4.2 With unknown exterior orientation 184
4.2.1 Separate orientation of the two images 185
4.2.2 Combined, single-stage orientation of the two images 188
4.2.3 Two-step combined orientation of a pair of images 189
4.3 Relative orientation 193
4.3.1 Relative orientation of near-vertical photographs 193
4.3.2 Relative orientation and model formation using highly tilted
photographs 197
xii Contents
4.3.2.1 Gauss-Helmert model of relative orientation 200

4.3.2.2 A combined, single-stage relative orientation 201
4.3.3 Alternative formulation of relative orientation 201
4.3.4 Relative orientation of near-vertical photographs by
y-parallaxes 205
4.3.4.1 Mountainous country (after Jerie) 206
4.3.4.2 Flat ground (after Hallert) 209
4.3.5 Critical surfaces in relative orientation 210
4.3.6 Error theory of relative orientation 213
4.3.6.1 Standard deviations of the elements of orientation . . 2 1 3
4.3.6.2 Deformation of the photogrammetric model 215
4.4 Absolute orientation 219
4.4.1 Least squares estimation 219
4.4.2 Error theory of absolute orientation 226
4.4.3 Determination of approximate values 228
4.5 Image coordinate refinement 230
4.5.1 Refraction correction for near-vertical photographs 230
4.5.2 Correction for refraction and Earth curvature in horizontal
photographs 233
4.5.3 Earth curvature correction for near-vertical photographs . . . . 2 3 5
4.5.4 Virtual (digital) correction image 237
4.6 Accuracy of point determination in a stereopair 238
5 Photogrammetric triangulation 246

5.1 Preliminary remarks on aerotriangulation 246
5.2 Block adjustment by independent models 248
5.2.1 Planimetrie adjustment of a block 248
5.2.2 Spatial block adjustment 256
5.2.3 Planimetrie and height accuracy in block adjustment by
independent models 259
5.2.3.1 Planimetrie accuracy 259
5.2.3.2 Height accuracy . . . 265
5.2.3.3 Empirical planimetric and height accuracy 267
5.2.3.4 Planimetric and height accuracy of strip triangulation 267
5.3 Bundle block adjustment 269
5.3.1 Basic principle 269
5.3.2 Observation and normal equations for a block of photographs . 270
5.3.3 Solution of the normal equations 273
5.3.4 Unknowns of interior orientation and additional parameters . . 274
5.3.5 Accuracy, advantages and disadvantages of bundle block
adjustment 274
5.4 GPS- and IMU-assisted aerotriangulation 276
5.5 Georeferencing of measurements made with a 3-line camera 277
Contents xiii
5.6 Accounting for Earth curvature and distortions due to cartographic

projections 280
5.7 Triangulation in close range photogrammetry 282
6 Plotting instruments and stereoprocessing procedures 286

6.1 Stereoscopic observation systems 286
6.1.1 Natural spatial vision 286
6.1.2 The observation of analogue and digital stereoscopic images . .288
6.2 The principles of stereoscopic matching and measurement 295
6.3 Analogue stereoplotters 297
6.4 Analytical stereoplotters 300
6.4.1 Stereocomparators 300
6.4.2 Electronic registration of image coordinates in the
monocomparator 302
6.4.3 Universal analytical stereoplotter 304
6.5 Digital stereoplotting equipment 306
6.6 Computer-supported manual methods of analysis 307
6.6.1 Recording in plan 308
6.6.2 Determination of heights 310
6.6.3 Recording of buildings 312
6.6.4 Transition to spatially related information systems 316
6.7 Operator accuracy with a computer assisted system 317
6.7.1 Measurement in plan 317
6.7.1.1 Point measurement 317
6.7.1.2 Processing of lines 318
6.7.2 Height determination 319
6.7.2.1 Directly drawn contours 319
6.7.2.2 Relationship between contour interval and heighting
accuracy 320
6.7.2.3 Contours obtained indirectly from a DTM 321
6.7.2.4 Measurement of buildings 322
6.7.3 Checking of the results 323
6.8 Automatic and semi-automatic processing methods 323
6.8.1 Correlation, or image matching, algorithms 323
6.8.1.1 Correlation coefficient as a measure of similarity . . .324
6.8.1.2 Correlation in the subpixel region 326
6.8.1.3 Interest operators 330
6.8.1.4 Feature based matching 331
6.8.1.5 Simultaneous correlation of more than two images . . 332
6.8.2 Automated interior orientation 334
6.8.3 Automated relative orientation and automated determination of
tie points 335
xiv Contents
6.8.3.1 Near-vertical photographs with 60% forward over-

lap taken over land with small height differences . . . 336
6.8.3.2 Near-vertical photographs with 60% forward over-
lap taken over land with large height differences . . . 337
6.8.3.3 Arbitrary configurations of photographs and objects
with very complex forms 337
6.8.3.4 Line-based (edge-based) relative orientation 338
6.8.3.5 Tie points for automated aerotriangulation 339
6.8.4 Automated location of control points 339
6.8.5 Inclusion of epipolar geometry in the correlation 341
6.8.5.1 Epipolar geometry after relative orientation using
rotations only 342
6.8.5.2 Epipolar geometry in normalized images 343
6.8.5.3 Epipolar geometry in original, tilted metric
photographs 345
6.8.5.4 Derivation of normalized images using the elements
of exterior orientation 346
6.8.5.5 Epipolar geometry in images which have been
oriented relatively using projective geometry 347
6.8.5.6 Epipolar geometry in three images 349
6.8.6 Automated recording of surfaces 350
6.8.7 Semi-automated processing for plan 352
6.8.7.1 Active contours (snakes) 353
6.8.7.2 Sequential processing 355
6.8.8 Semi-automatic measurement of buildings 360
6.8.9 Accuracy and reliability of results obtained by automated or
semi-automated means 363
6.8.10 Special features of the three-line camera 364
7 Orthophotos and single image analysis 366

7.1 Perspective distortion in a metric image 367
7.2 Orthophotos of plane objects 373
7.2.1 With vertical camera axis 373
7.2.2 With tilted camera axis 376
7.2.3 Combined projective and affine rectification 378
7.3 Orthophotos of curved objects 380
7.3.1 Production principle 380
7.3.2 Orthophoto accuracy 384
7.4 Analogue, analytical and digital single image analysis 393
7.4.1 Analogue, analytical and digital orthophoto analysis 393
7.4.2 Analytical and digital analysis of a tilted image of a flat object . 393
7.4.3 Analytical and digital single image analysis of curved object
surfaces 394
Contents XV
7.5 Photo models 396

7.6 Static and dynamic visualizations 399
8 Laser scanning 400

8.1 Airborne laser scanning 400
8.1.1 Principle of operation 400
8.1.2 Analysis and processing 404
8.1.2.1 Georeferencing 404
8.1.2.2 Derivation of terrain models 407
8.1.2.3 Generation of building models 411
8.1.3 Comparison of two paradigms and further performance
parameters of laser scanners 413
8.2 Terrestrial laser scanning 419
8.2.1 Principle of operation 419
8.2.2 Georeferencing 420
8.2.3 Connecting point clouds 422
8.2.4 Strategies for object modelling 423
8.2.5 Integration of laser data and photographic data 426
8.3 Short range laser scanning 428
Appendices 432
2.1-1 Three-dimensional rotation matrix 432
2.1-2 Mathematical relationship between image and object
coordinates (collinearity condition) 436
2.1-3 Differential coefficients of the collinearity equations 438
2.2-1 Derivation of Formula (2.2-5) using homogeneous coordinates . 440
4.1-1 Estimation by the method of least squares 441
4.2-1 Direct Linear Transformation (DLT) with homogeneous
coordinates 444
4.3-1 Differential coefficients for the coplanarity equations 445
4.6-1 The empirical determination of standard deviations and
tolerances 447
Completion of the references 449
Index 451
Chapter 1
Introduction
1.1 Definitions
Photogrammetry allows one to reconstruct the position, orientation, shape and size of
objects from pictures; these pictures may originate as photochemical images (conven-
tional photography) or as photoelectric images (digital photography). Laser scanner
images, a third group, have arrived in recent years; laser scanner images have distance
information associated with every picture element. The results of a photogrammetric
analysis may be:
• numbers—coordinates of separate points in a three-dimensional coordinate sys-

tem (digital point determination),
• drawings (analogue)—maps and plans with planimetric detail and contour lines
together with other graphical representation of objects,
• geometric models (digital)—which are fed in to information systems,
• images (analogue and/or digital)—above all, rectified photographs (orthopho-
tos) and, derived from these, photomaps; but also photomontages and so-called
three-dimensional photomodels, which are textured CAD models with textures
extracted from photographs.
That branch of photogrammetry which starts with conventional photographs and in

which the processing is by means of optical-mechanical instruments is called analogue
photogrammetry. That which is based on conventional photographs but which resolves
the whole process of analysis by means of computers is called analytical photogram-
metry. A third stage of development is digital photogrammetry. In that case the light
falling on the focal plane of the taking camera is recorded not by means of a light-
sensitive emulsion but by means of electronic detectors. Starting from such digital
photographs, the whole process of evaluation is by means of computers—human vi-
sion and perception are emulated by the computer. Especially in English, digital pho-
togrammetry is frequently called softcopy photogrammetry as opposed to hardcopy
photogrammetry which works with digitized film-based photographs1. Photogramme-
try has some connection with machine vision, or computer vision, of which pattern
recognition is one aspect.
'See PE&RS 58, Copy 1, pp. 49-115, 1992.

2 Chapter 1 Introduction
In many cases interpretation of the content of the image goes hand in hand with the
geometrical reconstruction of the photographed object. The outcome of such pho-
tointerpretation is the classification of objects within the images according to various
different characteristics.
Photogrammetry allows the reconstruction of an object and the analysis of its character-
istics without physical contact with it. Acquisition of information about the surface of
the Earth in this way is known nowadays as remote sensing. Remote sensing embraces
all methods of acquiring information about the Earth's surface by means of measure-
ment and interpretation of electromagnetic radiation 2 either reflected from or emitted
by it. While remote sensing includes that part of photogrammetry which concerns itself
with the surface of the Earth, if the predominant interest is in geometric characteristics,
one speaks of photogrammetry and not of remote sensing.
1.2 Applications
The principal application of photogrammetry lies in the production of topographic
maps in the form of both line maps and orthophoto maps. Photogrammetric instru-
ments function as 3D-digitizers; in a photogrammetric analysis a digital topographic
model is formed, which can be visualized with the aid of computer graphics. Both the
form and the usage of the surface of the Earth are stored in such a digital topographic
model. The digital topographic models are input in a topographical information sys-
tem as the central body of data which, speaking veiy generally, provides information
about both the natural landscape and the cultural landscape (as fashioned by man). A
topographic information system is a fundamental subsystem in a comprehensive geoin-
formation system (GIS). Photogrammetry delivers geodata to a GIS. Nowadays a very
large proportion of geodata is recorded by means of photogrammetry and laser scan-
ning.
Close range photogrammetry is used for the following tasks: architectural recording;
precision measurement of building sites and other engineering subjects; surveillance
of buildings and documentation of damage to buildings; measuring up of artistic and
engineering models; deformation measurement; survey of moving processes (for ex-
ample, robotics); biometric applications (for example, computer controlled surgical
operations); reconstruction of traffic accidents and very many others.
If the photographs are taken with specialized cameras, photogrammetric processing is

relatively simple. With the help of complex mathematical algorithms and powerful
software, however, the geometric processing of amateur photographs has now become
possible. This processing technology is becoming more and more widely used, espe-
cially now that many people have their photographs available on their computers and,
in addition to manipulation of density and colour, are frequently interested in geometric
processing.
2
See DIN 18716/3.
Section 1.3 Some remarks on historical development 3
1.3 Some remarks on historical development3

Technologies arise and develop historically in response both to need and to the emer-
gence and development of supporting techniques and technologies. With the invention
of photography by Fox Talbot in England, by Niepce and Daguerre in France, and
by others, the 1830s and 1840s saw the culmination of investigations extending over
the centuries into optics and into the photo-responses of numerous chemicals. Also
at that time, rapid and cost-effective methods of mapping were of crucial interest to
military organizations, to colonial powers and to those seeking to develop large, rela-
tively new nations such as Canada and the USA. While the practical application of new
technology typically lags well behind its invention, it was very quickly recognized that
cameras furnished a means of recording not only pictorial but also geometrical infor-
mation, with the result that photogrammetry was born only a few years after cameras
became available. Surprisingly, it was not the urgent needs of mapping but the desire
accurately to record important buildings which led to the first serious and sustained
application of photogrammetry and it was not a surveyor but an architect, the German
Meydenbauer 4 , who was responsible. In fact it is to Meydenbauer that we should be
grateful, or not, for having coined the word "photogrammetry". Between his first and
last completed projects, in 1858 and 1909 respectively, on behalf of the Prussian state,
Meydenbauer compiled an archive of some 16000 metric images of its most important
architectural monuments.
Meydenbauer had, however, been preceded in 1849 by the Frenchman Laussedat, a

military officer, and it is he who is universally regarded as the first photogrammetrist
despite the fact that he was initially using not a camera but a camera lucida, working
on an image of a facade of the Hotel des Invalides in Paris. The work of both of
these scientists had been foreshadowed by others. In 1839 the French physicist Arago
had written that photography could serve "to measure the highest and inaccessible
buildings and to replace the fieldwork of a topographer". Earlier than this, in 1759,
Lambert, a German mathematician, had published a treatise on how to reconstruct
three-dimensional objects from perspective drawings.
The effective production of maps using photogrammetry, which was to become a tech-
nological triumph of the 20 th century, was not possible at that time, nor for decades
afterwards; that triumph had to wait for several critical developments: the invention
of stereoscopic measurement, the introduction of the aeroplane and progress in the de-
velopment of specialized analogue computers. For reasons which will become clear to
readers of this book, buildings provided ideal subjects for the photogrammetric tech-
niques of the time; topographic features most certainly did not. Without stereoscopy,
measurement could be made only of very clearly defined points such as are to be found
on buildings. Using cameras with known orientations and known positions, the three-
dimensional coordinates of points defining a building being measured photogrammet-
3
Permission from the publishers to use some of the historical material from "Luhmann, T., Robson,
S., Kyle, S., Harley, I.: Close Range Photogrammetry. Whittles Publishing, 2006" is gratefully acknowl-
edged. That material and this present section were both contributed by one of the translators, Ian Harley.
4
A limited bibliography, with particular reference to historical development, is given at the end of this
chapter.
rically were deduced using numerical computation. The basic computational methods
of photogrammetry were established long ago.
By virtue of their regular and distinct features, architectural subjects lend themselves
to this technique which, despite the fact that numerical computation was employed,
is often referred to as "plane table photogrammetry". When using terrestrial pictures
in mapping, by contrast, there was a major difficulty in identifying the same point on
different photographs, especially when they were taken from widely separated camera
stations; and a wide separation is desirable for accuracy. It is for these reasons that so
much more architectural than topographic photogrammetry was performed during the
19th century. Nonetheless, a certain amount of topographic mapping by photogramme-
try took place during the last three decades of that century; for example mapping in the
Alps by Paganini in 1884 and the mapping of vast areas of the Rockies in Canada by
Deville, especially between 1888 and 1896. Jordan mapped the Dachel Oasis in 1873.
In considering the history of photogrammetry the work of Scheimpflug in Austria

should not be overlooked. In 1898 he first demonstrated double projection, which
foreshadowed purely optical stereoplotters. In particular his name will always be asso-
ciated with developments in rectification.
The development of stereoscopic measurement around the turn of the century was a
momentous breakthrough in the history of photogrammetry. The stereoscope had al-
ready been invented between 1830 and 1832 and Stolze had discovered the principle
of the floating measuring mark in Germany in 1893. Two other scientists, Pulfrich in
Germany and Fourcade in South Africa, working independently and almost simultane-
ously 5 , developed instruments for the practical application of Stolze's discovery. Their
stereocomparators permitted stereoscopic identification of, and the setting of measur-
ing marks on, identical points in two pictures. The survey work proceeded point by
point using numerical intersection in three dimensions. Although the landscape could
be seen stereoscopically in three dimensions, contours still had to be plotted by inter-
polation between spot heights.
Efforts were therefore directed towards developing a means of continuous measure-

ment and plotting of features, in particular of contours—the "automatic" plotting ma-
chine, in which numerical computation was replaced by analogue computation for re-
section, relative and absolute orientation and, above all, for intersection of rays. Digital
computation was too slow to allow the unbroken plotting of detail, in particular of con-
tours, which stereoscopic measurement seemed to offer so tantalisingly. Only analogue
computation was fast enough to provide continuous feedback to the operator. In several
countries during the latter part of the 19th century, much effort and imagination was di-
rected towards the invention of stereoplotting instruments, necessary for the accurate
and continuous plotting of topography. In Germany Hauck proposed such an apparatus.
In Canada Deville developed what was described by Ε. H. Thompson as "the first auto-
matic plotting instrument in the history of photogrammetry". Deville's instrument had
several defects, but its design inspired several subsequent workers to overcome these,
5
Pulfrich's lecture in Hamburg announcing his invention was given on 23 rd September 1901, while
Fourcade delivered his paper in Cape Town nine days later on 2 nd October 1901.
including both Pulfrich, one of the greatest contributors to photogrammetric instrumen-

tation, and Santoni in Italy, perhaps the most prolific of photogrammetric inventors.
Photogrammetry was about to enter the era of analogue computation, a very foreign
idea to surveyors with their long tradition of numerical computation. Although many
surveyors regarded analogue computation as an aberration, it became a remarkably
successful one for a large part of the 20th century.
In Germany, conceivably the most active country in the early days of photogrammetry,
Pulfrich's methods were very successfully used in mapping; this inspired von Orel in
Vienna to design an instrument for the "automatic" plotting of contours, leading ulti-
mately to the Orel-Zeiss Stereoautograph which came into productive use in 1909. In
England, F. V. Thompson was slightly before von Orel in the design and use of the
Vivian Thompson Stereoplotter; he went on to design the Vivian Thompson Stereo-
planigraph, described in January 1908, about which Ε. H. Thompson was to write that
it was "the first design for a completely automatic and thoroughly rigorous photogram-
metric plotting instrument". The von Orel and the Thompson instruments were both
used successfully in practical mapping, Vivian Thompson's having been used by the
Survey of India which bought two of the instruments.
The advantages of photography from an aerial platform, rather than from a ground
point, are obvious, both for reconnaissance and for survey; in 1858 Nadar, a Paris
photographer, took the first such picture, from a hot-air balloon 1200 feet above that
city, and in the following year he was ordered by Napoleon to obtain reconnaissance
photographs in preparation for the Battle of Solferino. It is reputed that balloon pho-
tography was used during the following decade in the American Civil War. The rapid
development of aviation which began shortly before the first World War had a decisive
influence on the course of photogrammetry. Not only is the Earth, photographed ver-
tically from above, an almost ideal subject for the photogrammetric method, but also
aircraft made almost all parts of the Earth accessible at high speed. In the first half, and
more, of the 20th century these favourable circumstances allowed impressive develop-
ment in photogrammetry, although the tremendous economic benefit in air survey was
not fully felt until the middle of that century. On the other hand, while stereoscopy
opened the way for the application of photogrammetry to the most complex surfaces
such as might be found in close range work, not only is the geometry in such cases of-
ten far from ideal photogrammetrically but also there was no corresponding economic
advantage to promote its application.
In the period before the first World War all the major powers followed similar paths in
the development of photogrammetry. After the war, although there was considerable
opposition from surveyors to the use of photographs and analogue instruments for map-
ping, the development of stereoscopic measuring instruments forged ahead remarkably
in very many countries; while the continental European countries broadly speaking
put most of their effort into instrumental methods, Germany and the Austro-Hungarian
Empire having a clear lead in this field, the English-speaking countries focused on
graphical techniques. It is probably true that until about the 1930s the instrumen-
tal techniques could not compete in cost or efficiency with the British and American
methods.
Zeiss, in the period following WWI, was well ahead in the design and manufacture
of photogrammetric instruments, benefiting from the work of leading figures such as
Pulfrich, von Orel, Bauersfeld, Sander and von Gruber. In Italy, around 1920, Santoni
produced a prototype, the first of many mechanical projection instruments designed
throughout his lifetime, while the Nistri brothers developed an optical projection plot-
ter, shortly afterwards founding the instrument firm OMI. Poivilliers in France began
the design and construction of analogue photogrammetric plotters in the early 1920s.
In Switzerland the scene was dominated by Wild whose company began to produce in-
strumentation for terrestrial photogrammetry at about the same time; Wild Heerbrugg
very rapidly developed into a major player, not only in photogrammetric instrumenta-
tion, including aerial cameras, but also in the wider survey world. As early as 1933
Wild stereometric cameras were being manufactured and were in use by Swiss po-
lice for the mapping of accident sites, using the Wild A4 Stereoautograph, a plotter
especially designed for this purpose. Despite the ultra-conservative establishment in
the British survey world at that time, Ε. H. Thompson was able to design and build a
stereoplotter in the late 1930s influenced by the ideas of Fourcade. While the one such
instrument in existence was destroyed by aerial bombing, the Thompson-Watts plotter
was later based on this prototype in the 1950s.
Meanwhile, non-topographic use was sporadic for the reasons that there were few suit-
able cameras and that analogue plotters imposed severe restrictions on principal dis-
tance, on image format and on disposition and tilts of cameras.
The 1950s saw the beginnings of the period of analytical photogrammetry. The expand-
ing use of digital, electronic computers in that decade engendered widespread interest
in the purely analytical or numerical approach to photogrammetry as against the pre-
vailing analogue methods. While analogue computation is inflexible, in regard to both
input parameters and output results, and its accuracy is limited by physical properties,
a numerical method allows virtually unlimited accuracy of computation and its flex-
ibility is bounded only by the mathematical model on which it is based. Above all,
it permits over-determination which may improve precision, lead to the detection of
gross errors and provide valuable statistical information about the measurements and
the results. The first analytical applications were to photogrammetric triangulation, a
technique which permits a significant reduction in the amount of ground control re-
quired when mapping from a strip or a block of aerial photographs; because of the very
high cost of field survey for control, such techniques had long been investigated. In
the 1930s, the slotted template method of triangulation in plan was developed in the
USA, based on theoretical work by Adams, Finsterwalder and Hotine. Up until the
1960s vast areas were mapped in the USA and Australia using this technique in plan
and one of the many versions of the simple optical-projection Multiplex plotters both
for triangulation in height and for plotting of detail. At the same time, precise analogue
instruments such as the Zeiss C8 and the Wild A7 were being widely used for analogue
triangulation in three dimensions.
Analytical photogrammetric triangulation is a method, using numerical data, of point

determination involving the simultaneous three-dimensional orientation of all the pho-
tographs and taking all inter-relations into account. Work on this line of development
had appeared before WWII, long before the development of electronic computers. An-
alytical triangulation demanded instruments to measure photo coordinates. The first
stereocomparator designed specifically for use with aerial photographs was the Cam-
bridge Stereocomparator designed in 1937 by Ε. H. Thompson. Electronic recording
of data for input to computers became possible and by the mid-1950s there were five
automatic recording stereocomparators on the market and monocomparators designed
for use with aerial photographs also appeared.
Seminal papers by Schmid and Brown in the late 1950s laid the foundations for theo-
retically rigorous photogrammetric triangulation. A number of block adjustment pro-
grams for air survey were developed and became commercially available, such as those
by Ackermann.
Subsequently, stereoplotters were equipped with devices to record model coordinates

for input to electronic computers. Arising from the pioneering ideas of Helava, comput-
ers were incorporated in stereoplotters themselves, resulting in analytical stereoplotters
with fully numerical reconstruction of the photogrammetric models. Bendix/OMI de-
veloped the first analytical plotter, the AP/C, in 1964; during the following two decades
analytical stereoplotters were produced by the major instrument companies and others.
Photogrammetry has progressed as supporting sciences and technologies have supplied

the means such as better glass, photographic film emulsions, plastic film material, aero-
planes, lens design and manufacture, mechanical design of cameras, flight navigation
systems. Progress in space technology (both for imaging, in particular after the launch
of SPOT-1 in 1986, and for positioning both on the ground and in-flight by GNSS) and
the continuing explosion in electronic information processing have profound implica-
tions for photogrammetry.
The introduction of digital cameras into a photogrammetric system allows automation,

nowhere more completely than in industrial photogrammetry, but also in mapping.
Advanced computer technology enables the processing of digital images, particularly
for automatic recognition and measurement of image features, including pattern cor-
relation for determining object surfaces. Procedures in which both the image and its
photogrammetric processing are digital are often referred to as digital photogrammetry.
Interactive digital stereo systems (e.g. Leica/Helava DSP, Zeiss PHODIS) have existed
since around 1988 (Kern DSP 1) and have increasingly replaced analytical plotters.
To some extent, photogrammetry has been de-skilled and made available directly to a
wide range of users. Space imagery is commonplace, as exemplified by Google Earth.
Photogrammetric measurement may be made by miscellaneous users with little or no
knowledge of the subject—police, architects, model builders for example.
Although development continues apace, photogrammetry is a mature technology with

a history of remarkable success. At the start of the 20 th century topographic mapping
of high quality existed, in general, only in parts of Europe, North America and India.
Although adequate mapping is acknowledged as necessary for development but is still
lacking in large parts of the world, such deficiencies arise for political and economic
reasons, not for technical reasons. Photogrammetry has revolutionized cartography.
Further reading. Adams, L.P.: Fourcade: The centenary of a stereoscopic method

of photographic surveying. Ph.Rec. 17(99), pp. 225-242, 2001 · Albertz, J.: A Look
Back—Albrecht Meydenbauer. PE&RS 73, pp. 504-506, 2007 . Atkinson, K.B.:
Vivian Thompson (1880-1917): not only an officer in the Royal Engineers. Ph.Rec.
10(55), pp. 5-38, 1980 · Atkinson, K.B.: Fourcade: The Centenary—Response
to Professor H.-K. Meier. Correspondence, Ph.Rec. 17(99), pp. 555-556, 2002 .
Babington-Smith, C.: The Story of Photo Intelligence in World War II. ASPRS, Falls
Church, Virginia, 1985, reprint · Blachut, T.J. and Burkhardt, R.: Historical de-
velopment of photogrammetric methods and instruments. ISPRS and ASPRS, Falls
Church, Virginia, 1989 · Brown, D.C.: A solution to the general problem of multiple
station analytical Stereotriangulation. RCA Data Reduction Technical Report No. 43,
Aberdeen, 1958 · Brown, D.C.: The bundle adjustment—progress and prospectives.
IAPR 21(3), ISP Congress, Helsinki, pp. 1-33, 1976 · Deville, E.: Photographic
Surveying. Government Printing Bureau, Ottawa, 1895. 232 pages · Deville, E.:
On the use of the Wheatstone Stereoscope in Photographic Surveying. Transactions
of the Royal Society of Canada, Ottawa, 8, pp. 63-69, 1902 . Fourcade, H.G.: On
a stereoscopic method of photographic surveying. Transactions of the South African
Philosophical Society 14(1), pp. 28-35, 1901. Also published in: Nature 66(1701),
pp. 139-141, 1902 · Fourcade, H.G.: On instruments and methods for stereoscopic
surveying. Transactions of the Royal Society of South Africa 14, pp. 1-50, 1903 ·
Fräser, C.S., Brown, D.C.: Industrial photogrammetry—new developments and recent
applications. Ph.Rec. 12(68), pp. 197-216, 1986 . von Gruber, 0., (ed.), McCaw,
G.T., Cazalet, F.A., (trans.): Photogrammetry, Collected Lectures and Essays. Chap-
man & Hall, London, 1932 · Gruen, Α.: Adaptive least squares correlation—a pow-
erful image matching technique. South African Journal of Photogrammetry, Remote
Sensing and Cartography 14(3), pp. 175-187, 1985 · Harley, I.A.: Some notes on
stereocomparators. Ph.Rec. IV(21), pp. 194-209, 1963 · Helava,U.V.: New principle
for analytical plotters. Phia 14, pp. 89-96, 1957 · Kelsh, H.T.: The slotted-template
method for controlling maps made from aerial photographs. Miscellaneous Publica-
tions no. 404, U.S. Department of Agriculture, Washington, D.C., 1940 · Landen,
D.: History of photogrammetry in the United States. Photogrammetric Engineering
18, pp. 854-898, 1952 · Laussedat, Α.: Memoire sur l'emploi de la photographie
dans le leve des plans. Comptes Rendus 50, pp. 1127-1134, 1860 · Laussedat, Α.:
Recherches sur les instruments, les methodes et le dessin topographiques. Gauthier-
Villars, Paris, 1898 (vol. 1), 1901 (vol. 2 part 1), 1903 (vol. 2 part 2) · Luhmann, T„
Robson, S., Kyle, S. and Harley, I: Close Range Photogrammetry. Whittles Publishing,
2006. 510 pages · Mason, K.: The Thompson stereo-plotter and its use. Survey of
India Departmental Paper No. 5, Survey of India, Dehra Dun, India, 1913 · Meier,
H.-K.: Fourcade: The Centenary—Paper by L.P. Adams. Correspondence, Ph.Rec.
17(99), pp. 554-555, 2002 · Meydenbauer, Α.: Handbuch der Messbildkunst. Knapp,
Halle, 1912. 245 pages · Poivilliers, G.: Address delivered at the opening of the His-
torical Exhibition, Ninth International Congress of Photogrammetry, London, 1960.
IAPR XIII(l), 1961 · Pulfrich, C.: Über neuere Anwendungen der Stereoskopie und
über einen hierfür bestimmten Stereo-Komparator. Zeitschrift für Instrumentenkunde
22(3), pp. 65-81, 1902 · Sander, W.: The development of photogrammetry in the
light of invention, with special reference to plotting from two photographs. In: von
Gruber, O., (ed.), McCaw, G.T., Cazalet, F.A., (trans.): Photogrammetry, Collected
Lectures and Essays. Chapman & Hall, London, pp. 148-246, 1932 · Santoni, E.:
Instruments and photogrammetric devices for survey of close-up subject with particu-
lar stress to that of car bodies and their components photographed at various distances.
Inedito. 1966. Reprinted in Selected Works, Ermenegildo Santoni, Scritti Scelti 1925—
1968. Societä Italiana di Fotogrammetria e Topographia, Firenze, 1971 · Schmid,
H.: An analytical treatment of the problem of triangulation by stereophotogrammetry.
Phia XIII(2/3), 1956-57 · Schmid, H.: Eine allgemeine analytische Lösung für die
Aufgabe der Photogrammetrie. BuL. 1958(4), pp. 103-113, and 1959(1), pp. 1-12 ·
Thompson, E.H.: Photogrammetry. The Royal Engineers Journal 76(4), pp. 432^44,
1962. Reprinted as Photogrammetry, in: Photogrammetry and surveying, a selection
of papers by E.H. Thompson, 1910-1976. Photogrammetric Society, London, 1977
• Thompson, E.H.: The Deville Memorial Lecture. Can.Surv. 19(3), pp. 262-272,
1965. Reprinted in: Photogrammetry and surveying, a selection of papers by E.H.
Thompson, 1910-1976. Photogrammetric Society, London, 1977 · Thompson, E.H.:
The Vivian Thompson Stereo-Planigraph. Ph.Rec. 8(43), pp. 81-86, 1974 · Thomp-
son, V.F.: Stereo-photo-surveying. The Geographical Journal 31, pp. 534-561, 1908
• Wheatstone, C.: Contribution to the physiology of vision—Part the first. On some
remarkable, and hitherto unobserved, phenomena of binocular vision. Philosophical
Transactions of the Royal Society of London for the year MDCCCXXXVIII, Part II,
pp. 371-394, 1838
Chapter 2
Preparatory remarks on mathematics

and digital image processing
Section 2.1 concerns itself entirely with introductory mathematics. Section 2.2 includes
notes on digital image processing in preparation for procedures of digital photogram-
metric processing.
2.1 Preparatory mathematical remarks

The various techniques for photogrammetric processing assume knowledge of basic
mathematics. While important mathematical matters may have been treated in lec-
tures and textbooks, some mathematical themes of importance for photogrammetry are
compiled in what follows.
2.1.1 Rotation in a plane, similarity and affine transformations

Given, a point P(x, y) in a plane coordinate system (see Figure 2.1-1) which has been
rotated through an angle a in an counterclockwise direction relative to a fixed coordi-
nate system, we wish to find the coordinates (X, Y) of the point Ρ with respect to the
fixed coordinate system.
X = χ cos a — y sin α
(2.1-1)
y = χ sin a + y cos a
If we introduce the cosines of the angles between the coordinate axes and use matrix
notation we have:
cos(ZxX) cos(ZyX)
(2.1-2)
cos(ZxY) cos(ZyY)
Representing matrices and vectors by means of bold symbols this becomes:
X = Rx, (2.1-3)
R is called a rotation matrix. It is square but not symmetric. The elements of R are the
cosines of the angles between the coordinate axes.
Section 2.1 Preparatory mathematical remarks 11
Figure 2.1-1: Plane rotation Figure 2.1-2: Introduction of the unit

vectors
Properties of the rotation matrix R
The question arises whether the four elements r ^ may be freely chosen or whether they
must satisfy certain conditions. To answer this question we introduce the unit vectors
i and j (Figure 2.1-2) along the coordinate axes χ and y. We express their components
in the XY system as:
i= /cosoA (2.1-4)
J v
y sin a J \ cos a J '
A comparison between Equations (2.1-1) and (2.1-4) shows that the elements r,k of the
rotation matrix are none other than the components of the unit vectors i and j.
R = ( i , j) (2.1-5)
The two mutually orthogonal unit vectors must, however, satisfy the orthogonality con-
ditions in Equation (2.1-6)'. These conditions are formulated in Equation (2.1-6) as
inner, or scalar, products of the two vectors in which transposition is signified by the
superscript T .
iTi = cos2 α + sin2 α = 1 = r 2 , + r\x

T
j j = 2
sin α + cos α = 1 = r\2 + r 2 2
2
(2.1-6)
T
i j = — cos α sin α + sin α cos α = 0 = τ \ \ τ η + τι\τ22
A matrix which satisfies the orthogonality conditions is known as an orthogonal ma-
trix2. If the four elements of the rotation matrix must satisfy the three orthogonality
1
More strictly, one orthogonal condition and two normalizing conditions.
2
In amplification of the preceding footnote, when d e t R = 1 the matrix is said to be a proper or-
thogonal matrix and when det R = — 1 the matrix is said to be improper. In the latter case the matrix
represents a rotation and a reflection.
12 Chapter 2 Preparatory remarks on mathematics and digital image processing
conditions, only one parameter may be freely chosen; in general this parameter is the
rotation angle a.
Numerical Example.
x 0 3 6 0 6 9 x
f \ - f ^ f \
[ Υ J ~ \ 0.19 0.27 J [yj
In the above case, the orthogonality conditions are not fulfilled; the transformation does
not represent a rotation. We deal with this transformation at the end of this section.
( X \ _ ( 0.6234 -0.7819 \ f x \
\Y ) ~ V 0.7819 0.6234 ) \y )
Here the orthogonality conditions are fulfilled; this means that under this transforma-
tion a field of points will be rotated.
Exercise 2.1-1. Consider a rectangle in an xy system, the vertices being transformed

into the XY system using the matrices of both numerical examples. Using the results,
consider the characteristics of both transformations.
Exercise 2.1-2. Think about the characteristics of the transformation when just one of
the three orthogonality conditions of Equation (2.1-6) is not fulfilled.
Exercise 2.1-3. Consider a matrix which brings about both a rotation and a mirror
reflection. (Answer: r n = cos a ; r n = sin a ; r2\ = sin α; Γ22 = — cos a).
Inverting the rotation matrix R
By definition, multiplication of the inverted matrix R _ 1 by the matrix R gives the unit
matrix I:
R_1R — I
On the other hand multiplication of the transposed matrix R T with the matrix R also
gives the unit matrix (Equations (2.1-5) and (2.1-6)):
As a consequence, we see that the following important result holds for the rotation
matrix:
R_1 = RT (2.1-7)
Reverse transformation
If one wishes to transform points from the fixed XY system into the xy system, one
obtains the desired rotation matrix as follows:
From Equation (2.1-3):
X = Rx
Premultiplication by R T gives:
R T X = R T R x = Ix = χ
Rewriting this result:

χ = RTX = r\ι r21
r 12 r22
Exercise 2.1-4. How would Equations (2.1-1), (2.1-2), (2.1-4) and (2.1-5) appear if
the rotation of the xy system had been made in a clockwise sense with respect to the
X Y system?
A transformation with a non-orthogonal matrix (see the first numerical example) is

known as an affine transformation. It has the following characteristics:
• orthogonal straight lines defined, for example, by three points in the xy system
are no longer orthogonal after the transformation.
• parallel straight lines defined, for example, by four points in the xy system, re-
main parallel after the transformation.
• line-segments between two points in the xy system exhibit a different length after
the transformation.
• on the other hand, the ratio of the lengths of two parallel line segments is invariant
under the transformation.
The affine transformation is of the form:
(2.1-8)
in which
• αιο and a2o are two translations (or, more exactly, the X Y coordinates of the
origin of the xy system) and
• oil, a n , a 2i and a 2 2 are four elements which do not satisfy the orthogonality
conditions of Equation (2.1-6) and which consequently allow not only different
scales in the two coordinate directions but also independent rotations of the two
coordinate axes.
In order to determine the six parameters α ^ one requires at least three common points
in both coordinate systems.
Numerical Example. Given three points with their coordinates in both systems, we
wish to find the six parameters a ^ of the affine transformation.
Pt.No. X y X Y
23 0.3035 0.5951 3322 1168
24 0.1926 0.6028 3403 2061
50 0.3038 0.4035 1777 1197
Using the linear Equations (2.1-8) we obtain the following system of equations
/I
1
0.3035 0.5951
0.3035 0.5951
\ ( «10 \ / 3322 \
1168
O20
1 0.1926 0.6028 an 3403
1 0.1926 0.6028 an 2061
1 0.3038 0.4035 a21 1777
V 1 0.3038 0.4035 y \ «22 / 1,1197 )
of which the solution is
fX \ _ f-1425 \ / - 1 7 1 . 5 8063.4 λ f x\
3713 / \ —8063.7 164.0 J J
One obtains a similarity transformation by replacing the nonorthogonal matrix A of

Equation (2.1-8) with an (orthogonal) rotation matrix R and introducing a unit scale
factor, m. The similarity transformation is of the form:
In order to determine the four parameters of the plane similarity transformation (two
translations αιο and 020, a scale factor m and, for example, a rotation angle a of the
rotation matrix R ) one requires at least two common points in each coordinate system.
The solution of this problem is discussed in Section 5.2.1.
Note: a square in the xy system remains a square after the transformation; it is simply
shifted, rotated and changed in scale. Against that, after an affine transformation it
becomes a parallelogram.
2.1.2 Rotation, affine and similarity transformations in

three-dimensional space
Based on Equation (2.1-2), the rotation in space of a point Ρ with coordinates (x, y, ζ)
in a fixed coordinate system, X Y Z , may be formulated as follows, using the cosines
of the angles between the coordinate axes:
f X \ / c o s (ZxX) cos (ZyX) cos {ZzX) \ f x\

y = cos (ZxY) cos (ZyY) cos (ZzY) 2/ (2.1-10)
\Z J \ cos (ZxZ) cos (ZyZ) cos (ZzZ) ) \z J
Figure 2.1-3: Rotation in three-dimensional space
m Π2 n 3
X = Rx; R . = I Γ21 T-22 r23 (2.1-11)
r-ii r 3 2 r 3 3
In a similar manner to that of Equation (2.1-5), the matrix R can be formed from the
three unit vectors shown in Figure 2.1-3; R = (i, j, k). 3 It is simple to write out the
following six orthogonality 4 relationships among the nine elements rik for the three-
dimensional case.
iTi - jTj = k ' k = 1
T (2.1-12)
i j = jTk = k i = 0
T
That is, a rotation in three dimensions is prescribed by three independent parameters.

In photogrammetry we frequently use three rotation angles ω, φ and κ about the three
coordinate axes. In this case, a hierarchy of axes is to be observed, as can be clearly
demonstrated with gimbal (or Cardan) axes (Figure 2.1-4):
The three unit vectors with their components n t are related, through their vector products, as fol-
lows:
Γ22 Γ23
T32 f"33
Γ22Γ33 — Γ32Γ23 j = k χ i
r 12 ΤΊ3
=j x k = r32r-l3 — Γ12Γ33
?"32 ^33 k = ix j
rnrn — Γ22Π3
r\2 ri3
+
Γ22 V23
4
More accurately, three orthogonality conditions and three normalizing conditions (see also the foot-
note to Equation (2.1-6) on page 11).
κ
ci?
<T>
φ
Figure 2.1-4: Rotations about the axes in gimbals
X =X(o
primary secondary tertiary
Figure 2.1-5: Hierarchy of the three rotations about the coordinate axes
If one performs an ω rotation, the attitudes in space of the other two axes are changed
accordingly. If, however, one rotates in φ, only the κ axis, and not the ω axis, is
affected. Rotation about the κ axis changes the attitude of neither of the other two axes.
An arbitrary rotation of the xyz system as illustrated in Figure 2.1-3 can therefore be
effected by means of three rotations ω, φ and κ. In each case the rotation is to be seen
as counterclockwise when viewed along the axis towards the origin.
The transformation into the X Y Z system of a point P , given in the xyz coordinate
system, may therefore be defined in terms of the three rotation angles ω, φ and κ. In
this case the matrix R of Equation (2.1-11) has the form5 (see Appendix 2.1-1):
5
In the following equation, the functions cos and sin are abbreviated by c and s, respectively.
'ωφκ — (2.1-13)
Exercise 2.1-5. Show, using trigonometrical relationships, that the nine elements of
the rotation matrix (2.1-13) fulfill the orthogonality conditions (2.1-12).
If the sequence of the rotations is defined in a different order, the elements of the matrix
(2.1-13) are also changed (see Appendix 2.1-1 and especially Section Β 3.4, Volume 2).
As in Equation (2.1-7) the inverse of the rotation matrix, R _ 1 , is the transposed matrix
R T , by virtue of the orthogonality conditions (2.1-12). To summarize, three differ-
ent interpretations have been given for the elements of the three-dimensional rotation
matrix R:
• cosines of the angles between the axes of the two coordinate systems
• components of the unit vectors of the rotated coordinate axes with respect to the
fixed system
• trigonometric functions of rotation angles about the three axes of a gimbal system
Two successive rotations

First rotation: X] = R i x
Second rotation: X2 = R2X1
Complete rotation:
Xi = R2 R1 χ = R x (2.1-14)
By multiplication, the rotation matrices, Ri and R2, are combined as a single rotation
matrix R:
R — R2R1 (2.1-15)
Since matrix multiplication is not commutative, the order of multiplication of the ma-
trices must be strictly observed. In the transposed matrix R T the sequence must be
reversed:
RT — ( R 2 R i ) t — R { R j (2.1-16)
Example (of Equation (2.1-13)).

Given: ω = —1.3948 gon = - Γ 1 5 Ί 9 "
ψ= 0.1041 gon = 5'37"
κ = -0.8479 gon = -45'47"
Required: elements of the rotation matrix R
Checks (see also Appendix 2.1-1):
sin</j = rj3 =>• ψ — 0.1041 gon 5'37"

tan κ = —r\2/r\\ =>· κ = - 0 . 8 4 7 9 gon —45'47"
tana; = — r23/r33 =φ· ω — - 1 . 3 9 4 8 gon —1°15'19"
The nine elements r ^ must satisfy the orthogonality conditions 6 ; one checks that
R T R = I:
( 0.999910 0.013319 0.001635

-0.013351 0.999671 0.021907
\ -0.001343 -0.021927 0.999759
0.999910 0.013351 - 0 . 0 0 1 3 4 3 \ / 1.000000 0.000001 0.000000

0.013319 0.999671 -0.021927 0.000001 1.000000 0.000000
0.001635 0.021907 0.999759 V 0.000000 0.000000 1.000001
Example (of Equation (2.1-14)). We are given a point Ρ in an xyz system which is
rotated by ωι, φ\, κ\ relative to an X\Y\Z\ coordinate system. The X\Y\Z\ system is
then rotated by ωι, ψ2, «2 relative to an X2Y2Z2 system.
We wish to find the final coordinates Χι, Y2, Z2 of the point Ρ and the angles ω, ψ, κ
by which the xyz system is rotated relative to an X2Y2Z2 system.
Given coordinates of P:
-43.461
χ = -83.699
152.670
First rotation:
ωι - 1 . 3 9 4 8 gon = -1°15'19"
Ψ\ +0.1041 gon = 5'37"
Κ\ - 0 . 8 4 7 9 gon = -45'47"
0.999910 0.013319 0.001635

Ri = -0.013351 0.999671 0.021907
-0.001343 - 0 . 0 2 1 9 2 7 0.999759
Second rotation:
U>2 - 0 . 1 7 2 6 gon = -9'19"
Ψ2 - 1 . 0 8 5 3 gon = -58'36"
K2 -101.3223 gon = -91°11'24"
0.020770 0.999639 -0.017047'

r 2 = ι - 0.999782 - 0 . 0 2 0 7 2 7 0.002710
0.002355 0.017100 0.999851
For didactical reasons, the notation of a matrix multiplication A B = C is often written in "Falk's
Β
scheme":
A C
First solution (two stage)

First rotation of Ρ (to XxY\ Zx)
-43.461
-83.699
152.670
0.999910 0.013319 0.001635' -44.3223'
-0.013351 0.999671 0.021907 -79.7467 = Xi = R i x
-0.001343 -0.021927 0.999759 154.5268
Second rotation of Ρ (to X2Y2Z2)
-44.3223
-79.7467
154.5268
-0.020770 0.999639 -0.017047 \ '-81.432'
-0.999782 -0.020727 0.002710 46.384 X2 — R2X1
0.002355 0.017100 0.999851 / 153.036
Second solution (one stage)

Rotation matrix R = R2R1 for the combined rotations:
0.999910 0.013319 0.001635
-0.013351 0.999671 0.021907
-0.001343 -0.021927 0.999759
-0.020770 0.999639 -0.017047 \ -0.034091 0.999407 0.004822

-0.999782 -0.020727 0.002710 -0.999419 -0.034096 0.000621
0.002355 0.017100 0.999851 J 0.000784 -0.004798 0.999988
Transformation of Ρ (to X2Y2Z2)'·
-43.461
-83.699
152.670
-0.034091 0.999407 0.004822 -81.432'
-0.999419 -0.034096 0.000621 46.384 = X2 = Rx
0.000784 -0.004798 0.999988 153.036
From the definition of the elements of the rotation matrix R (see Equation (2.1-13))
calculate the angles ω, ψ and κ through which the point Ρ has been rotated with respect
to the X2Y2Z2 system.
ω = -0.0396 gon = -2Ό8"

φ = 0.3070 gon = 16'35"
κ = -102.1708 gon = - 9 Ρ 5 7 Ί 3 "
Note: ωι + is not equal to ω; similarly for φ and κ.

Exercise 2.1-6. Transform the rotated point Ρ (X2, Yi, Ζ2) back into the xyz system,
in two stages and in one stage.
In three-dimensional space also, a transformation with a non-orthogonal matrix is

called an affine transformation. The characteristics of the plane affine transforma-
tion apply also to a three-dimensional affine transformation (see Section 2.1.1). The
three-dimensional affine transformation has the form:
X\ /αιο\ ί an ai2 a i 3 \ /
γ = Ö20 + I a-21 0,22 0,23 I 2/ I ; X = ao + A x (2.1-17)
Ζ) \ α30 / \ α31 <232 a33 / \z )
in which:
• aio. a20 and aô are three translations (alternatively, the X Y Z coordinates of the
origin of the xyz system).
• αϊ ι, a i 2 , . . . , 033 are the nine elements of R; they do not satisfy the orthogonality
conditions (2.1-12) and consequently they admit not only different scales in the
three coordinate directions but also six independent angles of rotation of the three
coordinate axes (Note: a coordinate axis is defined by two angles).
In order to determine the 12 parameters a ^ , one requires at least four corresponding

points in each coordinate system (by reason of this number of parameters one speaks
sometimes of a 12 parameter transformation. Section 4.4.3 contains a numerical exam-
pie).
The three-dimensional similarity transformation follows from this if one substitutes
an (orthogonal) rotation matrix R for the non-orthogonal matrix A and introduces
a uniform scale factor m. The three-dimensional similarity transformation has the
following form:
X\ ( flio \ / Πι Π2 rl3\ / x\
Y = fl20 + rn r 2 i r22 r23 y ; X = ao + m R x (2.1-18)
z J \ <230 / \ r 31 7-32 r-33 / \z )
For the determination of the seven parameters of a three-dimensional similarity trans-

formation (three translations αιο, «20 and 030, one scale factor m and three rotation
angles ω, ψ and κ defining R as in Equation (2.1-13)), at least seven suitable equations
are required. These equations (for example, two in X, two in Y and three in Z) may
be obtained from three corresponding points in the two systems. The solution of this
non-linear problem is dealt with in Section 4.1.1.
Note: a cube remains a cube after a three-dimensional similarity transformation; it
is simply translated, rotated and changed in scale. After an affine transformation it
becomes a parallelepiped.
1f
c = 99.16
.PP
V
Μ
J <
A k
Figure 2.1-6: Positive and Figure 2.1-7: Metric image

negative positions
2.1.3 Central projection in three-dimensional space

To be able to reconstruct the position and shape of objects we must know the geometry
of the image forming system. Many of the cameras used in photogrammetry, some-
times known as metric cameras, produce photographs which can be considered, with
adequate accuracy, as central projections of the three-dimensional objects in view. (In
Sections 2.1.3 to 2.1.7 we generally assume that we are dealing with analogue pictures.)
Figures 2.1-6 and 2.1-7 show some definitions.
Ο ... centre of perspective of a three-dimensional bundle of rays (also, the

camera location)
PP ... principal point with coordinates ξο,ηο
c ... principal distance (sometimes referred to as the camera constant)
Μ ... fiducial centre (as a coarse approximation, the point of intersection of
the straight lines joining the fiducial marks)
The relationship between the coordinates ξ and η of an image point P' and the coordi-
nates Χ,Υ,Ζ of an object point Ρ is illustrated in Figure 2.1-8 and is mathematically
formulated in Equation (2.1-19)7 (for the derivation of these collinearity equations, as
they are usually called, see Appendix 2.1-2).
7
Notation used in this book:
ξ,η ... two-dimensional image coordinates
x,y,z ... coordinates in a local three-dimensional coordinate system (frequently model
coordinates)
Χ,Υ,Ζ ... coordinates in a control coordinate system (sometimes called a global system;
frequently the national coordinate system)
Figure 2.1-8: Relationship between image and object coordinates
r „ (Χ - X0) + r 2 i (Y - Yp) + r 3 i {Z - Zp)

ξ ξ0 C
m {X - Xo) + 7-23 (Y - Yo) + r 3 3 (z - Zo)
(2.1-19)
m (X - Xo) + 7-22 (Y - Yo) + r 3 2 (Z - Zo)
η ηο C
ru(X-Xo) + r23(Y-Yo) + r33(Z-Zo)
The parameters r,^. appearing in Equations (2.1-19) are the elements of the rotation
matrix R which in this case describes the three-dimensional attitude, or orientation,
of the image with respect to the X Y Z object coordinate system. If so desired, the
elements rik can be expressed in accordance with Equation (2.1-13) in terms of the
three angles ω, φ and κ, which are, respectively, rotations about the X axis, the Υω axis
and the Ζωφ axis, as defined in Figure 2.1-5.
Solving the Equations (2.1-19) for the object coordinates X and Y gives:
X = X0 + (Z- Zo) Πι (ξ - ξο) + Π2 (η - ηο) - rue

Γ31 {ξ - ξο) + 7-32 (ν - ηο) - T-33C
(2.1-20)
7-21 (ξ ~ ξθ) + 7-22 (η ~ ηο) ~ T-23C
Y = Y0 + (Z- Z0)
Γ31 (ξ ~ ξθ) + 1*32 {η ~ ηο) - 7*33C
Equations (2.1-19) mean that to each object point there is one image point. Equa-
tions (2.1-20) draw our attention to the fact that, because the Ζ coordinates are on the
right hand side, to each image point there are infinitely many possible object points.
From a single metric image alone it is not possible to reconstruct a three-dimensional
object. To do so one also needs either a second metric image of the same object taken
from a different place or additional information about the Ζ coordinate (for example
the information that all object points lie on a horizontal plane of known height).
The transformations formulated in Equations (2.1-19) and (2.1-20) assume a know-

ledge of the following independent values:
ξο, Vo • · · image coordinates of the principal point P P ^ j 21)

c . . . principal distance
The above three parameters are known as the elements of interior orientation 8 . They
fix the centre of projection of the three-dimensional bundle of rays with respect to the
image plane.
The following six parameters are the elements of exterior orientation. They define the
position and attitude of the three-dimensional bundle of rays with respect to the object
coordinate system.
Χο,Υο,Ζο . . . object coordinates of the camera station ^

3 parameters defining the rotations of the image (for example, ω, ψ, κ)
To specify the central projection of an image a total of nine parameters is required,

which may be determined in various ways. The values of the three constants of interior
orientation are specific to the camera and are normally determined, at least in the first
instance, by the manufacturer in the laboratory. He tries to ensure that, as closely
as possible, the fiducial centre coincides with the principal point (ξο = ηο = 0). In
terrestrial photogrammetry the six elements of exterior orientation can be established
directly. On the other hand, the elements of exterior orientation of an individual image
from a photographic flight are not known with sufficient accuracy—unless GPS (Global
Position System) and an IMU (Inertial Measurement Unit), both very expensive, are
installed. An alternative, indirect method must be used, involving control points; these
are points for which both image coordinates and object coordinates are known. If the
interior orientation is known one requires three control points, for each control point
yields two Equations (2.1-19) from which the exterior orientation may be computed.
8
In normal English, the orientation of an object implies direction or angular attitude. Photogrammet-
ric usage, deriving from German, applies the word to groups of camera parameters. Exterior orientation
parameters incorporate this angular meaning but extend it to include position. Interior orientation para-
meters, which include a distance, two coordinates and a number of polynomial coefficients, involve no
angular values; the use of the terminology here underlies the connection between two very important,
basic groups of parameters.
Example (with Equations (2.1-19)).

Given:
Interior orientation:
c — 152.67 mm
ξο = 0.00 mm
ηο — 0.00 mm
Projection centre Ο:
/ 362530.603 \
X 0 [m] = 61215.834
\ 2005.742 J
Rotation matrix:
/ -0.034091 0.999407 0.004822 \
R = -0.999419 -0.034096 0.000621
V 0.000784 -0.004798 0.999988 J
Object coordinates of two points:
/ 363552.124 \ / 362571.087'
Pi [m] : 61488.048 P2 [m] : 61198.320
\ 588.079/ \ 596.670
To find: image coordinates of both points
Solution with Equations (2.1-19):
D/[ , f ζ — -33.288 λ p/r , ί ξ = 1.628

P\ [ m m ] : U = 110.074) :
L = 5.182
2.1.4 Central projection and projective transformation of a plane
Without restricting the generality of the statements, one can consider all the object
points to lie in a plane (Z = 0 in Figure 2.1-8)9. Equations (2.1-20) then read as
follows:
= αιξ + α2η + ä 3
ci£ + ο2η + c3
(2.1-23)
Υ _ + b2T? + h
διξ + δ2η + c3
9
One finds the derivation for a sloping plane, which leads to the same result, in Mikhail, E., Bethel,
J., McGlone, C.: Modern Photogrammetry. John Wiley & Sons, 2001.
The coefficients ait bi and c.t are related to the parameters of Equations (2.1-20) as
follows:
ä\ — Xorn - Z0rn
02 = Χθ?32 ~ ZqT\2
Dividing the numerators and denominators of Equations (2.1-23) by C3, we obtain the
following expressions for the relationship between image ξ,η and object coordinates
X,Y:
_ α\ξ + α2η + a^
+ 02η+ I
(2.1-24)
ci£ + 02η + 1
From these equations it follows that:
• a single picture suffices for the reconstruction of a plane object.
• eight independent parameters define the central projection of a plane object.
The reduction in the number of independent parameters from nine to eight may, at
first sight, seem surprising. It arises from the fact that, in the case of a plane object,
relationships exist among the original nine elements. In the special case when the
photograph and the plane object are parallel, it is easy to see (Figure 2.1-9) that Z0 and
c are no longer independent of each other; we need only to know the ratio ZQ JC.
Figure 2.1-9: Two geometrically identical metric images with different values of Zq
and c, but with the same value of the ratio Z 0 / c
We now turn to the question of how the eight parameters can be found in the general
case and how other details may be determined from the photographs. Supposing that
we have four control points10 (image and object coordinates known), then the eight co-
efficients can first be determined from Equations (2.1-24). Then the object coordinates
Xi and Yi of each new point Pi can be determined from its image coordinates and
m-
Numerical Example. We are given both the image coordinates and the object coordi-
nates of four control points A, B, C, D and the image coordinates of P, one of many
new points for which the object coordinates are to be found.
Image coordinates Object coordinates

ξ [mm] η X [ m] Y
A -33.288 110.074 1488.05 3552.12
Β 32.183 101.785 2229.38 3507.46
C -45.762 -74.337 1376.40 1899.76
D 28.472 -96.643 2086.48 1600.12
Ρ 1.628 5.182 ? ?
After multiplication by the denominator from Equations (2.1-24) we obtain eight si-
multaneous linear equations of the form:
ξαι + ηα2 + α3 - Χξο\ - Χηο2 = Χ

ξ6ι + φ2 + b3 - Y^ci - Υηο2 = Υ
/ -0.033288 0.110074 1 0 0 0 49.534 -163.796 \ / 0.1 \ ( 1488.05 \

0 0 0 -0.033288 0.110074 1 118.243 -390.996 αι 3552.12
0.032183 0.101785 1 0 0 0 -71.748 -226.917 a3 2229.38
0 0 0 0.032183 0.101785 1 -112.881 -357.007 bi 3507.46
-0.045762 -0.074337 1 0 0 0 62.987 102.317 ί>2 1376.40
0 0 0 -0.045762 -0.074337 1 86.937 141.222 b3 1899.76
0.028472 -0.096643 1 0 0 0 -59.406 201.644 Cl 2086.48
V 0 0 0 0.028472 -0.096643 1 -45.559 154.640 / \C2 J \ 1600.12 /
The solution of these equations is:
ai = 8021.065 6, = -4066.292 Cl = -1.330

a2 = -1084.217 b2 = 7360.815 c2 = -0.728
a 3 = 1821.069 b3 = 2479.221
From Equations (2.1-24) the object coordinates of Ρ are:
_ 8021.065 χ 0.001628 - 1084.217 χ 0.005182 + 1821.069

-1.330 χ 0.001623 - 0.728 χ 0.005182 + 1
= 1839.43 m
_ -4066.292 χ 0.001628 + 7360.815 χ 0.005182 + 2479.221
-1.330 χ 0.001628 - 0.728 χ 0.005182 + 1
= 2525.74 m
10
If the interior orientation is known, three control points suffice (see Section 4.2.1).
In the special case where the image and object planes are parallel (that is to say, ω
ψ = 0), the three-dimensional rotation matrix takes the form:
cos κ — sin κ 0 '

R = I sin κ cos κ 0
0 0 1
Inserting R in Equation (2.1-20) gives:
Zo
X = X0 + — (cos κ (ξ - ξ0) - sin κ(η - η0))
Zo
Υ = Yq + — (sin«(ξ - ft) +cosk(t? - η0))

Xo cos κ — sin κ
mB (2.1-25)
Writing this in matrix notation sin κthecos
YQand introducing κ
quantity mV-Vo
B = Zo/c we arrive at:
Taking Section 2.1.1 into account, one reaches the conclusion that in this special case,
the photographic image is geometrically equivalent to a map; it is only an object plane
reduced in scale (as well as rotated and translated). Equation (2.1-25) is a plane simi-
larity transformation (2.1-9). The image scale, or the map scale 11 , is 1 : where:
Zo
mB = (2.1-26)
Equation (2.1-26) can also be derived geometri- ___

c
cally (Figure 2.1-10): j_
If we assume not only that the image plane is par-

allel to the X Y plane of the object coordinate sys-
z
tem but also that °
K = X0 = Y0= ξ0 = ηq = 0
then
s c 1
— = — — const. = Figure 2.1-10: Image as map
S Z0 rnB
Example. Camera axis perpendicular to the fa£ade (c = 157.65 mm, Zo = 50.4 m).
See Figure 2.1-11.
11
A scale is defined by the ratio 1 : m e - We call m e the scale number.
Figure 2.1-11: Image scale = map scale
Exercise 2.1-7. What would be the maximum error to be found in this "photographic
map" (an orthophoto) if an element of the fagade projected 50 cm from the plane de-
fined as ZQ — 50.4 m? (For the solution see Equation (7.2-1) or Exercise 7.2-1.)
Comment: With the expression (2.1-24), which portrays the relationship between XY
coordinates in the object plane and ξη coordinates in the image plane, we depart from
the idea of a central perspective bundle of rays. In place of the central perspective
bundle of rays with its nine elements of interior and exterior orientation, projective
geometry provides the mathematical relationship between a plane object and its image.
(Employing projective geometry one can write the mathematical relationship between
an object (a plane or a straight line) and its image as created by central perspective
without making use of the position and orientation of the image in relation to the ob-
ject.)
Projective geometry is an alternative to central perspective with the following charac-
teristics:
• the mathematical relationship is linear as opposed to the non-linear bundle Equa-

tion (2.1-19).
• as a result, no initial values are necessary when, for example, the transformation
parameters have to be computed by means of control points.
• the elements of interior orientation are not necessary; this means that projective
geometry is very suitable for non-metric images, such as amateur pictures.
• even if the elements of interior orientation are known, as with metric images, one
cannot make use of them, except indirectly.
• the affine transformation, the characteristics of which are enumerated in Sec-

tion 2.1.1, is included as a special case in the projective transformation Equa-
tions (2.1-24). (If one puts c\ = C2 = 0, one obtains the affine transforma-
tion with its six parameters. Against that, central perspective (collinearity) equa-
tions (2.1-19) or (2.1-20) cannot be applied to images which have been produced
by means of an affine process.)
Exercise 2.1-8. Given four control points in the ξη image system (1(0,0), 2(1,0),
3(0,1), 4(1,1)) and in the X Y object system (1(0,0), 2(2,0), 3(0,4), 4(2,4)), as
well as the ξη coordinates of a new point 5(0.5,0.5), find the eight parameters of the
projective transformation using Equations (2.1-24) and also the object coordinates of
the new point. (Solution: αϊ = 2, ü2 = 0, 03 = 0, &ι = 0, 62 = 4, 63 = 0, c\ = 0,
C2 = 0; X Y coordinates of the new point, 5, are (1,2). Comment: The projective
transformation (2.1-24) even deals with this typical affine deformation.)
2.1.5 Central projection and projective transformation of the straight

line
Without loss of generality, the X axis (Y = 0) and the ξ axis (η — 0) can be adopted
as the corresponding straight lines in the object plane and the image plane respectively.
From the first of the two Equations (2.1-24) we then obtain the following equation:
χ = 0^+03
+ 1
The three coefficients 01, <23 and ci describe the central projection of a straight line.
They can be determined from three control points; every other point on the line in the
image can subsequently be transformed into the object line.
Equation (2.1-27) evolves from the projective transformation (2.1-24) in which the
elements of interior and exterior orientation are unknown. In projective geometry one
may prefer to use the cross-ratio, rather than Equation (2.1-27); the cross-ratio for four
collinear points is invariant under projective transformation and, it goes without saying,
under central projection (see the second solution of the exercise below for a definition
of the cross-ratio).
Example (of the reconstruction of a straight line in the object space). We are given a
photograph of a street with lane markings in a straight line APBQC (Figure 2.1-12).
The lane marking AB is 4.50m in length and the separation BC is 5.00m. Points Ρ
and Q on the skid marks are to be reconstructed.
1 st solution (using Equation (2.1-27))
Image and object coordinates of the three control points A, Β and C:
ξ [mm] X[m]
A 0.00 0.00
Β 38.40 4.50
C 68.30 9.50
Figure 2 . 1 - 1 2 : Straight lines in the image
ojO + a3
( 2 . 1 - 2 7 ) for A: 0 =
ci0+ 1 a3 = 0
α ι 3 8 . 4 0 + α3
( 2 . 1 - 2 7 ) for B: 4.50 = ai = 0.09747
ci 3 8 . 4 0 + 1
ci = - 0 . 0 0 4 3 8
α ι 6 8 . 3 0 + ο3
( 2 . 1 - 2 7 ) for C: 9.50 =
ci68.30 + 1
Substitution of the image coordinates ξρ and in Equation ( 2 . 1 - 2 7 ) gives Xp and

Xq'·
ξ [mm] X [ m ]
Ρ 33.50 3.83
Q 44.90 5.45
2 nd solution (using the cross-ratio)

AB AC _
~PB'^C~ P'B' ' P'C'
4.50 9.50 38.40 68.30
Ρ Β = 0.67
PB 5.00 + P B 4.90 34.80
AB AC _ A'B' A'C'
QB QC~
:
~ΟΒ' ' ~QC'
4.50 9.50 38.40 68.30
QB = 0 . 9 5
QB 5 . 0 0 - QB 6.50 23.40
Exercise 2.1-9. Consider the solution if, instead of point C, the vanishing point F ,
with coordinate ξρ — 2 2 1 . 0 mm is given; that is, Xf = 00. (Hint: X F must be infinite
in Equation (2.1-27), as is achieved when 2 2 1 . 0 c i + 1 = 0 . ) (Solution: Xp = 3 . 8 2 m,
X Q = 5 . 4 6 m.)
Exercise 2.1-10. Show that for four points the four expressions (2.1 -27) can be inserted
in the cross-ratio formulation. (Hint: Choose the origin of coordinates for both X and
ξ in the point A: aj = 0.)
2.1.6 Processing a stereopair in the "normal case"

Photogrammetry is used above all for the reconstruction of three-dimensional objects
from metric images. In Section 2.1.3 we reached the conclusion that two photographs
of the same object are necessary for this. The analysis proves to be especially simple
if both camera axes are normal to the base (the base is the straight line between the
two perspective centres) and parallel to each other (Figure 2.1-13). This is known
as the "normal case". This condition is very difficult to achieve in the case of aerial
photographs although one seeks to approximate it as closely as possible.
Xo\ = Yoi — Υθ2= Ζοι = ζ02= 0

X-02= Β
ξοι = ηοι = ξθ2 = = 0
ωι = UJ2 = Ψ\ = Ψ2 = Κι = Κζ
In the "normal case" the three-dimensional rotation matrix R (2.1-13) becomes the unit
matrix for both pictures:
1 oo1
R= I 0 10
0 0 1
In the "normal case" the relationship between image and object coordinates as formu-
lated in Equations (2.1-20) simplifies as follows:
Image 1
X = Z— (2.1-28)
—c
Y = Z— (2.1-29)
Image 2
X = B + Z^- (2.1-30)
—c
Y = Z— (2.1-31)
—c
Equations (2.1-29) and (2.1-31) imply that:
Vi = V2 => Vi ~ V2 — Ρη = 0 (no 77-parallaxes)
p
Γ i*
Figure 2.1-14: Two metric images without η-parallaxes
The final formulae for the calculation of the object coordinates Χ, Υ, Ζ from the
image coordinates ξ and η follow from Equations (2.1-28) to (2.1-31). We begin with
£1 £2
the Equations (2.1-28) and (2.1-30); that i s , - Z — = Β - Z - :
c c
cB
Ζ - - —
6 - 6 Ρς
Y = -Z^- = (check) (2.1-32)

c c
X= -Z^
c
The difference £1 — £2 — Ρς (£-parallax) can be measured directly in some photogram-
metric instruments (Section 6.4.1); in others the original image coordinates £t and £2
are measured and the difference ρζ computed. Not only can the Formulae (2.1-32) be
derived from Equations (2.1-20) but also, more simply, using ratios directly obvious
from Figure 2.1-15.
Figure 2.1-15: Geometrical derivation of the relationships (2.1-32)
Exercise 2.1-11. A statue is photographed with a stereometric camera (Section 3.8.2)

conforming exactly to the "normal case" with base Β — 1.20 m and principal distance
c = 64.20 mm. The values in the table below are measured in the image for points
Ρ and Q. Calculate the length of the line-segment PQ in object space. (Answer:
2.655 m.)
6 [mm] 771 [mm] Pi mm]

Ρ -3.624 34.202 18.321
Q 29.876 14.809 16.983
2.1.7 Error theory for the "normal case"

Formulae (2.1-32) describe how we arrive at the object coordinates X, Y, and Ζ from
the quantities which can be measured in the picture, ξι, η\ and £2 or, in other cases,
Ρζ (the image coordinate ryi is usually also measured and provides a check). In this
section we enquire into the accuracy of these indirectly acquired object coordinates. In
so doing we assume that the values for both the principal distance c and the base Β are
without error.
From the first relationship in the Equations (2.1-32) one obtains the following expres-
sion for the mean-square error σζ·
cB Ζ Ζ
σζ — -yCp, = — (2.1-33)
pi 4
cΒ κ
The ratio Β / Ζ is known as the base/distance ratio (or in the case of aerial photographs
the base/height ratio). The ratio Z/c is called, from Equation (2.1-26), the photo-scale
number TUB although only in the special case of parallelism between object and im-
age planes can one speak with justification of a scale number for the whole picture
(Section 2.1.4). The mean-square errors σχ and ay of the X and Y coordinates are
then derived using the rules of error propagation applied to the corresponding Equa-
tions (2.1-32).
Ζ Ζ2
σζ = ™ΒβσΡξ =
σ γ
= ]J + (§'ση) = +(τηΒση) 2
(2.1-34)
σχ = \l{^mBBap() +
(msc7«)2
Numerical Example.
Given 12 : Image coordinates ξ\ = ηι = 50 mm ± 7 μπι
Accuracy of measured parallaxes σΡί = ± 5 μπι
Principal distance c = 150 mm
To be found: the root mean square errors of the object coordinates as functions of
the photo scale number and the base/distance ratio Β / Ζ . (Take note of the units of
measurement shown in the last column.)
B/Z = 1 : 1 Β/Ζ = 1 : 3 Β/Ζ = 1 : 10 Β/Ζ = 1 : 20

mB σχ,γ σζ σχ,γ σζ σχ,γ σζ σχ,γ σζ Units
50000 0.36 0.25 0.43 0.75 0.90 2.50 1.70 5.00 m
10000 0.72 0.50 0.86 1.50 1.81 5.00 3.41 10.00 dm
1000 0.72 0.50 0.86 1.50 1.81 5.00 3.41 10.00 cm
100 0.72 0.50 0.86 1.50 1.81 5.00 3.41 10.00 mm
25 0.18 0.13 0.22 0.38 0.45 1.25 0.85 2.50 mm
Table 2.1-1: Accuracy of photogrammetry as a function of photo scale 1 : and of

base-distance ratio Β jZ
Using this table and the Formulae (2.1-34) we can make the following generalized
statements concerning photogrammetric accuracy:
• assuming a constant base/distance ratio, the root mean square errors in all three
coordinates are directly proportional to the photo scale number. Any desired
accuracy can, therefore, be achieved by means of an appropriate choice of photo
scale.
12
Measurement accuracy will be more closely considered in Sections 4.6 and 6.1.1.
Section 2.2 Preliminary remarks on the digital processing of images 35
• for a constant photo scale, the root mean square error in the Ζ coordinate is in-
versely proportional to the base/distance ratio. The root mean square error in the
XY coordinates, however, increases only slowly as the base/distance ratio de-
creases. If the base ratio is somewhat less than 1 : 1, all three object coordinates
will be equally accurate.
• with a constant base, the root mean square error in the Ζ coordinate increases as
the square of the distance, Z, from the camera.
For very rough estimation of accuracy:
• for the moment, ignore the first term in the expressions for σχ and σγ from
(2.1-34) and
• replace the accuracies σξ, ση and σΡ( with a generalized accuracy figure for im-
age measurement σ Β ·
As a result we derive the following easily remembered rules of thumb:
Ζ
σζ — τηΒ~σΒ
B
(2.1-35)
σχ = σγ = τηΒσΒ
Exercise 2.1-12. Using a metric camera and the "normal" disposition of photographs a
land-slide is to be monitored from a slope on the opposite side of a valley. An accuracy
of σχ = σγ = σζ = ±10 cm is required. What is the greatest distance from which
this land-slide can be monitored (measurement accuracy σΒ = ±6μιυ)? What base
length should be chosen? Hint: One should use the Formulae (2.1-35). (Answer:
Ζ — Β — 1666 m.) Supplementary exercise: How much can the base be reduced for
areas at half the distance? (Answer: Β = 416 m, Β/Ζ = 1 : 2; in this case the root
mean square error in the X and Y coordinates will be smaller than that demanded:
σχ = σγ = ±5 cm.)
Exercise 2.1-13. What is the root mean square error in the length of the line-segment
PQ of Exercise 2.1-11? The values in the above table should be used for σξ, ση and
σρ . (Answer: 2.655 m ± 1.0 mm.)
2.2 Preliminary remarks on the digital processing of images

The mathematical methods described in Section 2.1 are also, of course, necessary in
digital photogrammetry. In addition one requires some procedures from digital image
processing. In this section we cover some aspects of digital image processing which
are of interest in photogrammetry.
2.2.1 The digital image

An analogue photograph appears in a light-sensitive coating on a (tough) supporting
layer. (In analogue or analytical photogrammetry the supporting layer is usually film.)
A digital photograph is recorded by electronic means. 13
In analogue pictures geometrical shapes such as points, straight lines and so on are
interpreted visually and are defined in an abstract way, as is customary in analytical
geometry. The left of Figure 2.2-1 shows a straight line and six points as represented
in analytical geometry or in an illustration. In Sections 2.1.3-2.1.6 everything is dealt
with on the basis of such geometry.
In digital images there are no abstract points and lines. Picture elements, usually
square, take the place of image points. Instead of a straight line, which in analytical
geometry runs from a point, we find a stepwise string of adjacent picture elements.
The usual word for picture element is pixel.
The right hand side of Figure 2.2-1 shows a digital image corresponding to the six
points A-F and the straight line G. A digital image can be formed such that all pixels
which lie on points and lines receive a value which stands out against the value of the
background pixels. The two images in Figure 2.2-1 approach each other more closely
the smaller the size of the picture element becomes; that is, the better the geometrical
resolution of the digital image. The high degree of abstraction of analytical geometry
is, however, never achieved in pixel geometry. 14
D.'E
C
' . F
A· B·
Figure 2.2-1: Analytical geometry (left) and pixel geometry (right)
A few special features of pixel geometry should be touched on with the help of Fig-
ure 2.2-1:
a) the two separate points D and Ε merge together in the digital image.
13
For that reason digital photogrammetry is also known in English as "soft copy photogrammetry"
as opposed to "hard copy photogrammetry" which uses photography recorded on film (PE&RS 58(1),
pp. 49-115, 1992).
14
Limits to the numerical representation of numbers in a computer, e.g. the image coordinates ξ and
η, lead to similar (negative) effects in analytical photogrammetry as does the finite pixel size in digital
photogrammetry.
b) it is strictly necessary that the pixels belonging to one geometrical shape be spe-
cially labeled, for example with what is called a chain code; starting from the
position of the initial pixel this defines the next pixel by means of a pointer (right,
left, above, below, . . . ) . The criterion "all neighbouring pixels" is in fact inad-
equate. In pixel geometry, for example, two neighbouring points of the straight
line G are joined together just as much as the two separate points A and Β (Fig-
ure 2.2-1).
c) the formation of geometrical objects with the help only of neighbour relation-
ships, in this context better described as connectivity relationships, leads to dif-
ferent results, depending on the criterion used. 4-neighbourhoods (4N) and 8-
neighbourhoods (8N) are common in digital image processing (Figure 2.2-2).
1- s
Figure 2.2-2: 4N and 8N relationships
8N
Using a 4N the points Β and C remain as separate points; using an 8N all the
points A — F are combined. Incidentally, if the pixels representing the points A -
F were to be shifted one pixel to the right, then using an 8N connectivity all the
relevant pixels of Figure 2.2-1 would be assigned to a single geometrical object.
d) In analytical geometry, separations are usually determined by means of Eucli-

dean distance defined in the direction of the normal to the straight line. In pixel
geometry, on the other hand, distance measure is determined using criteria which
are indicated in Figure 2.2-2. In Figure 2.2-2 the Euclidean distance of point Β
from the straight line G is somewhat less than the corresponding distance of
point F ; that is to say, point F is further removed from the straight line than
point B. Application of an 8N, however, results in the pixel representing the
point F being closer to the string of pixels representing the straight line G than
the pixel representing point B 1 5 .
The picture elements arranged in a raster or a matrix carry information. Their value
range depends on the recording equipment and on the computer used. Very commonly
the values range between 0 and 255 which clearly exceeds the ability of the human
eye which can discriminate between about 50 different shades. Information with 256
different states can be represented by 8 bits (2 8 combinations of bits). A group of eight
bits is combined as a byte in most computers. In very sensitive recording equipment the
15
In this connection one also speaks of a Euclidean metric and of raster data metrics (see, for example,
Bill, R.: Grundlagen der Geo-Informationssysteme. Band 2, Wichmann, 1999, and the literature cited
therein for further reading).
image data may even be registered with 16 bits, that is with 2 bytes (216 = 65536 com-
binations).
In information technology, besides binary representation, one also comes across octal
and hexadecimal forms. The table below sets out pixel values in the different notations
opposite each other:
Decimal Octal Hexadecimal Binary bit pattern

0 000 00 00000000
1 001 01 00000001
2 002 02 00000010
127 177 7F 01111111

128 200 80 10000000
254 376 FE 11111110

255 377 FF 11111111
Table 2.2-1: Different representations of pixel values
In normal black and white images the pixel values are known as grey values (usually
black is coded as 0 and white as 255). In binary images there are only two grey values;
the value zero can represent the extraneous background and the value one the signif-
icant image information. The right hand side of Figure 2.2-1 shows a binary image.
Colour images have three spectral layers which are registered on three image matri-
ces of equal size. These images are usually represented by 24 bits (eight bits for each
colour layer with Red, Green and Blue as the primary colours for a so-called RGB
reproduction).
Exercise 2.2-1. In Figure 2.2-1 place a line below and parallel to the existing line
such that in the corresponding digital picture it does not touch the string of pixels
representing the straight line G. Avoidance of contact should be established once in a
4N pattern and once in an 8N pattern.
2.2.2 A digital metric picture

A digital metric image is a digital photograph which meets the requirements of pho-
togrammetry. The basic prerequisites are that the image should have been formed by
central projection and that the perspective centre should be fixed in relation to the im-
age. Such a metric image was defined in Section 2.1.3 in which the analogue image
was discussed. We now turn to the digital metric image.
The digital metric image is formed in the image plane of a digital (metric) camera. The
digital metric image consists of a (two-dimensional) matrix G with picture elements
gij (Figure 2.2-3). The row index i runs from 1 in steps of 1 to I, the corresponding
column index j from 1 to J. The picture element dimensions are Δξ χ Αη.
112 η
9u
PP
1 ^ j J
Δη
Figure 2.2-3: Digital metric image
In a metric digital image a relationship is required between pixel position and ξη im-
age coordinate values. In Figure 2.2-3 we have introduced an image coordinate system
lying half a pixel width outside the image matrix and rotated 100 gon ( = 90°) clock-
wise from our accustomed image coordinate system. In a digital metric camera no
special fiducial marks are necessary to define the image coordinate system as are re-
quired in a metric film camera.
Multiplication of the index i of the matrix by Αξ gives the image coordinate ξ of the
mid-point of the pixel gij; correspondingly the index j multiplied by Αη gives the im-
age coordinate η. The conventional measurement of image coordinates in analogue
photographs is replaced in digital photographs by location of the respective pixel. Lo-
cation of the pixel by a human operator is completed with a mouse click. If possible,
however, pixel identification and measurement should be automatic—even into the sub-
pixel range; that is to say, with accuracy better than the pixel dimensions.
For photogrammetric processing of digital photographs one requires the interior ori-
entation exactly as with analogue photographs. In Figure 2.2-3 the position of the
principal point PP is given in the ξη image coordinate system. In the case of appropri-
ately small pixels it is adequate just to know the pixel in which the principal point lies.
One can continue with this idea by regarding the indices i and j directly as the image
coordinates ξ and η. In this case, assuming square pixels, the principal distance c is in-
troduced in units of Δξ(= Αη). It is easy to see that the collinearity equations (2.1-20)
and the Equations (2.1-32) relating to the "normal case" retain their validity despite
these unfamiliar units of measurement in both image and camera space.
2.2.3 Digital processing in the "normal case" and digital projective

rectification
After finding homologous points in two digital metric images which conform to the
"normal case", and using the relationships (2.1-32), the object coordinates Χ , Y and Ζ
of individual points can be found from the image coordinates, the base (which is part
of the exterior orientation) and the interior orientation. (Note: the image coordinates
in (2.1-32) are referred to the principal point P P as origin.) It should be remembered
that, in the processing of a "normal" image pair as outlined, the same mathematical
relationships are called upon in the case of digital metric images as in the analogue
case. In the processing of digital image pairs the identification of homologous points
in the two photographs invites complete automation. Not until Section 6.8.3 do we
approach this subject.
In this section we turn to the evaluation of an individual digital photograph of a plane

object. We have already learnt, in Section 2.1.4, not only the mathematical relation-
ship between the image plane and the object plane but also the analytical procedure.
Four control points are necessary for the determination of eight independent parame-
ters; after the solution of the relevant equations the object coordinates X and Y can
be calculated for new points from their measured image coordinates ξ and η (see the
numerical example in Section 2.1.4).
With the support of the same mathematics we can also, as we shall shortly see, transfer
the whole digital photograph into the object plane. Since projective geometry comes
into use (see the comment at the end of Section 2.1.4), in this case we speak of projec-
tive rectification. The result is a digital orthophoto, therefore a geometrically correct
(undistorted) digital photograph. The conversion of the distorted digital image (the
original picture) into the digital orthophotograph, which is in the (two-dimensional)
object coordinate system, is done with the help of Equations (2.1-24). This conversion,
typical of operations with digital images, will be clarified in what follows.
Because of their arrangement within the image matrix, we know the ξη coordinates
of the mid-points of all the pixels of the original distorted image. In Figure 2.2-4
these mid-points are labeled with small black-filled circles. We find the corresponding
positions in the object coordinate system using Equations (2.1-24); in Figure 2.2-4 this
is signified by the abbreviated notation (Χ, Υ) = / ( ξ , η). Naturally, on account of the
distortion of the original image, the X Y positions found for all the original pixels do
not fall in an orthogonal raster. The required mid-points of the pixels in the resultant
image, the digital orthophoto, with their indices i and j, are shown in Figure 2.2-4 with
small open circles. We have a set (a sample) of arbitrarily arranged points which are to
be brought into a new pattern, an orthogonal raster; in digital image processing this is
known as resampling.
Nearest neighbourhood assignment of grey values
One of the possible techniques for assigning the grey value within such a resampling
is the nearest neighbour method. In this technique the grey value of the original pixel
is given to that position in the new image matrix which lies nearest, as measured by
coordinate differences, to the transformed mid-point of the pixel. In the example of
Figure 2.2-4 the grey value <72,4 of the original image will be brought to position (i,j) =
(2,3), that of <?3i4 to position (i,j) — (3,2), and so on.
direct
η • J
(Χ,Υ) = ί(ξ,η)
0 0 0 0 0
0 0 0 0 0
0 > 0 0 0
0 0 0 0 0
0 0 0 0 0
-•X
Figure 2.2-4: Direct rearrangement
The rectification outlined in Figure 2.2-4 is called a direct rectification or, in a general
form, as a direct rearrangement. It has the disadvantage among other things that it can
lead to gaps in the resulting digital image. The much more widely used technique is
that of indirect rectification or rearrangement. It is summarized in Figure 2.2-5.
indirect
(ξ,η) = ηχ,Υ)
• • • • 1- Ο 0 0 0 0
0 0 0 0
• • • cr •
O*• 0 0 0
• • •
Υ* 0 0 ND 0 0
• • • • 0 0 0 0 0
-•x
Figure 2.2-5: Indirect rearrangement
It starts with an originally empty image matrix with indices i and j in the object coor-
dinate system. In general the number of pixels in this image matrix will be somewhat
larger than that in the original image. Subsequently the mid-points of these pixels are
transformed into the original image using the inverse transformation, which we go into
more closely at the end of this section. The grey value at that position in the origi-
nal image can be found using, for example, the nearest neighbour method and can be
copied into the new image, the digital orthophotograph. For example, in position i = 3,
j = 3 of the resulting image of Figure 2.2-5, the grey value computed at the inverse
point shown within the cell ξ = 2, η = 3 will be inserted; likewise in position i = 4,
j = 3, the grey value computed in ξ = 3, η = 4 will be inserted, and so on.
Bilinear interpolation of grey values
A disadvantage of the nearest neighbour method is that in the most unfavourable case
the grey values assigned may be displaced from their correct positions by as much as
half a pixel. These displacements may lead to a shift, in the worst case, of up to one
pixel of the transformed image; in general it leads to a sawtooth effect and to variable
widths of bands of pixels representing lines.
An interesting alternative to nearest neighbour is bilinear interpolation, in which case
the desired grey value is found, as a function of ξη coordinates, from the four neigh-
bouring values g\, <72, <73 and 94 using bilinear interpolation (Figure 2.2-6).
1 Λ
Figure 2.2-6: Bilinear grey value interpolation
The equation for bilinear interpolation is:
9 = 9(ξ, V) = αο + αχξ + α2η + α3ξη (2.2-1)

If one were to consider the grey value as height above the ξη plane, then Equation
(2.2-1) would describe a surface of which the height would be linear with respect to
ξ for a fixed η and vice versa. Such a surface is known as a hyperbolic paraboloid.
We introduce local coordinates ξ'η' referred to the upper left-hand corner of the square
comprised of the mid-points of the four neighbouring pixels of the original image. The
determination of the parameters ao, a\, and «3 from the four grey values g\, g2, <73
and (74 then becomes especially simple. If the pixel size is Δ (Figure 2.2-6), then:
(\ 0 0 0 \ ί a0\
92 1 Δ 0 0 αϊ
(2.2-2)
93 1 0 Δ 0 α2
\9iJ \ l AAA2 J \ °3 /
The simple structure of the coefficient matrix makes it easy to write a general solution:
f °o\ ( 1 0 0 0 \ fgi\
αϊ - 1 / Δ 1/Δ 0 0 92 (2.2-3)
a2 -1/Δ 0 1/Δ 0 93
\ «3 / ^ 1/Δ 2 - 1 / Δ 2 — 1/Δ 2
iMV w
Exercise 2.2-2. Check that the general solution given in (2.2-3) is correct. (The prod-
uct of the matrices of (2.2-2) and (2.2-3) should be the unit matrix.)
Substituting expressions for ao, a\, ai and aj from Equation (2.2-3) in Equation (2.2-1)
leads to the desired interpolation formula:
cv
92+
Δ2
(2.2-4)
+ 1 9 3 +
Δ " W ) Δ^54
Numerical Example. We choose Δ = 1; ξ' and η' vary be-

160 140 tween 0 and 1. With respect to an origin at the upper left-
hand corner, the coordinates of the point whose grey value we
Ο wish to interpolate are ξ' = 0.75 and η' — 0.25. It follows
from Equation (2.2-4) that: 5(0.75,0.25) = 0.1875 χ 160 +
180 150 0.5625 χ 180 + 0.0625 χ 140 + 0.1875 χ 150 = 168
Exercise 2.2-3. Check the result found in this numerical example by means of linear
interpolation along lines parallel to the sides of the square and passing through the
interpolated point.
Bilinear interpolation involves more computation than nearest neighbour but has the
advantage that no shifts appear. The contrast in the original image is, however, some-
what reduced. If this attenuation in contrast is to be avoided one should move to higher
order interpolation including, for example, 16 pixels 16 .
Inverse transformation equations for indirect projective rectification
The inverse transformation equations (ξ, η) — / " ' (Χ, Y), which are necessary for in-
direct transformation (Figure 2.2-5), are still to be formulated. In the projective trans-
formation, from a mathematical view point, neither image plane nor object plane takes
precedence over the other. That is to say, in the Equations (2.1-24) ξ and X as well as
η and Y can be interchanged. The eight parameters for the inverse transformation can
16
See, for example: Luhmann, T., Robson, S., Kyle, S., Harley, I.: Close Range Photogrammetry.
Whittles Publishing, 2006.
therefore be found in the same way from eight linear equations, assuming four control
points.
It is also possible, however, to formulate the inverse transformation equations for in-
direct rectification in terms of the eight parameters of the direct projective transfor-
mation, (Χ,Υ) — / ( ξ , η), Equations (2.1-24). After much re-arrangement of Equa-
tions (2.1-24) one obtains:
^ _ (&2 ~ c2b3)X + (a 3 c 2 - a2)Y + (α2&3 ~ aib2)

(&1C2 - b2ci)X + (a2ci - aic2)Y + (α^2 - a2b\) ^ ^^
_ (63C1 - bi)X + (ai - a3C])y + (0361 - 0163)
^ (b\c2 - b2c\)X + (a2c\ - ajc2)Y + {a\b2 - a2b\)
Exercise 2.2-4. Starting with Equations (2.1-24) derive Equations (2.2-5). (A very
elegant solution is possible using homogeneous coordinates. (Appendix 2.2-1).
Numerical Example. Let us take up once more the numerical example of Section
2.1.4. Equations (2.2-5) give the following inverse transformation equations, in which
both denominator and numerator have been divided by (a\b 2 — a2b])\
0.00016776X - 0.00000441 Y - 0.2945585

ξ =
0.0002333IX + 0.00013323Y - 1
η =
0.00001409X + 0.00019114Y - 0.4995346
0.0002333 I X + 0.00013323y - 1
If one exchanges the ξη coordinates and the XY coordinates in Equations (2.1-24)
and solves the corresponding linear equations as shown in Section 2.1.4, one arrives at
identical transformation parameters. With object coordinates X = 1839.43 m and Y —
2525.74m for the point P , it emerges that the image coordinates are ξ — 1.628 mm
and η = 5.182 mm as already known from Section 2.1.4.
A practical example of digital projective rectification
Figure 2.2-7 is an oblique picture of a fagade, which can be regarded as a plane. The
photograph was taken with a Kodak DCS 460c digital camera (Section Ε 3.5, Vol-
ume 2). Coordinates were known for four control points in the object plane; their image
coordinates were found after identification of the corresponding points in the original
image matrix. The transformation parameters for a projective rectification were subse-
quently evaluated and the whole of the picture was digitally rectified, the result being
shown in Figure 2.2-8. The size of the pixels at object scale is 2 cm χ 2 cm.
Instead of the projective transformation of an arbitrary quadrangle, it seems at first
glance that affine transformation of two triangles might offer an alternative. The two
triangles to be transformed in this way are reproduced in Figure 2.2-7. Figure 2.2-9
illustrates the error occurring in the case of the affine transformation. This error can be
estimated with the help of Equation (2.1-27) or of the cross-ratio. It is that much larger,
the larger the triangle chosen and the larger the tilt of the picture (see Section 7.3.2d).
Figure 2.2-7: Oblique digital photograph
Figure 2.2-8: Rectangular section of the fa9ade projectively rectified
Figure 2.2-9: Affine transformation of two triangles

Exercise 2.2-5. In the above example estimate the error of the affine transformation
along the diagonal running from top left to bottom right. Hint: Choose a point in
Figures 2.2-7 and 2.2-8 somewhere in the middle of the diagonal. One then has three
points on a straight line with known image and object coordinate ξ and X, respectively;
using the method of Section 2.1.5 one can state the projective transformation relation-
ship for points on this diagonal. One then transforms the midpoint of this diagonal into
the object plane and compares its X coordinate with that of the midpoint in the object
plane. The difference can be verified as the error arising as a result of moving from a
projective to an affine transformation (Figure 2.2-8 to Figure 2.2-9).
Chapter 3
Photogrammetric recording systems

and their application
Metric images are produced by a metric camera which lies at the centre of Section 3.1.
If the photons arriving at the image plane of a metric camera are recorded by a chemical
sensor, then an analogue metric image is produced. This analogue technology is dis-
cussed in Section 3.2. If the photons in the image plane are recorded by an electronic
sensor, then a digital metric image is obtained. Digital image recording is handled in
Section 3.3. Section 3.4 describes a hybrid technique. It starts with an analogue film in
the camera; a subsequent digitization of the film negative or corresponding photograph
also results in a digital metric image.
3.1 The basics of metric cameras

A metric photograph has been defined so far as an exact central projection (Sections
2.1.3 and 2.2.2) in which the perspective centre is at a distance c from the principal
point of the photograph. The parameters of this simplified mathematical-geometric
model, namely the principal distance c and the image coordinates ξο and ηο of the prin-
cipal point (PP) of the photograph, are defined as the elements of interior orientation.
This idealized model does not correspond exactly to reality, however. The inevitable
errors of the lens, the camera and the photograph itself must be considered if the highest
accuracy is to be achieved.
3.1.1 The interior orientation of a metric camera

The geometric theory of optical systems postulates for a combination of lens elements
two principal planes Η, H ' (the object-space and image-space principal planes) in
which the one reproduces the other at a scale of 1 : 1 perpendicular to the axis. For
an optical system consisting of air-glass-air the two optical principal points, i.e. the
intersections of the principal planes with the optical axis OA, coincide with the two
nodal points Ν and N'. These are defined in such a way that the central rays, to be
discussed in more detail with the help of Figure 3.1-3, pass through the system without
deviation and form the same angles τ to the optical axis at Ν and N ' (Figure 3.1-1,
r ' = τ).
48 Chapter 3 Photogrammetric recording systems and their application
Optical
principal planes Image plane
Η Η'
Object space Image space P'
Image point
Optical axis N=0 PP

Principal point
Object point
p Object distance V Image distance
s e — s'=c
Figure 3.1-1: Idealized geometric image formation of an optical system. Ν, N ' =

nodal points = perspective centres O, O'
In this idealized case, Ν is the object-space perspective centre Ο, N ' is the correspond-
ing image-space perspective centre O' and the principal distance c is equal to the image
distance s'. The image distance s' and the object distance s are always measured from
the principal planes. In practical amateur photography, however, the distance to be set
on the focusing mechanism of the camera is D = (s + e + s'), i.e the distance of the
object from the image plane.
The optics of photogrammetric cameras are thick, usually asymmetric objectives. The
individual lenses are made from different types of glass so as to ensure that imaging
errors are corrected to the greatest possible extent. The aperture stop AS is usually
not in the centre of the objective (Figure 3.1-2). We must therefore pose the question:
where is the physical perspective centre?
All the rays from an object point that pass through the objective must pass through the
aperture. The apparent image of the aperture stop, as seen from the object, therefore
Optical axis
Aperture stop
Figure 3.1-2: Cross-section of a typical photogrammetric objective (Wild 21 NAg II,

//4)
Section 3.1 The basics of metric cameras 49
limits the effective bundle of rays forming the image point; this apparent image is
called the entrance pupil (EP). Its centre is the object-space perspective centre O. The
analogous exit pupil ( E P ' ) lies in the image space of the objective.
Exercise 3.1-1. How can the distance VO of the entrance pupil from the vertex V of
the objective be determined by theodolite?
EP AS
Figure 3.1-3: The definition of the centre of the entrance pupil Ε Ρ as the object-space
perspective centre O. Dot/dashed lines show central rays. V = vertex
The real photogrammetric objective and the idealized model of it described above (Fig-
ure 3.1-1) differ significantly:
a) the optical axis should contain the centres of all spherical lens surfaces. After the
cementing and assembly of all individual lenses and the mounting of the objec-
tive in the camera relative to the mechanical focal-plane frame, small errors will
inevitably have accumulated. The reference axis of photogrammetry is therefore
not the optical axis OA, but a calibrated, (i.e. standardized) principal ray PRA
which in object space is perpendicular to the image plane and passes through the
centre of the entrance pupil (Figure 3.1-4). Its physical extension intersects the
image plane in the principal point of autocollimation PPA (see explanation and
definition below).
b) the angles τ are defined at the centre of the entrance pupil and not at the nodal
points. Since the entrance pupil usually does not lie in the principal plane H, it
follows that τ ' is not equal to r.
c) the mechanically realized principal distance s'm defined by the focal-plane frame
of the camera differs slightly from the optical principal distance s' which pro-
vides the sharpest image.
d) the image plane is not rigorously perpendicular to the optical axis.
In essence, the result of these small errors is that the angles r ' in image space are not
equal to the angles r in object space. We therefore define a mathematical perspective
EP EP' IP
Figure 3.1-4: The definition of the image-space perspective centre 0'M. PRA = (auto-
collimation) principal ray, PPa = principal point of autocollimation in the image plane
IP, EP = entrance pupil, EP' = exit pupil, 0'P = physical perspective centre, c =
principal distance, ρ = image height = 1, Ap = (radial) optical distortion
centre 0'M which lies at a perpendicular distance c, the principal distance, from the
principal point of autocollimation PPa and which reproduces the angles τ as closely
as possible. Residual errors lead to optical distortions Ap.
The elements of interior orientation, so far defined as ξο, r?o and c, must therefore be
extended to include the radial optical distortion Ap:
Equation of interior orientation
ρ = c tan τ + Ap (3.1-1)
Photogrammetric cameras are mostly calibrated in a laboratory with the help of an op-
tical goniometer (Figure 3.1-5). Firstly, before mounting the camera in the instrument,
the observing telescope T\ is set in its zero position as defined by autocollimation with
the telescope T2. The camera is then mounted with the centre of the entrance pupil
E P , i.e. its object-space perspective centre O, in the axis of rotation and rotated about
E P until the mirror image of the cross-hairs of the telescope T2 is superimposed on
the cross-hairs (autocollimation). A flat glass plate with a small reflecting surface is
placed on the image plane (focal-plane frame) of the camera for this purpose. On the
side facing object space this plate also carries precise graduations. In the previously
defined zero position of telescope T\, the operator now observes the principal point of
autocollimation PPA which will be considered as the origin of the p-scale. The oper-
ator then points T\ to various graduations on the p-scale along each semi-diagonal of
the square image frame and observes the corresponding angles r . The radial optical
distortions can then be computed from the differences Ap — ρ - co tan r , where co is
the best known value of the principal distance (Figure 3.1-6).
Λ
Collimating telescope (fixed)
Τ
Reflecting flat glass plate
. with graduated surface
f" in image plane (p-scale)
Photogrammetric camera (fixed)
EP = rotation axis
of telescope Ti
x-measurements
Measuring telescope T,
Figure 3.1-5: Schematic diagram of a photo-goniometer

The results are usually asymmetric, the curves do not coincide. Amongst other reasons,
this asymmetry is caused by errors of centering of the individual lens elements. The
asymmetry can be greatly reduced by choosing another reference point, slightly dif-
ferent from the principal point of autocollimation PPA- The reference point resulting
from best symmetry is known as the principal point of best symmetry PPS.
The juxtaposition of two reference points, the principal point of autocollimation PPA
and principal point of best symmetry PPS, has the following practical consequences:
a) the distortions must be recalculated and given with respect to the PPs (Fig-
ure 3.1-7).
b) when correcting for distortion (Section 3.1.3), the PPs is used as reference point.
c) the mathematical perspective centre lies a distance c in front of the PPA', ξο and
ηο in the central projection equations (2.1-19) and (2.1-20) are the coordinates of
the PPA·
d) the rotations incorporated in the matrix elements r ^ of the central projection

equations (2.1-19) and (2.1-20) are with respect to the principal axis of autocol-
limation. (Its component in object space, determined by autocollimation (Fig-
ure 3.1-5), is perpendicular to the image plane, as required by the central projec-
tion (Figure 3.1-4).)
Figure 3.1-7: Radial distortion of the four semi-diagonals of Figure 3.1-6, referred to
PPs, and the resulting mean curve
The change of principal distance Ac is finally computed in such a way as to bring the
mean curve as close as possible to the p-axis (balanced radial distortion, Figure 3.1-8).
Exercise 3.1-2. How does a change of principal distance affect distortion? (Solution:
linearly dependent on p.)
The camera manufacturers are obliged to deal with radially symmetric distortion (e.g.
Figure 3.1-8). For the most demanding accuracies the radial distortion is separately
given for the 4 semi-diagonals, which effectively means details are given about radially
Δρ [μπι]
ΔρΓηαχ ' Δρ min

Ρ [mm]
30 60 90 -rar 150
Figure 3.1-8: Mean radial distortion with |Ap max | = |Δρ,^ η |
asymmetric distortion. In addition to radial distortion there also exists tangential dis-
tortion. This originates principally from the centering errors of individual lenses in the
objective. It is always an asymmetric distortion and is generally an order of magnitude
less than radial distortion.
Modern photogrammetric objective lenses have a radial distortion within ± 5 μηι, and
in film-based aerial metric cameras, radial distortion is, in fact, smaller than ± 3 μηι 1 .
Photogrammetric images taken with old objective lenses have radial distortion up to
30 μιη. Objectives which have not been specially developed for metric cameras can
display radial distortions of up to 100 μιη (Figure 3.1-9).
Figure 3.1-9: Distortion curve for Sonnar 4/150 objective from Rollei (focal length^
150 mm, smallest aperture (Section 3.1.4)= 4)
From the point of view of analytical and digital photogrammetry, larger optical dis-
tortions, if accurately known, are not a significant disadvantage. However, changes
in distortion values are certainly a problem since, in general, such changes cannot be
determined. These changes occur mainly in unstable cameras due to changes of focus,
vibrations and impacts, etc.
In a metric camera there must exist a coordinate system in which, amongst other para-
meters, the principal point is given. This photo or image coordinate system is realized
differently in analogue and digital cameras.
'E.g. Light, D.: PE&RS 58, pp. 185-188, 1992.

I·
A -»η
L« LJ
£ •»ξ
Π Π
Figure 3.1-10: Image coordinate system in an analogue metric camera (film camera,
left) and a digital metric camera (digital camera, right)
In analogue metric cameras which work with film, fiducial marks are located in the
image plane and are imaged on every photograph (Figure 3.1-10, left). With the help
of these fiducial marks, measurements taken in the image, and normally located in
an arbitrary comparator coordinate system, can be related to the camera used to take
them (Section 3.2.1). In digital metric cameras, the mathematical relationship between
image matrix and camera is never lost: the image points are measured within the image
matrix whose coordinate system is defined in the camera (Figures 2.2-3 and 3.1-10,
right).
The interior orientation of a metric camera is specified in a calibration certificate by
the following data. Note that there are differences between film cameras (F) and digital
cameras (D).
• F/D: date of calibration

• F/D: principal distance, c
• F: calibrated coordinates of the fiducial marks stated in a coordinate system
which, in principle, can have an arbitrary origin and rotation but in practice cor-
responds very closely to the ideal configuration (Figure 3.1-10, left)
• F: the principal point of autocollimation PPA, and the principal point of best
symmetry PPS, are given in this (arbitrary) coordinate system. In most metric
cameras, both these principal points and the fiducial centre FC (intersection of
connecting lines between fiducial marks, see Figure 3.1-10, left), lie within a
circle of radius < 0.02 mm.
• D: coordinates of principal point of autocollimation PPA and principal point of
best symmetry PPs are given in the ξη image coordinate system (Figure 3.1-10,
right).
• F/D: radial distortion, for example in the form of a piecewise linear function
(Figure 3.1-11). In higher quality cameras, the radial distortion is given on 4
semi-diagonals.
Δ ρ [μΓΠ]
Figure 3.1-11: Radial distortion of a 21 NAgll objective lens manufactured by Wild

(now Leica Geosystems)
• F/D: information on image quality (see Section 3.1.5)
• D: exact pixel sizes Αξ and Αη (Figure 2.2-3)
• D: for a 3-line camera (see Section 3.3.1) at least the ξ coordinates of the three
linear arrays
• F/D: type of light used in the laboratory calibration, since the wavelength influ-
ences the imaging properties of the optics and therefore also the distortion. (From
Figure 3.1-3 it can be seen that the aperture defines the perspective centre O; the
physical perspective centre plays a decisive role in the calibration, Figure 3.1-4.)
• F/D: the range used in the calibration, since focusing influences the imaging
geometry of the optics and therefore also the distortion
The last three mentioned items of information are of some importance to close-range
photogrammetry. Standard photogrammetric analysis methods do not require this in-
formation. A short selection of further reading related to the specialized methods in
close-range photogrammetry: Fräser, C., Shortis, M.: PE&RS 58, pp. 851-855, 1992.
Fryer, J., Brown, D.: PE&RS 52, pp. 51-58, 1986. Luhmann, T., Robson, S„ Kyle, S.,
Harley, I.: Close Range Photogrammetry. Whittles Publishing, 2006.
3.1.2 Calibration of metric cameras

In the previous section we learnt about laboratory calibration with a goniometer. This
is used for very high quality cameras in the expectation that the elements of interior
orientation remain unchanged over a long period of time.
Calibration can also be carried out by photographing a test field. Such test fields have
a relatively large number of control points with known X Y Z object coordinates. The
ξη photo coordinates of the control points are first measured. With this information,
and by means of the equations of central projection (2.1-19), both the elements of
exterior orientation and the elements of interior orientation can be calculated in a single
analysis. (For details, see Section 5.3.4). The system of equations (2.1-9) can also be
extended by polynomial coefficients in order to determine also the unknown optical
distortions (Section 3.1.3). When using test field calibration it is assumed that, between
making the test field exposures and recording the region of interest or object to be
measured, the elements of interior orientation remain unchanged.
In many photogrammetric tasks, a calibration of the camera during the execution of the
project is expected. The procedure in this case may be called self-calibration or on-the-
job calibration. Mathematically, self-calibration proceeds in the same way as a test field
calibration. Self-calibration merges together the elements of calibration and objection
reconstruction. The self-calibration method makes use not only of control points with
known X Y Z coordinates but also of the unknown target points of interest which appear
in multiple images. Self-calibration at close-range additionally makes use of conditions
of orthogonality and planarity imposed by artifacts and also, or alternatively, an array
of plumb lines.
Volume 2 in this series of textbooks has a full Chapter Ε devoted to calibration. Con-
cepts and practical guidelines for close-range calibration can be found in: Luhmann, T.,
Robson, S., Kyle, S., Harley, I.: Close Range Photogrammetry. Whittles Publishing,
2006.
Exercise 3.1-3. Without determining distortion, how many control points are required
for a test field calibration using a single image? (Solution: 5.) What happens if all
control points lie in a plane perpendicular to the camera axis? (Solution: singularity.)
3.1.3 Correction of distortion

A correction for the objective lens distortion, as determined by laboratory calibration
with a goniometer, by test field calibration or by self-calibration, must be applied to the
image coordinates of the measured target points. The simplest case in which the image
coordinates are corrected for radially symmetric distortion in the form of a piecewise
linear function is first discussed (Figure 3.1-11).
For each individual image point P ' , whose image coordinates ξ and η have been de-
termined, the calculation is as follows. It is assumed that the image coordinate system
has provisionally been located with its origin at the principal point of best symmetry,
PPS (Figure 3.1-12):
• calculation of radial offset ρ — χ/ζ 1 + η2
• extraction of radial distortion value Ap at position ρ on the distortion curve of

Figure 3.1-11
• reduction of radial distortion correction Ap into components Αξ and Αη for image

coordinates ξ and η (these relationships can be seen in Figure 3.1-12):
η
Δξ = -Ap Αη = -Ap (3.1-2)
Ρ Ρ
When correcting distortion in a digital metric image, the finite size of the pixels must
be taken into account. The correction values Αξ and Αη (3.1-2) do not, in general, cause
η
A
Δη
PPS
Figure 3.1-12: Correction of radial distortion
an "integer shift" of the original pixel; a resampling of the original image matrix on
the orthogonal grid of a new matrix would be required, as explained in Section 2.2.3.
Since some potentially significant loss of information is associated wtih every resam-
pling, this should not be done solely for distortion correction. Instead, all image-related
corrections (e.g. also refraction, Section 4.5.1) should be accumulated and only at the
end of the process should the unavoidable resampling take place (Section 4.5.4).
The test field calibration or self-calibration generates a radial distortion polynomial.

The following polynomial has proved itself in practice 2 :
Δρ = gi3p(p2 - pi) + 9up(p4 - Po) (3.1-3)
<7i3, 5i4 · · · (known) polynomial coefficients (indices have been deliberately

chosen to emphasize the link with Equations Β (5.2-9) and
Β (5.2-10) in Volume 2).
ρ . . . Radial offset of image point from the principal point of best sym-
metry PPs
Po •·• (known) radial offset at which Ap is zero. In the distortion curve
of Figure 3.1-8, po — 85 mm, in the distortion curve of Figure
3.1-9, po = 20 mm.
Correction of radial distortion defined by a polynomial curve is done in the same way as
distortion defined by a piecewise linear function. In the former case, Ap is determined
from Equation (3.1-3) at position p, in the latter case from the piecewise linear function
in Figure 3.1-11. The process then continues with Equation (3.1-2).
2
E.g. Fryer, J.: In Atkinson (ed.): Close Range Photogrammetry and Machine Vision. Whittles
Publishing, pp. 156-179, 1996.
Individual polynomials are applied in cases of radially asymmetric distortion compo-

nents or a possible tangential distortion (e.g. Equations Β (5.2-11) and Β (5.2-12) in
Volume 2) 3 .
Exercise 3.1-4. A point with image coordinates ξ = - 8 . 3 4 2 mm and η — 23.593 mm,

referred to the principal point of best symmetry PPs, is to be corrected for distortion
as defined in Figure 3.1-11. (A graphical estimation of Ap from Figure 3.1-11 is suffi-
cient.) (Solution: ξ = - 8 . 3 4 3 mm, η = 23.595 mm.)
3.1.4 Depth of field and circle of confusion

This section is devoted to the question of the extent to which an objective lens can
sharply image an object which is extended in depth, i.e. in the direction of imaging.
The familiar basic lens equation of optics for sharp imagery is:
f ... (constant) focal length of the optical system

s ... (variable) object distance
s' ... image distance, which only corresponds to principal distance c in ideal-
ized case (Figure 3.1-1). In reality, it agrees only approximately with the
principal distance which is purely geometrically defined (Figure 3.1-4)
In the alternative Newtonian form, with s = (x+f) and s' = (a;'+/), the lens equation
is:
xx' = f2 (3.1-5)
Thus, for every object distance s there is a defined optical image distance s', at which,
apart from the effects of lens errors and diffraction effects, all rays emanating from an
object point at distance s and passing through the aperture of diameter d meet to form a
theoretical image point (Figure 3.1-13). This focused optical image plane I P therefore
contains the sharpest possible image of the corresponding focused object plane OP.
Objects in front of and behind the preset object distance will therefore be imaged with
some degree of unsharpness, since the rays meet in a point behind or in front of IP.
In IP there exists therefore a "circle of confusion" of diameter u. We seek now those
object distances sn (near limit of depth of field) and s/ (far limit) which produce the
same diameter u of a circle of confusion, so that we can say that all objects between
Sf and sn, the depth of field, will be imaged in IP with a circle of confusion less than
or equal to u. The dimension u is the measure of the sharpness of the image (circle of
confusion), and will not be exceeded in the range of the depth of field. We try to achieve
a diameter of the circle of confusion of less than 20 μπι in a film camera. The pixel size
of CCD sensors may vary between 10 and 2 μπι. Therefore, the diameter of the circle
3
An example of distortion correction using piecewise linear functions along the 4 semi-diagonals can
be found in: Kraus, K„ Stark, E.: BuL 41, pp. 50-56, 1973.
Figure 3.1-13: Derivation of depth of field and circle of confusion
of confusion should vary between 2 and 10 pixel, respectively. From Equation (3.1-4)
we have:
*' = — t s'n = — f 1
= (3.1-6)
8 - J S n - f Sf - f
We see from Figure 3.1-13 that:
d u
(3.1-7)
S'n ~ S Sf S' - S'f
Introduce Equation (3.1-6) in (3.1-7) and rewrite to obtain:
d S
f ο
Sf = Tt
d S
7f J\ CX 1 Q\
(3.1-8)
J
df + u ( s - f ) d f - u { s - f )
We normally multiply numerator and denominator by f / d , so that the /-number or

aperture stop k = f/d(= focal length / aperture diameter) appears in one position,
only, with a clear effect:
sf2 sf2
Sn = f Sf = f (3.1-9)
2 J 2 J
/ + -u{s - f ) f - -u(s - f )
If sn and sj are given for a particular standpoint, we can find s from Equation (3.1-9),
as follows:
s = 2 S ^ L · (3.1-10)
S n + Sf
So we have the object distance to be set for which u is as small as possible.

To derive the (finite) distance s for which s/ = oo, so that the depth of field extends
from sn to infinity, we must set the denominator of Equation (3.1-9 b) to zero:
soo = / ( - + Λ « / - = (3.1-11)
W ) u
An important practical question is the diameter of the circle of confusion Ui when the
object distance is s* and the preset focusing distance is s. We define ut <0 for Si > s
and solve Equation (3.1-9) for w^:
tn = - ^ '-γ- (3.1-12)
Numerical Example.
a) Given: sn — 16.8 m, Sf — 36.8m, / = 100mm, f / d = 8. Required: s, un, Uf.

Equation (3.1-10) gives s = 23.2m. Equation (3.1-12) gives un — 0.020mm
and Uf must equal un.
b) Given: A camera of focal length 64 mm with a fixed focusing distance of 25 m
(Wild P32, Section 3.8.3). It is to be used to photograph objects at distances from
si = 8 m to S2 — 12.5 m.
Required: u\, «2 and for the aperture stops / / 8 , 11, 16, 22.
With the assumption that c « s' « / , which can be used for larger object distances as
indicated by Equation (3.1-4), Equation (3.1-12) leads to the results in Table 3.1-1.
f / d s 1 = 8 m S2 — 12.5 m S3 = 00
8 44 21 -21
11 32 15 -15
16 22 11 -11
22 16 8 -8
Table 3.1-1: Diameters of circle of confusion for the Wild P32 camera for s = 8, 12.5
and 00 (see Section 3.8.3). Tabulated values of u in μπι
Exercise 3.1-5. What is the significance of the negative sign in the last column of
Table 3.1-1? (Solution: The object distance, in this case S3 = 00, is larger than the
object distance covered by the focusing mechanism, in this case s = 25 m.)
Finally, we need a relation which will yield the setting for that object distance, known
as the hyperfocal distance, which will give a depth of field from sn to 00. In Equa-
tion (3.1-10) divide denominator and numerator by s / ; then, as s j tends to infinity:
β
sao=2sn Sn = ^ - (3.1-13)
The corresponding circle of confusion ua0 is derived from Equation (3.1-11):
/2 (3.1-14)
Exercise 3.1-6. A metric camera has a focal length of 64 mm and a fixed focus of 10 m.
Compute the depth of field (s„, Sf) for a maximum permissible circle of confusion of
0.05 mm diameter for the aperture stops f/4,5.6, 8,11,16,22. (Solution: For example,
for k = 8, sn — 5.1 m, Sf — 336 m.)
Exercise 3.1-7. An aerial camera with c = 305 mm is required to give equally sharp
photographs for the photo scales of 1 : 1500 and 1 : 50000. Compute the best focusing
distance s and the maximum circle of confusion to be expected for this range for the
aperture stops f / 4 , 5.6, 8, 11. (Solution: s = 900m, u = 12μπι, e.g. for / / 8 . )
The necessity to consider depth of field very closely is an important constraint in close-
range photogrammetry which does not occur in "long range photogrammetry" where
fixed focus cameras are used (Section 3.7.1 ff.). The following possibilities exist in
close-range photogrammetry to accommodate the optical requirement for sharp imag-
ing with the photogrammetric requirement for interior orientation which, to achieve
optimal accuracy, does not, in principle, change:
a) fixed-focus cameras: s and c are fixed, the camera is used only for objects within
its depth of field (Example: Wild and Zeiss stereocameras, Section 3.8.2).
b) calibrated focusing rings: An additive constant to the principal distance is spec-

ified for each ring. The rings fit so precisely that the other elements of interior
orientation are not significantly changed (Example: Wild P31, Figure 3.1-14,
Section 3.8.3).
c) cameras with variable principal distance: Certain principal distances are cali-
brated and the corresponding settings indicated by index marks on the focusing
knob of the camera (Example: Zeiss UMK, Figure 3.1-15, Section 3.8.3 and the
widely used semi-metric cameras, Section 3.8.4).
In practice, a circle of confusion of 50 μπι is often adopted. Depending on the pixel

size, the corresponding dimension in pixel units can be derived. More stringent de-
mands are often impossible to meet in close-range photogrammetry.
Exercise 3.1-8. The Wild P31 precision terrestrial camera (see Table 3.8-2) has, with
the focusing ring for s = 25 m, a principal distance of 99.13 mm. What is the focal
length of the lens? (Solution: 98.74 mm.) There is another focusing ring which is
3.05 mm thicker. At what object distance will this ring give the sharpest images? (So-
lution: 2.93 m.) What is then the depth of field (sf - sn) at f / 8 and with u = 0.05 mm?
(Solution: 0.69 mm.) What is the constant to be added to 99.13 mm to give the princi-
pal distance for sharp images of objects at a distance of 10 m? (Solution: +0.59 mm.)
«It I t t
Figure 3.1-14: Circle of confusion of the Wild P31 precision metric camera (c =
100 mm) for the various focusing rings at s = | and for an aperture stop f/22. (The
line u = k is explained in Section 3.1.5.1 on diffraction blurring.)
tttt t t t
Figure 3.1-15: Circle of confusion of the focusable Zeiss UMK 10/1318 universal
metric camera (c = 100 mm) for the calibrated distances s = | and for an aperture stop
/ / 3 2 . (The line u = k is explained in Section 3.1.5.1 on diffraction blurring.)
3.1.5 Resolving power and contrast transfer

The wave nature of light and the consequent diffraction at a circular aperture make
it impossible to generate an ideal, i.e. dimensionless, image point. Instead, the wave
energy forming a "point" image is distributed in a central diffraction disc containing
about 84% of the total energy, together with interference rings around the disc contain-
ing 7%, 3 % , . . . (see Figure 3.1-16). In general, the energy in the interference rings is
insufficient to have an effect on the emulsion or sensor located in the image plane of
the camera; the sensor therefore records only the central diffraction disc. Also, at the
edge of the central disc the energy is too low to produce an image. Experience shows
that of the theoretical total diameter u t k of the central diffraction disc4
uth [μπι] = 2.44Α;λ where k — f / d aperture stop „ _

λ = wavelength in [μπι]
only a diameter of about

u — O.TSuth (3.1-16)
is visible on a photographic emulsion used as sensor (see Figure 3.1-16).
3.1.5.1 Diffraction blurring
A very practical rule of thumb can be derived for the diffraction blurring u: if we
set Equation (3.1-15) in (3.1-16), we have for the average wavelength of visible light
λ = 0.55 μηι
= ^ (3.1-17)
Diffraction theory shows, therefore, that no photographically imaged point can be

smaller than the aperture stop number, expressed in [/im].
Numerical Example (using a photographic emulsion as sensor).
Jfc = 8 Equation (3.1-15): uth = 2.44 χ 8 χ 0.55 = 11 μπι

Equation (3.1-16): w(photograph)= 0.75 χ 11 = 8 μπι
k — 22 Equation (3.1-16): «(photograph) = 22 μπι
The minimum distance Ämm between two diffraction discs which can just be distin-
guished as two points depends on the diffraction blurring u and the minimum energy
difference which can be distinguished in the image from the corresponding sensor (Fig-
ure 3.1-16).
<5min « I (3.1-18)
This minimum distance due to diffraction cannot quite be achieved in practice.

4
Serway/Beichner: Physics for Scientists and Engineers. Saunders College Publishing, 5 lh ed., p.
1222, 2000.
Radiation
energy
Figure 3.1-16: Distribution of light energy for two object points caused by diffraction
at the aperture
3.1.5.2 Optical resolving power
The imaging quality of an optical system such as a metric camera is generally defined
by its resolving power. The optical resolving power, or optical resolution, states how
many dark lines per millimetre can just be distinguished in the image from the equally
wide gaps between them. Figure 3.1-17 shows a test card used to determine resolving
power which is expressed in lines per millimetre ([L/mm]) or line pairs per millimetre
Figure 3.1-17: Resolution test chart (Leica Geosystems)
([Lp/mm]); both expressions are equivalent.5

Based on the presentation above, it is possible to state the influence of diffraction on the
resolving power. The diffraction-limited point separation of ό^η [μ·πι] = k/2 (Equa-
tions (3.1-17) and (3.1-18)) results in a resolution limit for an optical system with
photochemical image recording, i.e. film-based photography, of:
„ Γτ . , 103 [/xm/mmlJ 2000 ,

tfmax Lp/mm - / V 1 = —Γ- (3.1-19
(5min [/im] k
Numerical Example.
k = 22 (small aperture) (3.1-19): Rmax = 91 Lp/mm = 2300dpi
Width of line pair = 1/91 = 0.011 mm = 11 μπι
Width of dark line and gap = 11/2 = 5.5 μπι
(Ignoring blooming, which can amount to a few micrometres
(Section Β 2.2, Volume 2))
k = 8 (medium aperture) (3.1-19): fimax = 250 Lp/mm = 6350 dpi

Width of line pair = 1 /250 = 0.004 mm = 4 μπι
Width of dark line and gap — 4/2 — 2 μπι
(Ignoring blooming, which can amount to a few micrometres
(Section Β 2.2, Volume 2))
Diffraction is only one factor which influences the resolution of an optical system. The
blurring resulting from spherical and chromatic aberration also impairs the resolution.
5
Particularly in the printing industry the resolution (R) is expressed not in [L/mm] or [Lp/mm] but
in [dpi] (dots per inch). Since 1 inch is equivalent to 25.4 mm, the following relationship is obtained:
1 : 25.4 = R [L/mm] : Η [dpi].
Aberration and other lens errors can be reduced by using a small aperture, i.e. by elim-
inating the edge rays, though this procedure increases diffraction blurring.
The optimum resolution is achieved with that aperture at which the sum of the opti-
cal and diffraction blurring is a minimum (Figure 3.1-18). We call this the "critical
aperture", since any smaller aperture will increase the diffraction blurring.
Figure 3.1-18: The definition of critical aperture for two different, good objectives
The optical resolving power depends on the contrast of the line pattern in object space,
i.e. from the difference in brightness between the lines and the gaps between them.
When object contrast is high, the resolving power can be twice as good as when contrast
is low. (Figure 3.1-21 goes into this in more detail.)
The optical resolving power decreases from the centre of the image towards the edge.
Reasons for this are the light fall-off (Section 3.1.6) which reduces contrast towards
the edge of the image field and the increasing blur from centre to edge as a result of
spherical and chromatic aberration.
A practical example will help to illustrate the variation in optical resolving power
within the image plane. This example will also demonstrate the derivation of a rep-
resentative optical resolution using a weighted averaging technique which gives the
area weighted average resolution, AWAR. A Wild P31 terrestrial metric camera (Sec-
tion 3.8.3) was tested with the chart shown in Figure 3.1-17, the optical resolving
power determined in a number of annular zones (Figure 3.1-19, left) and the results
finally averaged (Figure 3.1-19, right).
Film-based aerial metric cameras currently have an optical resolution of 90Lp/mm 6 .
The resolution is, in fact, determined by exposure of a glass plate supporting a photo-
graphic emulsion with a much higher resolution (e.g. 400Lp/mm).
For users of photogrammetry in particular, the quality in the image plane is not of
primary importance. They are instead more interested in the transfer from the image
6
E.g. Light, D.: PE&RS 58, pp. 185-188, 1992.
Area A [mm2] Average R [L/mm]

230 93
730 85
1280 75
1660 75
1990 75
2420 71
1640 63
580 53
ΣΑ = 10530 L(AR) = 759070
AWAR = (Σ(ΛΑ)/(ΣΛ) = 72L/mm
(a) optical resolution for individual (b) calculation of averages to determine
annular zones AWAR
Figure 3.1-19: Additional details: aperture number 8, principal distance 10 cm, dis-
tance of test chart from camera 25 m, high contrast in daylight
plane to the object. The question to be answered is, which details can still be recognized
with a metric camera? This question will be answered for the P31 terrestrial camera
(Figure 3.1-19, Section 3.8.3) and for aerial metric cameras.
Wild P31:
• optical resolution = 72Lp/mm, principal distance = 10 cm, imaging range

= 25 m
• width of line pair in image = 1 /72 = 0.014 mm = 1 4 μιη
• width of dark line and gap in image = 14/2 = 7 /im
• width of dark line and gap at object = 7 χ 25/0.1 = 1750 μπι = 1.7 mm.
Aerial metric camera:
• optical resolution = 90Lp/mm, principal distance = 30 cm, flying height

- 2 km.
• width of line pair in image = 1/90 = 0.011 mm = 1 1 μιη
• width of dark line and gap in image = 11/2 = 5.5 μπι
• width of dark line and gap at object = 5.5 χ 2000/0.3 = 36000μπι =
3.6 cm.
Terrestrial cameras can therefore identify objects in the millimetre range, aerial cam-
eras correspondingly in the centimetre range.
Exercise 3.1-9. Recalculate the estimates for aerial cameras with the following para-
meters:
a) same flying height but a principal distance of 15 cm, (solution: 7.2 cm),
b) same principal distance but with flying heights of 1 km and 5 km (Solution:
1.8 cm and 9 cm).
3.1.5.3 Definition of contrast
Contrast is a measure of the difference in intensities I between (neighbouring) parts

of an object or image (Figure 3.1-20). Intensity differences can either be expressed
as a direct ratio or a relative difference. Using the largest and smallest intensities, the
contrast as direct ratio is designated Κ and defined as:
tf = W / m i n (3-1-20)
The contrast as relative difference is designated C. The difference between the largest
and smallest intensity is usually given with respect to the intensity sum as follows (see
also Equations (3.1-24) and (3.1-25):
^max -^min /"II
O - -
J
— — (3.1-21)
max ι ^min
Contrast in digital images is defined by the differences in grey values g. The grey level
range, generally 255 (Section 2.2.1), is used to normalize the contrast:
C = (3 1 22)
' "
In an analogue photograph (Section 3.2.2.3), the contrast is expressed using the den-
sity differences in the exposed and developed film. The density difference ΔD is the
logarithm of contrast Κ above:
AD = \ogK = log(7max/-imin) (3.1-23)
For example, if / max = 100 / m m , then AD = 2.
3.1.5.4 Contrast transfer function
The resolving power is an insufficient, although important, measure of imaging quality.

The resolving power gives an indication of the smallest details which can be seen in the
image. The extent to which differences between bright and dark areas of larger details
are transferred can only be answered by the more refined theory of the contrast transfer
function.
The theory of the contrast transfer function uses a contrast as defined in Equation
(3.1-20). Figure 3.1-20, left, defines the object contrast C as:
C = [l ~ [2 (0 < C < 1) (3.1-24)

-<i + h
-Object Image
' plane
Figure 3.1-20: Left: object contrast. Right: image contrast
In the imaging process the optical system reduces this object contrast. In the image
plane a reduced image contrast C' is observed (Figure 3.1-20, right):
c = (0 <C'< 1) (3.1-25)
ι + li
The reduction in contrast becomes even larger as the line pattern becomes narrower.
The object and image contrast must therefore be defined in terms of the width (wave-
length) Δ or ί of the pattern (Figures 3.1-17 and 3.1 -20). In the transfer function theory,
the reciprocals of these values are used, i.e. the frequencies 7 F and / , instead of the
values themselves (Figure 3.1-20). The frequency indicates the number of line pairs
(Lp) per unit of length.
The relationship between image contrast C and object contrast C is, by definition, the
contrast transfer function CTF(f) or CTF(F). A dependency on image frequency /
is preferred, which in general is stated in line pairs per millimetre ([Lp/mm]):
CTF(f) = ^ (3.1-26)
It is common to normalize the contrast transfer function CTF(f) by setting CTF(0) —

1. (CTF(0) means the contrast transfer for infinitely wide lines.) Figure 3.1-21 shows
the contrast transfer function for a single optical system and three different object con-
trasts (the various definitions of contrast are investigated in Section 3.1.5.3). The sig-
nificantly better imaging properties for objects with high contrast can clearly be seen.
The resolution is shown in Figure 3.1-21. It indicates the narrowest line pair which
can still be detected, in other words the line pattern with the highest frequency. The
resolution therefore indicates the limiting case; it provides no information about the
imaged quality of object patterns with lower frequencies. Figure 3.1-22 shows, for
example, the contrast transfer function for two optical systems which, by chance, have
the same resolving power. System 1, however, is significantly better than system 2.
7
It is more accurate to speak of spatial frequency rather than temporal frequency. Temporal frequency
is used for phenomena dependent on time (e.g. the propagation of electromagnetic radiation) and spatial
frequency for effects dependent on spatial location (in other words, from coordinates in a coordinate
reference system).
a) High contrast
Κ = 1000 : 1
AD = 3
C = 0.998
b) Medium contrast
Κ = 6.3 : 1
AD = 0.8
C = 0.73
c) Low contrast
Κ = 1.6 : 1
AD = 0.2
R R C = 0.23
Figure 3.1-21: Contrast transfer function CTF(f) and resolution (R) for three differ-
ent object contrasts
CTF
Figure 3.1-22: Contrast transfer functions for two optical systems with the same re-
solving power R
There is a large amount of literature available on the contrast transfer function and the
related modulation transfer function. For a very small selection see: Inglis, Α., Luther,
Α.: Video Engineering. 2nd ed., Mc Graw-Hill, New York, 1996. Graham, R.: Digital
Imaging. Whittles Publishing, 1998.
3.1.6 Light fall-off from centre to edge of image

If a camera were constructed using only a single (thin) lens, then the irradiance would
reduce from the centre of the image to the edge according to the cos4 r law (r is the
off-axis angle). This means that areas with the same surface properties (e.g. a newly
Section 3.2 Photochemical image recording 71
cropped field), and under the same conditions of illumination, at an off-axis angle of
50gon (45°), would have only 25% of the image brightness at the edge of the image
compared with the centre.
With multiple element lenses (compound lenses), the angle of view can be reduced near
the aperture and the light fall-off reduced. Objective lenses in modern metric cameras
have a light fall-off which can be described by a cos™ τ law where the exponent η lies
between 1.5 and 2.5.
Light fall-off can be compensated by the use of a circular graduated filter; this is a grey
filter whose density decreases towards the edge. Light fall-off can also be compen-
sated by post-processing, for example using digital image processing (Section 3.5.1.3
or Section C 1.3.1.2, Volume 2).
Numerical Example. A newly cropped field has a grey value of g = 170 when imaged
at the centre of the field of view by a metric camera with a digital sensor. According to
the cos 2 5 τ rule, a newly cropped field at an off-axis angle of 50gon (45°) has a grey
value of 170 χ 0.707 2 5 = 170 χ 0.42 = 71. Before cropping, the corresponding values
are: image centre: g — 60; at an off-axis angle of 50gon (45°): g = 60 χ 0.42 = 25.
The contrast reduction, from image centre to image edge, of a field before and after
cropping is therefore (Equation (3.1-22)):
J-jq gQ 7j
Image centre: C = — — — = 0.43 Image edge: C = ——-— = 0.18
3.2 Photochemical image recording

A light-sensitive, photographic emulsion is an extremely efficient medium for record-
ing the incoming photons at the image plane of a metric camera. In Section 3.2.1 we
examine the issues relating to the use of glass or film for supporting the emulsion. This
section also evaluates the correction of film deformation. A wide range of optical and
chemical aspects of photography are discussed in Section 3.2.2. Film resolution, con-
trast and suitability for aerial photography complete the discussion in the final three
sections.
3.2.1 Analogue metric image
An image recorded using a film-based metric camera is, after exposure, taken to a
darkroom, developed, fixed, washed, dried and then stored under various conditions of
temperature and humidity. Also, the diapositives and paper prints of the photographs
must be developed and dried. For further photogrammetric analysis it is necessary to
take into account the changes of scale and shape of the photographs resulting from
these processes.
3.2.1.1 Glass versus film as emulsion carrier
The deformations of photographic materials are, above all, dependent on the emulsion
carrier (Figure 3.2-1) which must be chemically inert and highly transparent.
111111111111111111 1 Μ 11 Μ 1111 1 MI Protective coating

Light-sensitive layer (7-20 μιτι)
also called the emulsion
As such, glass plates and films are used in photogrammetry (see Table 3.2-1). Glass
plates are no longer used in aerial photogrammetry. Glass is heavy, easily breakable,
relatively unflat, but stable. The adhesion of the light-sensitive emulsion is less efficient
than with film and therefore deformations of the emulsion tend to be larger8.
Type of base Thickness [mm] Flatness [μπι] Note on flatness

Glass plates 1.3-3.0 30-50 Flat
(15 χ 15 to 24 χ 24 cm) 1.3-3.0 20-30 Ultraflat
6.0 5-10 Mirror glass
ground
Films9 Polyester base 0.06 ± 0.003 5-20 Depending
(Width to e.g. Estar thin to on the
24 cm) 0.18 ±0.005 vacuum or
thick pressure
plate
Table 3.2-1: Thickness and flatness of films and plates (aerial photogrammetry)
Glass plates, usually with a smaller format than used in aerial photogrammetry, are
sometimes used in terrestrial photogrammetry since only a few photographs at a time
are required. However, a range of terrestrial cameras make use of film cassettes or
magazines and a pressure plate for film flattening (see Section 3.8.3).
Film in aerial metric cameras (Section 3.7.2.1) is flattened by means of a vacuum de-
vice. Film is less stable than glass. Its deformations are dependent on: temperature,
relative humidity, tension in handling, storage and ageing (Table 3.2-2). There is a
8
Calhoun, J.M. et al.: Photogr. Eng. 29, pp. 661-672, 1960.
9
Kodak Data for Aerial Photography. Kodak Puhl. M-29, Rochester USA, 1982.
distinction between reversible and irreversible deformations. Age deformations are

irreversible. Deformations can be conformal, corresponding to a similarity transfor-
mation, Equation (2.1-9), or affine, corresponding to an affine transformation, Equa-
tion (2.1-8). The linear deformations shown in Table 3.2-2 can be 30% different in
the longitudinal direction compared with the lateral direction. Systematic deforma-
tions can be corrected. Non-systematic film deformations10, which cannot be corrected
without excessive effort, lie in the range ± 3 — 7 μνα.
Base Unit Relative change of length caused by

Temperature Rel. humidity Development and
per 1°C per 1 % RH ageing
Black/ Colour 1 week 1 year
white film at 50° C at 25°C
film & 20% & 60%
RH RH
PETP 0.06 %0 0.02 0.035 0.050 0.9 0.4
/im/20cm 4 7 10 180 80
PETP 0.18 %c 0.02 0.015 0.020 0.2 0.2
μηι/20 cm 4 3 4 4 4
Table 3.2-2: Film-base deformations for polyethylene terephthalate (PETP), thin

(0.06 mm = 2.5 mil), thick (0.18 mm = 7 mil). (Kodak states film thickness in
1 /lOOOinch = 1 mil.) RH = rel. humidity11
3.2.1.2 Correcting film deformation
This section considers the task of re-creating the imaging bundle of rays from the ana-
logue metric image. For this purpose, fiducial marks are required whose reference co-
ordinates are available in a calibration certificate (Section 3.1.1). These fiducial marks,
which are also imaged on the photograph, are measured with a comparator, essentially
in an arbitrary coordinate system. This results in measured coordinates for the fiducial
marks. Transformation parameters are then calculated which enable the conversion of
all measured image points from the current comparator coordinate system of the pho-
tograph into the reference coordinate system of the metric camera. Within the scope of
this transformation, systematic film deformations are corrected.
Various transformations are available for the correction of systematic film deformation.
In order to achieve a reliable result, due regard must be taken with respect to the number
and arrangement of fiducial marks (Section Β 7.2.2.1, Volume 2). Table 3.2-3 gives the
recommended transformation for typical arrangements of fiducial marks:
10
ASP-Manual of Photogrammetry. 4lh ed., p. 335, Falls Church, 1980.
"Rüger, W. et al.: Photogrammetrie. 4th ed., VEB Verlag für Bauwesen, Berlin, 1978.
Number and Film deformation

arrangement of Transformation which can be
fiducial marks eliminated
Number of Number of (significant part
measurements unknowns only)
4 fiducial marks at the Plane similarity transform (2.1-9)12 conformal part, i.e.
centres of the image a uniform scale
edges (Figure 3.1-10) 8 measurements 4 unknowns deformation
4 fiducial marks in the Plane affine transformation (2.1-8) affine part, i.e.
corners of the image differing scale
changes on the
8 measurements 6 unknowns longitudinal and
transverse axes
8 fiducial marks Bilinear transformation (2.2-1) in affine part and
(Figure 3.7-7) both the ξ- and η directions13 additionally
different scale
changes on
16 measurements 8 unknowns opposite facing
image edges
Table 3.2-3: Correction of film deformation by means of plane transformations for

different arrangements of fiducial marks (see also Section Β 7.1.1, Volume 2)
Exercise 3.2-1. Calibration (camera) and measured (comparator) coordinates of 4 fidu-

cial marks in the centres of the image edges, together with measured coordinates of one
of a number of photo points P , are listed in the following table:
Fiducial marks Calibration/camera Measured/comparator

Photo point coordinates [mm] coordinates [mm]
£ η ? rf
1 1113.002 1000.008 1115.133 1000.985
2 1000.001 1112.998 1002.111 1113.956
3 886.994 999.995 889.103 1000.935
4 1000.004 887.002 1002.161 887.971
Ρ ? ? 1024.334 964.847
12
A plane affine transformation (2.1-8) would be mathematically possible, but it could be dangerous
to extrapolate into the image corners. If different scale changes along the axes are still of interest in
this configuration of fiducial marks, then differing scale corrections can be made using the calibrated and
measured separations of the fiducial marks along the ξ and η directions. (In the solution to Exercise 3.2-3
an adequate 5 parameter transformation is given for this.)
This bilinear interpolation would also be possible with four fiducial marks in the image corners, but
there would be no overdetermined solution in this case. A data error would not be detected and the result
would be in error. However, if the properties of a bilinear transformation using four fiducial marks remain
of interest, the data should first be checked for the presence of gross errors by means of an initial affine
or similarity transformation.
Find the coordinates ξ and η of the photo point in the coordinate system of the metric
camera. Hint: A similarity transformation is appropriate. Its four unknown transforma-
tion parameters can be calculated from the 8 measurement equations using the method
of least-squares adjustment of indirect observations (Appendix 4.1-1). The non-linear
Equations (2.1-9) should be linearized. A more elegant solution is achieved using a
Helmert transformation (Section Β 5.1.1.3, Volume 2):
ξ\ _ (-1.6529λ
η) \ -0.5320 J
1 0000000 0.0000221 λ / 1024.334 λ _ / 1022.198 λ
+
-+- 0v .9995493
y y y j w o ιΓ0 _' 0 0 0 0221 1.0000000) V 964.847 ) ~ \ 963.903 J
Exercise 3.2-2. Calibration (camera) and measured (comparator) coordinates of 4 fidu-

cial marks in the corners of the image, together with measured coordinates of one of a
number of photo points P , are listed in the following table:
Fiducial marks Calibration/camera Measured/comparator

Photo point coordinates [mm] coordinates [mm]
ξ η ξ! v'
1 1106.004 1106.003 1104.412 1104.313
2 893.994 1106.000 892.382 1100.602
3 894.002 894.000 896.107 888.550
4 1106.006 893.993 1108.123 892.241
Ρ ? ? 916.241 1009.421
Find the coordinates ξ and η of the photo point in the coordinate system of the metric
camera. Hint: An affine transformation is appropriate. It should only be used for data
control in the analysis. The final transformation is made, in this case, by a bilinear
transformation. Eight linear equations provide the solution for the eight unknowns.
(More exactly, the corresponding four unknowns for each coordinate axis are deter-
mined from the four corresponding linear equations.)
Solution:
ξ = - 1 7 . 4 9 8 + 0.999804ξ' +0.017688?/ -1.8565 χ 10" 7 ξ'τ/
η = 21.389 - 0.017239ξ' +0.999642?/ -2.1678 χ ΙΟ'1 ξ'η'
Coordinates for point Ρ: ξ = 916.246 mm, η = 1014.453 mm
In addition, the corrections to the ξ' and ^'-coordinates should be represented graph-
ically. Both translations need not be taken into account in the graphical presentation
which, for convenience, can be done separately for each coordinate in the form of
lines of equal correction value. The bilinear property should be visible in the graphical
result.
Exercise 3.2-3. Re-arrange the affine and similarity transformations such that there
are only two translations, a rotation and individual scale factors πΐξ and m , in both
coordinate directions.
Solution:
(ξ\ = (aoi + cos a — sin a πιξξ'
sin a cos a
\V J V ° J V sin a cos a J y τηηη'
a 2
In linearized form with small angle da as well as m^ = 1 + drri( and m , = 1 + dmv:

ξ = aoi + ξ'άτης - η'da + ξ'
η = ao2+ ξ'da + η'άπΐη + η'
Exercise 3.2-4. Repeat Exercise 3.2-1 using this 5 parameter transformation.
-2.2152 λ
-0.8235 )
1.0000000 -0.0001858λ /0.9999026 x 1024.334
0.0001858 1.0000000 J ^ 0.9999514 χ 964.847
1022.198 λ
963.880 J
In addition to film deformation, it may also be possible to account for departures from
planarity (Table 3.2-1). Its distorting effect on the bundle of rays is greater with wider
angle lenses. With an off-axis ray of 50gon (45°), a flatness error causes an error in
the ξη-image plane in the ratio 1 : 1 .
In cameras which do not have a vacuum back or pressure plate, lack of film flatness can
be corrected using a reseau. Semi-metric cameras which use such reseaus are discussed
in more detail in Section 3.8.4.
Exercise 3.2-5. How large is the correction Ap resulting from the spherical deforma-
tion of a glass plate emulsion carrier given by d = 30 μπι (principal distance c =
152 mm, format = 23 cm χ 23 cm)? (Solution: for ρ = 0,40,80,120 and 160 mm,
Ap — 0,7,8, —2, —30 μηι.)
Exercise 3.2-6. Consider a square in object space, placed symmetrically and perpen-
dicularly to the camera axis. Evaluate the deformations resulting from the following
distortion of the glass plate: a) spherical towards the object, b) spherical away from the
object, c) cylindrical towards the object.
3.2.2 Physical and photochemical aspects

Aerial photographs are taken from moving aircraft (40 - 220 m/s, 80 - 440 knots, 150-
800 km/h), sometimes from great heights, through the atmosphere. Since only short
exposure times can be allowed (1/250-1 /1000 s), there is often very little light avail-
able to form the photograph. The atmosphere causes strong reduction in contrast and
change in colours. Practical aerial photography therefore poses many photographic
problems, the most important of which are considered below.
3.2.2.1 Colours and filters
The visible part of the electromagnetic spectrum, i.e. the part to which the human
eye responds, extends approximately from a wavelength of 400 nm to 700 nm (Fig-
ure 3.2-2). The photographic process covers a range from 300 nm to 1000 nm, i.e.
includes part of the ultraviolet (UV) and near infra-red (IR) spectra. It does not cover
either the medium or the thermal infra-red. If all wavelengths of visible light are uni-
form in intensity, the eye sees white light. This white light can be split into a large
number of monochromatic spectral lines which are seen as saturated "spectral colours".
Photographic Visible light Photographic MONOCHROMATIC

ultraviolet infra-red SPECTRAL COLOURS
tu -
U ν£
300 400 500 600 700 800 λ [nm]
"MIXED" COLOURS
Blue Green Red of objects
< -
of filters
nF
Υ//////Λ//////Δ Blue
Additive
V/////A Υ//////Δ Green
primary colours
Κ///////////////! Red
Υ/Ζ//77π Yellow
Magenta Subtractive
V/////A primary colours
V//////A Cyan
Transparent
Y//////A Opaque
Figure 3.2-2: The spectral range of photography, together with the transmission and
absorption characteristics of filters for the additive and subtractive primary colours
If we group these in bandwidths of 100 nm, we obtain the three "mixed" colours blue,
green and red, the three additive primary colours. Other colours can be created by
adding various proportions of these three primary colours, for example by superim-
posing them in projection. A uniform mixture of the three primary colours produces
white. The addition of pairs of additive primary colours produces the subtractive pri-
mary colours cyan, yellow and magenta. These colours are called subtractive since
they are created by subtracting one primary colour from white light:
Cyan = Blue + Green = White - Red

Yellow = Green + Red = White - Blue (3.2-1)
Magenta = Red + Blue = White - Green
Such subtractions are achieved by absorption filters which absorb part of the light
reaching them and transmit the remainder. If white light falls on the filter, the trans-
mitted part gives the filter colour, the absorbed part, which if added to the filter colour
would give white, is called the complementary colour.
Subtractive and additive primary colours are complementary in pairs. Two colours
complementary to each other and superimposed in projection, i.e. added together, pro-
duce white.
Blue + Yellow = White
Green + Magenta = White
Red + Cyan = White (3.2-2)
Blue + Green + Red = White
Yellow + Magenta + Cyan = White
Figure 3.2-2 shows the relations of the six primary colours. Each colour filter in a
subtractive primary colour absorbs the complementary additive primary colour and
transmits the other two additive primary colours. A yellow filter, for example, transmits
green and red. A combination of filters of two subtractive primary colours (e.g. cyan
and magenta) transmits only the part of white light which is the additive primary colour
to both filter colours (blue in this example)14.
Exercise 3.2-7. Why is a yellow filter also known as a minus-blue filter?
Exercise 3.2-8. Which description of a UV-absorbing filter is more correct: a UV filter

or a UV-blocking filter?
Supplementary filters in front of the objective are used in practical photography to

prevent unwanted light from reaching the photographic emulsion (e.g. UV-blocking
filters in high mountains, yellow filters in hazy weather (Section 3.7.5)). The entire
visible spectrum must be blocked with infra-red filters for infra-red photography (Sec-
tion 3.2.2.6), so that only the infra-red rays form the photograph (Figure 3.2-10).
Filters are not, as shown in Figure 3.2-2, 100% effective. A green filter, for example,
can only transmit about 80% of green light, while the other colours are not completely
14
A detailed treatment of colour theory is given in: Wysecki, G., Stiles, W.: Colour Science. Wiley,
2000.
blocked. About 20% blue and 35% red will be transmitted (Figure 3.2-3). The filter
quality must be determined by measuring the intensities of the different transmitted
wavelengths. Figure 3.2-4 shows the opacity and transmission of typical filters used in
aerial photogrammetric cameras. These highly precise filters have a very steep "cut-
off" or edge and therefore block very exactly from a certain wavelength. They absorb
as much as possible of the unwanted light and as little as possible of the desired light.
Figure 3.2-3: Transmission characteristics of a green filter. Above: Transmission Τ in

% of light falling on the filter. Below: Similar, but on a semi-logarithmic scale
3.2.2.2 The photochemical process of black-and-white photography
The exposure (see Section 3.2.2.3) produces a latent image in the layer of light-sensi-
tive silver halide crystals (AgBr, AgCl, Agl etc., size 0.2 μηι) embedded in the gelatine
of the emulsion, which can be made visible by the negative development process (for
example with hydroquinone, alkalis and potassium bromide). Here the silver is sep-
arated from the bromine. The unexposed silver bromide is converted in a fixing bath
(sodium thiosulphate = hypo) into a silver salt easily soluble in water and released.
The remainder of about 5% is then removed by washing. Only metallic silver remains
finally in the exposed photograph, with the areas of high exposure blacker than those
of lower exposure. In the reversal process the exposed silver bromide is released in the
predevelopment (bleaching), while the unexposed silver bromide remains. It is then
uniformly exposed and finally developed as a positive, fixed and washed. The uniform,
intermediate exposure can be replaced by chemical processing with sodium sulphide
(instant-photo process!).
ZEISS filters
log (T[%])
F
0[%] W
τ
10
Dl Ε F 100
LEICA WILD filters
f/ r ( Γ
1/
if
φ I
10
B/JJ
I
Μ 1 1
Dark
0 100
300 400 500 600 700 800 900
λ [nm]
Figure 3.2-4: Spectral transmission/absorption curves of filters for aerial photogram-

metric cameras; Τ . . . Transmission, Ο ... Opacity
a) Exposure
Exposed AgBr b) Latent image

Unexposed AgBr (exposed AgBr)
c) Development to
Unexposed AgBr a negative
washed out
Exposed AgBr
removed d) Predevelopment
Unexposed AgBr
e) Development, after
uniform exposure,
to a positive
Figure 3.2-5: Black-and-white negative process (a,b,c) and reversal process (a,b,d,e)
3.2.2.3 Gradation
The relation (Figure 3.2-6) between exposure Η and density D of a negative or positive
is of great practical importance for the j u d g i n g of photographic materials, processes
and results. T h e density D is measured in a densitometer. A densitometer generates a
constant luminous flux Φο (measured in lumen ([lm])) and measures the proportion Φ
transmitted through the film. The ratio Φ/Φο is called the transparency τ , the reciprocal
Φ ο / Φ = 1 / r is called the opacity O. The logarithm of the opacity is the density D.
(3.2-3)
(3.2-4)
A density D — 2 means therefore that only 1 / 1 0 0 of the luminous flux Φο falling on

the film is transmitted through the film and that 9 9 / 1 0 0 is absorbed; for D — 1, one
tenth of the incident luminous flux is transmitted. A density D = 0 indicates complete
transparency or τ = 1 (see also Section 3.1.5.3).
The logarithmic definition of density accords with the W e b e r - F e c h n e r law which states
that the sensitivity of h u m a n sense organs, at least in the middle range of the usual
exciting intensities, is proportional to the logarithm to the base 10 of the exciting in-
tensity. A logarithmic grey wedge, i.e. a series of densities increasing logarithmically,
therefore appears as a linear scale to the eye.
The exposure Η is defined as the product of the illuminance Ε falling on the emulsion,
measured in [lux] = [lumen/metre 2 ] (1 lx = 1 l m / m 2 ) , and the exposure time t, and is
expressed in lux seconds (lx χ s). The gradation 7 is defined as 7 = tan a , w h e r e a is
the slope angle of the straight portion of the density curve (also called the characteris-
tic curve) in the range of normal exposure (see Figure 3.2-6). The gradation depends
upon the photographic material and its age, the development chemicals as well as the
developing temperature and time. If 7 > 1, w e have a "hard" photographic material or
process which increases contrast: small differences of exposure produce larger differ-
ences of density. If 7 < 1, i.e. for "soft" materials or processes, the relation is reversed.
If 7 = 1, i.e. "normal", the differences of density reproduce 1 : 1 the differences of
exposure. T h e correct exposure for hard films, which is preferred for use in aerial pho-
tography, is m u c h more difficult to j u d g e than for soft films. The solarization point
takes its n a m e f r o m the photographic effect in which the image of the sun is lighter in
a negative than that of the surrounding sky. Apparently an excessive exposure leads to
desensitization of the silver halide.
The normal exposure range between points P2 and P4 of the density curve (Figure
3.2-6) is known as the dynamic region of the density. In black-and-white films it is
around 0 . 2 - 2 . 5 Ζ λ
Since the m a x i m u m resolution in the photograph occurs in the lower third of the normal
range of exposure, photogrammetric photographs should be just barely exposed and not
too long developed. The practical consequences are that exposures must be carefully
measured and, usually, that test photographs must b e made.
Figure 3.2-6: The characteristic curve

E\ = Base point for film speed definition (DIN, ASA; ISO), AD = 0.1 above fog
E2 = Base point for film speed definition (AFS, EAFS), AD = 0.3 above fog
(see Section 3.2.2.4)
Exposure meters measure the illuminance Ε as spot or integral measurements. In spot

exposure measurement 15 , a series of spot measurements is taken along the direction of
flight to determine the minimum and maximum illuminances. The series is continually
refreshed. A photograph is so exposed that the minimum density lies always at point
Pi, but not below it, most of the photograph lies around point P3 and the maximum
density (except for specular reflections such as those of the sun off water surfaces) lies
above point P 3 , but never above point P4. Modern exposure meters regulate aperture
and exposure time automatically and give warnings if points P2 or P4 will be exceeded.
In this case, the film and development specifications must be changed. For integral
exposure measurements, which can only measure the average object brightness over
the entire photograph, one selects the aperture and exposure time for point P3 and
hopes that nothing will be under- or over-exposed.
3.2.2.4 Film sensitivity (speed)
The sensitivity, or speed, of a photographic emulsion is defined as the reciprocal of that

exposure H&d which produces a defined density difference AD above fog level, under
15
Zeth, U.: VT 32, pp. 147-151, 1984.
precise conditions of radiation, exposure and development16. The German Standards

Institute (Deutsches Institut für Normung (DIN)) defines a logarithmic system known
as DIN speeds, according to Equation (3.2-5) and based on that exposure HAD which
produces a density difference AD — 0.1 above fog level (point E\ in Figure 3.2-6). The
American National Standards Institute (ANSI)17 adopts an arithmetic system rather
than logarithmic, according to Equation (3.2-6). The International Standards Organi-
zation (ISO) standard adopts both the DIN and ANSI speeds. Table 3.2-4 shows a
comparison between the two systems.
H
SPIN = 10 log ° (3.2-5)
n&D=o.i
SANSI = 0.8 v H o (3.2-6)

tlhO=0.1
Ho = Unit exposure = 1 lx s (for visible light); Had in lx s
DIN ANSI HAD=0.1 DIN ANSI Had=0.1

0 0.8 1.0 16 32 0.025
1 1 0.79 17 40 0.020
2 1.2 0.63 18 50 0.016
3 1.6 0.50 19 64 0.013
4 2 0.40 20 80 0.010
5 2.5 0.32 21 100 0.0079
6 3 0.25 22 125 0.0063
7 4 0.20 23 160 0.0050
8 5 0.16 24 200 0.0040
9 6 0.13 25 250 0.0032
10 8 0.10 26 320 0.0025
11 10 0.079 27 400 0.0020
12 12 0.063 28 500 0.0016
13 16 0.050 29 640 0.0013
14 20 0.040 30 800 0.0010
15 25 0.032 40 8000 0.0001
Table 3.2-4: Comparison of DIN and ANSI film speeds. HAD=o.i in lux seconds is that
smallest exposure (= illuminance χ time) which just produces a density of AD = 0.1
above fog level.
There exists in the USA a further standard18 for black-and-white aerial films, not the
same as the DIN and ANSI standards. It is derived according to Equation (3.2-7) from
16
Alog Η = log HB ~ logHE\ = 1.30 must yield a AD = DB - DE ι = 0.80 ± 0.05. Β is the
corresponding point on the characteristic curve between points P2 and P4 (Figure 3.2-6).
"Formerly the American Standards Association (ASA).
18
See ANSI Standard PH 2.34-1969.
a unit of 3/2 times that exposure which produces, for a precisely specified "average"
development, a AD — 0.3 above fog level (point Ej in Figure 3.2-6, AFS = Aerial Film
Speed).
S AF s = I J ^ - (3.2-7)
i nAD=0.3
For development conditions different from those of the standard and for colour films,
the speed is quoted in EAFS (= Effective AFS). The advantage is that one can develop
in any way thought suitable, but must also supply the corresponding information19.
Exercise 3.2-9. Calculate the film speeds according to the various standards for the
density curve in Figure 3.2-6.
3.2.2.5 The colour photographic process
Black-and-white films record only grey tones, but colour films contain further, impor-
tant information. Colour films are composed of three light-sensitive layers which are
so treated in development that each layer becomes a colour filter.
Positive and negative colour films

The three light-sensitive layers are shown diagrammatically in Figure 3.2-7. Each layer
is sensitive to a particular spectral range, though each of them is sensitive to blue.
Therefore, a yellow filter layer Y is interposed between layers I and II, which prevents
blue light (Figure 3.2-2) from passing through to the lower layers and is dissolved
in development. Each colour-sensitive layer contains a colour-coupler or dye for the
complementary colour, with which in negative films the exposed part and in positive
films the unexposed part is dyed. Figures 3.2-7 and 3.2-8 show, in a simplified way,
the layers and the effects of development.
Positive colour infra-red film

Colour infra-red, or false-colour, film has layers sensitive to infra-red, green and red
rather than to blue, green and red light. These layers are coupled to the colours cyan,
yellow and magenta so that in white transmitted light the colours red, blue and green
appear, though these do not correspond to the natural colours of the object. This film
is therefore also known as false colour film. Since all three layers are also sensitive to
blue light, a yellow filter must be used (Figure 3.2-10). A yellow filter layer in the film
is therefore not necessary (Figure 3.2-9).
19
Graham, R., Read, R.E.: Manual of Aerial Photography. Focal Press, London, 1986.
Colour sensitivity: blue Dye: yellow

/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / /
green (and blue) magenta
red (and blue) cyan
blue green yellow red white black

1 J J J ϊ f
I
b) II
III
1 yellow yellow yellow yellow

0 II magenta magenta magenta
III cyan cyan cyan
• * • * • •
1 yellow yellow yellow yellow
d) II magenta magenta magenta
III cyan cyan cyan
• t * • t t
e) blue green yellow red white black
a) object radiation
b) exposure and development without colour-coupling
c) exposure to diffuse white light and development in a colour developer with
colour-coupling
d) bleach bath and fixing bath remove reduced silver and silver bromide, the dyes
remain
e) observed colours in transmitted white light (subtractive colour mixing
Figure 3.2-7: Positive colour film (colour reversal process). Red, green and blue are
the three additive primary colours, yellow is introduced as an example of a mixed
colour. Grey areas: reduced silver. Y = Yellow filter layer.
a) blue green yellow red white black
t * * t * *
1 yellow yellow
b) II magenta magenta magenta
III cyan cyan cyan
t • • * * *
c) yellow magenta blue cyan black white
a) object radiation
b) exposure and development with colour-coupling developer, followed by bleach
bath and fixing bath, produce complementary colours
c) observed colours in transmitted white light
Figure 3.2-8: Negative colour film
1 Colour sensitivity: infra-red + blue Dye: cyan

II green + blue yellow
III red + blue magenta
green yellow red infra-red white black
ι
b) II
cyan cyan cyan ψ cyan

yellow | yellow yellow
magenta ; magenta magenta
cyan cyan cyan cyan

yellow yellow yellow
magenta magenta magenta
* t Τ t * t
blue cyan green red white black
Grey areas: reduced silver. a ) - e ) as in Figure 3.2-7.

Blue light is absorbed by a yellow filter.
Figure 3.2-9: Positive false-colour film

3.2.2.6 Spectral sensitivity
The layers of black-and-white and colour films are sensitized by various chemicals
(mostly catalysers) for light of particular ranges of wavelength (Figure 3.2-10). The
layers are not, however, equally sensitive to each wavelength. The spectral sensitivity
S\ defines the sensitivity as a function of the wavelength Λ. It is, by analogy with the
overall sensitivity, expressed as the reciprocal of that exposure (in lx s) which produces
under standardized conditions a particular difference of density (usually AD = 1.0)
above fog level.
Ho
log 5«, (A) = log (3.2-8)
ttAD= 1.0{λ)
Ho is here not the unit exposure (Section 3.2.2.4), but a defined constant reference
exposure to which the exposures at all wavelengths are referred.
Orthochromatic emulsions are sensitive to blue-green light and therefore unsuitable
for red object details. Since the sensitive layer integrates light of all wavelengths,
the loss of the red influence leads to better differentiation in the blue-green range.
Orthochromatic emulsions are therefore frequently used in terrestrial photogrammetry.
They also have a practical advantage in that they can be developed under a red safety
light in often improvised darkrooms in the field.
Panchromatic emulsions reproduce the full visible spectral range in the naturally cor-
responding grey tones. They must be developed in complete darkness.
Infra-red sensitive black-and-white emulsions are also sensitive to wavelengths from
blue to red and must therefore be exposed through infra-red filters if one requires only
the infra-red details in the photograph (see also Figure 3.2-4 and Section 3.2.3). The
film must also be developed in complete darkness.
(Normal) colour and false-colour films are primarily used for photo-interpretation,
though they are all suitable for aerial photogrammetry since they have a sufficiently
fine grain and thin layers. Where film is still required, precision photogrammetry nev-
ertheless continues to use the fine-grained, medium-speed black-and-white films.
The human eye can resolve many more tones and hues of colour than it can of grey
tones. Colour photographs therefore contain significantly more information than do
black-and-white photographs and thus also help to simplify interpretation during ste-
reoscopic observation. The relatively small extra costs of colour photography are well
justified for interpretation tasks.
In (normal) colour film, the red sensitive layer receives not only the red object illu-
mination between 600 and 700 nm, as indicated in the idealized Figure 3.2-7, but also
part of the green object illumination (500 to 600 nm) and even part of the blue object
illumination (Figure 3.2-10). (Normal) colour is also, to some extent, a false colour
film.
In false colour film, the colour separation is worse than in colour film (Figure 3.2-10):
the infra-red sensitive layer, which appears as red, not only receives the infra-red object
illumination but also very large parts of the green and red object illumination.
logs.(λ)
Orthochromatic
black-and-white-film
300 400 500 600 700 800 900 [nm]
log S . (λ)
Panchromatic
black-and-white-film
1 -
Y = Yellow filter recommended
(see Sections 3.2.2.1 and 3.7.5)
at least a
-UV = UV-blocking filter
400 500 600 700 800 900 [nm]
log s . (λ)
Infra-red black-and-white film
Y = Yellow filter
R = Red filter
IR = Infra-red filter
-1 -
300 400 500 600 700 800 900 [nm]
log s t l (λ)
green Colour film
1 -
Λ/Λ UV- blocking filter recommended
(see Section 3.2.2.5a)
300 400 500 600 700 800 900 [nm]
log s (λ) Colour infra-red film

red Y = Yellow filter obligatory
\ ^ 9reen (see Section 3.2.2.5b)
infra-red
/ X red
300 400 500 600 700 800 900 [nm]
Figure 3.2-10: Spectral sensitivity of various types of film

Practical guidelines concerning the various types of film are given in Section 3.2.3.
3.2.2.7 Resolution of photographic emulsions
The decisive factor in the resolution of a photograph itself is the grain. The "grains" in
undeveloped emulsions are the small silver halide crystals embedded in the gelatine.
They range in size from a few nm (in emulsions of very low speed, but high resolution)
to a few /xm (in high speed emulsions). After development, the "grains" are clumps of
metallic silver molecules with diameters from 0.5 to 2.0 μπι.
Kodak specifies a procedure for measuring granularity indirectly 20 . A microdensito-
meter with a circular aperture of 48 μπι diameter is used to scan part of the film with a
relatively even density D = 1.0 and to derive the mean density distribution ση- This
value, multiplied by 1000, is called the Root Mean Square Granularity ( R M S G ) .
RMSG = 103<7D (3.2-9)
A value of RMSG of 13 means that D — 1.0 ± 0.013 under these conditions. The
smaller the value of RMSG, the higher the resolution of the film can be (see Ta-
ble 3.2-5). If a section of film with another density D had been used, instead of
D — 1.0, experience shows that another RMSG would result. According to Diehl 21 :
RMSG(D) = RMSG(D — (3.2-10)
For the best aerial films (Table 3.2-5) the RMSG = 6, i.e. D — 1.0 ± 0.006. Ac-
cording to Equation (3.2-10), for a section of film with D = 2.0 then RMSG(2.0) =
6(2.0 + 1.5)/2.5 = 8, i.e. D = 2.0 ± 0.008; for a section of film with D = 0.3 the cor-
responding values are RMSG (0.3) = 6(0.3 + 1.5)/2.5 = 4.3, d.h. D = 0.3 ± 0.0043.
The resolution of photographic emulsions is determined by test charts (Figure 3.1-17),
imaged by a camera onto the emulsion layer under test. This provides the total resolv-
ing power RT of the sensor combination "metric camera with photochemical record-
ing" which depends on the resolving power of the optics R o and the resolving power
of the photographic emulsion RP as follows 22 :
+ (3
W
±hrp = JW
IQ WIbp -2-n)
From the observed total resolving power RT and the known resolving power R0 of
the optics, determined for example with a very high resolution emulsion in a separate
calibration (Section 3.1.5.2), the resolving power of the photographic emulsion, Rp
can be calculated from Equation (3.2-11).
Numerical Example. Observed: RT = 72Lp/mm; known: R0 = 90Lp/mm (Sec-
tion 3.1.5.2); calculated from Equation (3.2-11): RP — 120Lp/mm.
Kodak Publication M-29. Rochester, USA, 1982.
21
Diehl, H.: IAPRS XXIX(Bl), pp. 1-6, Washington, 1992.
22
Meier, H.-K.: BuL 52, pp. 143-152, 1984.
With low object contrast, modern films have a resolving power between 40 and
250Lp/mm; when object contrast is higher, the resolving power is two to three times
better.
Numerical Example. The width of a line pair for an aerial metric camera with R0 =
9 0 L p / m m was determined in Section 3.1.5.2. The question now arises, to what extent
do these figures degrade when film with a resolving power Rp = 100Lp/mm is taken
into account?
Equation (3.2-11): \/R2T = 1/90 2 + 1/100 2 =» RT = 6 7 L p / m m
Line pair width in image = 1/67 = 0.015 mm = 15 μπι
Width of dark line (or of gap) in the image = 15/2 = 7.5 /im
Width of dark line (or of gap) at the object (image scale 1 : 6666) = 7.5 χ 6666 =
50000 μιη = 5.0cm (ignoring blooming, Section Β 2.2)
The contrast transfer function, introduced in Section 3.1.5.4, provides more detailed
information about imaging quality than the resolving power. Also in this section, the
contrast transfer function CTF(f)o for the optics of a metric camera was introduced.
The photographic emulsion has its own contrast transfer function CTF(f)p. Both
are combined to give the contrast transfer function CTF(f)T of the complete system
(optics plus photographic emulsion) as follows:
CTF(f)τ = CTF(f)0 CTF(f)P (3.2-12)
An example is shown in Figure 3.2-11. It shows the contrast transfer function for film,
in addition to two contrast transfer functions for the optics, one for the centre of the
CTF
t
0 « - f[Lp/mm]
25 50 75 100
Figure 3.2-11: Contrast transfer function of an objective lens for r = 0° and τ = 45°,
as well as for film. The thick lines show the resultant contrast transfer functions for a
film-based camera.
image (off-axis angle τ = 0°) and one for the edge of the image (r = 45°). The thick
solid line is the total contrast transfer function resulting from the combination of film
and optics (r = 0°) transfer functions according to Equation (3.2-12); the thick dashed
line is the corresponding combination at (r = 45°).
3.2.2.8 Copying with contrast control
The density range of an aerial photograph, AD = D m a x — Dmin, can reach a value as

high as AD = 2.5 (Section 3.2.2.3) as a result of shadows cast by objects or clouds on
the one hand and by snow, rock or fields on the other, or as a result of the gradation of
the film. This corresponds to a contrast of 320 : 1 (Section 3.1.5.3). Electronic copying
devices permit the global density differences ADg of large areas to be reduced while
at the same time increasing the detail contrast ADd (Figure 3.2-12). The negative is
scanned by a cathode-ray spot. The light passing through the copy is sensed and either
the intensity I or the scan velocity ν is instantaneously changed. If copying material
with a hard gradation is used (7 = 4, Figure 3.2-6), the detail contrast within the spot
area (variable size between 2 and 144 mm 2 ) will be greatly increased. The electronic
control of density reduction in larger areas is so adjusted as to give densities between
about 0.3 and 1.4. A disadvantage is that at the edges of the photograph, as well as at
other contrast edges, there is a 2 to 3 mm wide transition zone where illumination is
not optimal and detail resolution is impaired.
a) density distribution in negative

b) density distribution in positive with reduced global contrast ADg and increased
detail contrast ADd
Figure 3.2-12: Copying with contrast control (abscissa = image plane)
3.2.3 Films for aerial photography
Table 3.2-5 shows details of nine typical aerial films. The resolving power of each film
is given for two different object contrasts (Section 3.1.5.3). High contrast is represented
by Κ = / m a x : /min = 1000 : 1 (log Κ = density difference AD — 3). Similarly, low
92
red light.
Manu- Product name Type Spectral sensitivity Thickness General sensitivity Recommended Photogr; aphic resolving Gradation Granu-
facturer and film base of base DIN AFS Developer power [Lp/mm] for larity
[mm] EAFS χ objecit contrast Κ (RMSG)
λ= 400 600 800 nm
i 1000:1 | 1:6.1
+
fl
/
?
•—
+ '
Aviphot Pan 80 PE
s
Pan + 0.10 18-22 50-150 χ G 74 c
·— «—
Agfa
2
Polyester
1.3-2.1
Aviphot Pan 200 PE
s
Agfa 0.10 G 74 c 0.9-1.9
1
8
8
20-25
1+
Polyester
\
Chapter 3
f
I
I
--
Plus-X-Ae rographic 200 Versamat 1.3
©
©
Kodak Pan + 0.10

Κ
2402 Estar 125 - 400 χ 885 0.8-2.0
/
Ρ anatomic-X
40 Versamat
a.
u
s
Aerographic II 0.10
8TJ- ; 1.2-2.1
1
•a
32 - 64 χ 885
2412 Estar
ι
High Definition
ο
-
:>
Kodak Aerial 3414 Estar 0.06
8
1.6-2.5
1+
oo ~
1 88
Thin Base
[
/
Infra-red D-19
400 2.3
ο
ο
Kodak Aerographic 2424 0.10 Versamat
25
200-800 χ ; 1.0-2.1
Estar 885
/
Agfacolor
I/
*
Aviphot Color Color
©
Agfa 0.10 21-23 100- 160 χ AP70

κ
X 100 PEl Neg.

Process
\
Aerochrome II MS Color
V --
©
©
©
a
a
Kodak 0.10 AR-5

2448 Estar Pos.
Aerochrome III
f
J
Color
ο
0Γ0
s
s
•Γ
Kodak Infra-red AR-5
8
IR-Pos.
1443 Estar
1
are treated in Section 3.3.3. Pan + emulsions have an extended spectral sensitivity for
Photogrammetric recording systems and their application
Table 3.2-5: Films for aerial photography. The terms "resolution" and "granularity"
Section 3.3 Photoelectronic image recording 93
contrast is represented by a contrast Κ = 1 . 6 : 1 (log Κ = density difference AD =

0.2). The different sensitivities of black-and-white films are needed to allow pho-
tographs to be taken in different seasons, even in late autumn. Panchromatic black-and-
white films are used for topographic mapping, i.e. for wide-area recording of ground
detail. Black-and-white infra-red films are more suitable when only vegetation is of
interest, for example for a forestry inventory (Sections 3.2.2.6 and 3.7.5).
Colour and false colour (colour infra-red) films are primarily used for photo interpre-
tation 23 . The human eye can differentiate many more shades of colour than shades
of grey. Colour images contain significantly more information than black-and-white
images. Since they have been available with a sufficiently fine grain and thin film
base, all are suitable for the analysis of metric aerial photographs. Where film is still
required, fine grained, medium sensitivity black-and-white films are still preferred in
high precision photogrammetry.
Aerial films have a length from 26 to 140 m and are developed either in film developing
machines 24 in a continuous process or in tanks with automatic forward and reverse
winding. They are dried after processing in special drying machines in which the film
is drawn slowly over a drum while clean air is blown onto it. The handling of the film
must impose no mechanical stress and a leader and trailer, each of about 1.5 to 2 m
length, must be provided.
3.3 Photoelectronic image recording

It is becoming much less frequent for a photographic emulsion to be used as the sensor
in the image plane of a metric camera. Instead, an electronic sensor which directly
supplies a digital image (Section 2.2.2) is becoming more common. In Section 3.3.1
the principle of opto-electronic sensors is presented. These devices are suitable for
digitally recording optical radiation, that part of the electromagnetic spectrum covering
visible and infra-red light (Figure 3.2-2). Sections 3.3.2 and following discuss the
quality of image recording by opto-electronic sensors.
3.3.1 Principle of opto-electronic sensors

Solid state sensors are exclusively used nowadays for digital image recording in pho-
togrammetric cameras. They consist of a large number of detectors which register the
photons of the incident light falling on the image plane of the camera. The principle
of opto-electronic image recording by means of semi-conducting elements (preferen-
tially made of silicon) is illustrated in Figure 3.3-1 for a linear array sensor. In the
semi-conducting elements under the electrodes, the photons of the incident light build
up an electrical charge by creating electron-hole pairs. This charge is proportional to
the number of incident photons (the photo-electric effect).
^Literature: Arnold, R.H.: Interpretation of Airphotos and Remotely Sensed Imagery. Waveland PR
Inc., 2004.
24
For example, the Kodak Versamat Ektachrome and Aerochrome RT Processor 1811 (1984).
a) Photons of the incident light
Electrodes
Insolator
• · · • · • Semi-conductor
• · .
b) Active detectors
4 1 \ 1 + Transfer and readout registers
*
(shielded)
*>
a) generation of charge in semi-conductor according to the intensity (number of

photons) of the incident light.
b) transfer and read-out registers arranged next to the active detectors
c) read-out principle by means of CCD (charge coupled device): the charges are
shifted at regular time intervals (ίο, t\, t 2 , . . . ) by one element to the right and
quantized when they reach the last element.
Figure 3.3-1: Linear array sensor 25
The varying charges in the individual semi-conducting elements must be read out. For
this purpose the charges are passed in parallel to a transfer register and from there read
out in series (Figure 3.3-lb). Figure 3.3-lc shows the charges after they have been
passed to the transfer register. The quantities of charge here are the ones produced by
the photons in the active detectors.
25
With acknowledgement to Schenk, Τ.: Digital Photogrammetry. TerraScience, 1999. Figure 3.3-2 is
also taken from this source.
The transfer register also consists of semi-conducting elements which are sensitive
to light; they must therefore be covered. To determine the amount of each charge,
the transfer register has an additional element with its own electrode. At time to this
holds no charge. At time t\ all the charge packets are shifted by one element to the
right (Figure 3.3-lc). With the electrode at this additional element, the amount of
charge currently resident can be converted into an analogue electrical signal which can
subsequently be digitized. At time t\ this charge is the one generated by the last active
detector in the array, at time ^ it is the charge from the last but one detector, and so on.
This technology based on charges coupled with semi-conducting elements is today
dominant. This gives rise to the abbreviation CCD for charge coupled device, with the
abbreviation itself in common use, for example in reference to CCD cameras.
Figure 3.3-1 shows a CCD linear array sensor. For photogrammetric requirements,
CCD area arrays are more appropriate. Figure 3.3-2 shows such a CCD area array
sensor. It consists of:
• a number of active detectors arranged in rows
• transfer registers between the rows
• a serial read-out register
In this case only half the image plane is covered by active detectors; the other half is
taken up by the transfer registers. This configuration is known as an interline transfer
sensor. Since this development there have been other technical solutions in which the
entire image plane is covered with active detectors (full frame sensor).
Some current technical details for CCD sensors prove the high development level of
this technology but also show its limits. At present, the largest CCD area arrays have
4096 χ 4096 = 4K χ 4K = 16M detectors and CCD linear arrays have up to 20K
detectors.
To achieve a high resolution, which is crucially dependent on the number of detectors,
mapping flights dispense with CCD area arrays and instead use linear arrays, for exam-
ple with 20k detectors, in a push broom configuration. This ensures that 20K detectors
determine the resolution rather than 4K detectors; resolution is therefore increased by
a factor of 5.
A digital line camera with a CCD linear array is shown in Figure 3.3-3, left. For each
image line there are individual elements of exterior orientation. In a digital line camera,
the six elements of exterior orientation therefore change from imaged line to imaged
line. In metric cameras which use CCD area arrays or film, the six elements of exterior
orientation only change between each full two-dimensional image (from image plane
to image plane).
For the spatial recording of an object it is a condition that each object element is imaged
by two rays from different directions. The digital line camera (Figure 3.3-3, left) does
not meet this requirement within a flying strip. If instead a 3-line camera is employed
Serial readout register
Figure 3.3-2: CCD area array sensor in interline transfer mode
Figure 3.3-3: Line camera (left) and 3-line camera (right)
(Figure 3.3-3, right) then every object element within a flying strip is recorded from not
just two but three directions; objects recorded in this way can be spatially reconstructed.
If the time required to read out the charges is significantly longer than the reaction
time to build up the charges in the detectors, and if the CCD camera, for example in
an aircraft, moves during this time, then a mechanical shutter of the type used in film
cameras must be employed to limit the exposure time. A mechanical shutter is not
necessary in CCD cameras where the read-out frequency is very high, particularly in
line cameras.
Boyle, W.S., Smith, G.E.: Bell Systems Technical Journal, Volume 49, pp. 587-593,
1970 (a key publication relating to CCD technology). Graham, R.: Digital Imaging.
Whittles Publishing, 1998. Toth, C.: In Fritsch, D„ Spiller, R.: Photogrammetric Week
'99, Wichmann Verlag, pp. 95-107, 1999.
3.3.2 Resolution and modulation transfer

The recording quality of opto-electronic image sensors, in terms of their resolution and
contrast transfer, is primarily influenced by the size of the pixels Αξ χ Αη. The size of
the pixel at the object is given by projecting the surface of an individual detector ele-
ment through the camera optics onto the surface of the object, for example the Earth's
surface. For vertical images taken of flat ground, the size of a pixel Αχ χ Ay on the
ground surface, commonly referred to as "the footprint" (Figure 3.3-3), is given by:
Αχ=-Αξ Ay=-An (3.3-1)

c c
h... flying height, c . . . principal distance or focal length

For images taken with a CCD sensor, the illumination intensity I within a pixel is av-
eraged to a representative value I. Figure 3.3-4 shows the relationships for an intensity
profile along the y direction, i.e. at right angles to the direction of flight.
Figure 3.3-4: Average intensity value within a pixel
The illumination intensity function I ( y ) along a profile can be given as a Fourier series:
oo
I ( y ) = I0 + ^ Ck c o s ( 2 n F k y - ipk) (3.3-2)
fc= ι
J 0 . . . Average intensity value along the entire profile length L

Fk ... Spatial frequencies of the individual component waves of the series.
They are the reciprocals of the wavelengths Lk, i.e. Lk = 1 /Fk. The
wavelengths derive from Lk — L/k, where k runs from 1 to (theoret-
ically) oo.
Ck ... Amplitudes of the component waves
φκ . . . Phases of the component waves.
A representative mean value I within a pixel is given by the following integral taken
over the range [y — A y / 2 , y + A y / 2 ] :
=h1 v - t y
1{z)dz (3 3
' ~3)
The solution to this integral is:
m = Ό + ΣΓ=ι cos(2,Fky - V k ) ^
= +
Comparison with the original intensity function I(y) indicates that the amplitudes Ck
are attenuated to the values C'k as a result of imaging with the detectors. In Sec-
tion 3.3.5.3, the contrast transfer function C T F ( F ) was defined using the ratio of the
k
amplitudes of a square wave function. Since signal analysis is based on the theory of
Fourier series, the usage of square waves is replaced by sine waves, and consequently,
the term "contrast transfer function" by "modulation transfer function" M T F ( F k ) .
The MTF can be stated mathematically as 26
Here the pixel size in the ground is no longer designated by Ay but by Ayr, which is
the footprint of the detector element D.
In Figure 3.3-5 the modulation transfer function is plotted against F A y and L / A y o ·
k D k
The abscissa gives the intensity as a function of frequency Fk and the lower scale as a
function of wavelength Lk.
The modulation transfer function (3.3-5) becomes unity 27 , i.e. there is no attenuation
of the intensity profile I(y), when either the detector size AyD or the frequency Fk be-
comes zero. The first case implies an infinite number of infinitesimally small detectors,
the second case that the intensity profile I(y) has a constant value within the footprint,
i.e. an infinitely long wavelength.
26
The function sin x / x is known as sinc(.x) function, which resembles a damped sinusoidal oscillation.
Figure 3.3-5 shows the first positive lobe of the function.
27
The limit of sin(a;)/x is unity when χ —> 0.
MTF(F k Ay D ) a
theoretical R
1 1 . A>t
0 >• F k Ay D = L
0,25 0,5 k
actual R
00 2
4
Figure 3.3-5: Modulation transfer function MTF and resolution R of a detector
The modulation transfer function (3.3-5) becomes zero, i.e. the variations in the inten-
sity profile are eliminated in the imaging process, when FkAyD = 1. This occurs when
the wavelength of the intensity variations has the same size Ay η as the footprint, i.e.
L = Ayo- Any shorter wavelength, where L < Ayu, cannot be recorded correctly,
even though non-zero values may be obtained. The limiting case, where MTF be-
comes zero, corresponds to the point where the resolution can no longer be digitized
with that detector. (A related definition of resolution for an analogue photograph is
shown by Figure 3.1-21.) This is a theoretical resolution. Due to the limited sensitivity
of the detector, the actual resolution lies around the value FmaxAyD « 0.7 (where the
amplitude of function (3.3-5) reaches some 40% of its maximum) and is dependent on
the opto-electronic sensor used (Figure 3.3-5). In practice, the width of a line pair L ^ n
which can just be resolved by a detector with a projected extent on the object surface
of AyD is therefore:
Width of line pair ( = Lmm — 1/Fmax) « \AAyD (3.3-6)
Between the limits 0 < Fk < F m a x the intensity amplitudes are attenuated. The higher
the frequency (or smaller the wavelength), the larger the attenuation. This attenuation
can be corrected by use of an inverse filter (Section 3.5.2.2). The above considerations
are not sufficient for deciding if a certain wave pattern, with frequencies higher than the
limiting case, can potentially be reconstructed from a digital image, since another para-
meter, the sampling interval, also plays an important role. One should bear in mind that
the considerations in this section presume an infinitesimally narrow sampling interval,
i.e. continuous sampling. For a proper assessment of the frequency-related digitization
quality, detector size and detector spacing must not be treated independently.
3.3.3 Detector spacing (sampling theory)

The previous section concentrated on the imaging properties within the area of a sin-
gle detector element, which was continuously moved across the object. The shortest
intensity waves which can be resolved by a detector with dimension Αη 0 have a wave-
length of /min = ΙΛΑηο, corresponding to a frequency of / m a x = 0.7 /Αη0. Following
Equation (3.3-1), the dimensions Δ y D , Lmm and F max , which refer to the object, can be
replaced by Αηο, lmin and / m a x which refer to the image. The shortest wave compo-
nent which can, with certainty, be resolved by a detector of size Αηο is represented in
Figure 3.3-6.
In the previous section an infinitesimally narrow spacing Αη3 of the detector has been
presumed. In practice, this spacing is finite. The question now arises as to the spacing
required between detectors in order to record the shortest component wavelength in an
image. The sampling theorem states that the sampling interval Ar/S must be smaller than
half the wavelength of the sampled wave28. Figure 3.3-6 shows an adequate sampling
interval Αη8, which, in the case of a CCD camera, corresponds to the detector spacing.
It can be seen that the detector spacing must be significantly smaller than the detector
dimension Αη029.
The sampling interval or detector separation Αη3 can also be related to the resolution
R, expressed, for instance, in line pairs per millimetre (Lp/mm). Here the width of
a line pair corresponds to the wavelength Zmm in Figure 3.3-6, yielding the maximum
28
The sampling theorem is also known as the Shannon-Nyquist Theorem and the limiting frequency
as the Nyquist frequency. (Jerri, A.J.: The Shannon Sampling Theorem—Its Various Extensions and
Applications: A Tutorial Review. Proc. of the IEEE 65(11), 1977.)
29
In commercially manufactured CCD sensors, the reverse is the case (Figures 3.3-1 and 3.3-2). By
using an array of microlenses in front of the detectors, the effective (and larger) detector area is imaged
onto the actual and physically smaller detector surface.
detector spacing Αη3:

Αη3 [mm] < 1 / ( 2 R [Lp/mm]) (3.3-7)
In practice, the numerator 0.7 is chosen 30 , as indicated in Figure 3.3-6, i.e.
Αη3 [mm] = 0.7/(2R [Lp/mm]) (3.3-8)
If the intensity profile is sampled at an interval larger than the values derived from
Equation (3.3-8), the result is an undersampling. Artificial (low-frequency) patterns are
then generated in digital images, caused by the interference of the higher frequencies
in the image with the frequency of the sampling interval. This effect is known as
aliasing 31 . A demonstrative example can be seen in Figures 2.2-7 to 2.2-9, where
artifacts appear in the window blinds. The sampling interval has a value which clearly
cannot fully record the original texture in the slats of the blinds.
Equation (3.3-6) permits a comparison between the resolution of a CCD sensor and
photographic film. From Table 3.2-5, a film with a resolution of 100 Lp/mm can
be found 32 . This corresponds to a line pair width of 10 μτη in the image. In order
to achieve the same resolution with an electronic sensor, the pixels must have the
following size (derived from Equation (3.3-6): Ar)s = 0.7/(2 χ 100) = 3.5 μπι).
The analogue photograph provides the given resolution across an image format of
23 cm χ 23 cm. In order for CCD technology to deliver the same quality just by
replacing the film plane by a CCD chip, it would require 66000 detector elements
( = 230000/3.5) along one side of the image. It is difficult to imagine such a number
of pixels in area CCD arrays, even in the long term.
In area CCD arrays, sampling interval and detector size are more or less identical and
given by the detector arrangement on the sensor plane. In the case of linear array sen-
sors, as they are used in line cameras, the same is valid within the line, while perpendic-
ular to the sensor line, i.e. in the direction of flight, the sampling interval is variable and
depends on the relation between flying speed above ground (projected into the image
plane) and readout rate. Flight planning has to take this into consideration, particularly
in order to avoid undersampling.
Exercise 3.3-1. For a camera with an image format of 23 cm χ 23 cm, there is an

equivalence in resolution of 100 Lp/mm between film and a (fictitious) electronic
sensor having 66000 χ 66000 electronic detectors. Nowadays, CCD sensors with
10000 χ 10000 elements with 9 μτη pixel width are realistic. What is the correspond-
ing resolution in the image plane? If the flying height is kept the same as with a film
30
Note: the factor 0.7, which has been applied in the previous section to determine the largest detector
size, is not related to the factor 0.7 here. While in the first case the detector size has been reduced in
order to detect a reasonably high amplitude (noticeably greater than zero), here the detecting interval is
reduced in order to guarantee the reconstruction of a certain minimum wavelength (noticeably different
from the Nyquist frequency).
31
Smoothing filters, which suppress high frequencies (prior to digitization) in order to avoid aliasing,
are known as anti-aliasing filters.
32
Note: the specified resolution is valid for the film and not necessarily for an exposed image. A pho-
tograph's resolution also depends on negative effects due to the lens, filters and atmospheric disturbances.
How the cumulative resolution may be calculated, can be found in Section 3.2.2.7.
camera, how must the principal distance c be changed for the CCD camera in order to
achieve the same resolution on ground as film would deliver? What ground area would
be covered compared to the film camera? (Solution: c must be 2.57 times longer.
Area covered by the digital camera would be just 2.3% of the area covered by the film
camera of the same resolution.)
Exercise 3.3-2. Reconsider the situation in the case where the camera optics have
a resolving power of 90Lp/mm (Hint: Make use of Equation (3.2-11)). (Solution:
67 Lp/ mm.)
Further reading for Sections 3.3.2 and 3.3.3: Schenk, Τ.: Photogrammetry. Terra Sci-
ence, 1999. Holst, G.C.: CCD Arrays, Cameras and Displays. SPIE Press, JCD Pub-
lishing, 1996. Inglis, A.F., Luther, A.C.: Video Engineering. 2 nd ed„ McGraw-Hill,
1996.
3.3.4 Geometric aspects of CCD cameras

When CCD cameras are employed for photogrammetric purposes it is, as a rule, neces-
sary to have knowledge of the interior orientation. The required parameters are defined
in the calibration certificate (Section 3.1.1) and are:
• the principal distance
• the coordinates of the principal point of autocollimation and principal point of

best symmetry
• the objective lens distortion.
In addition, CCD cameras must also provide information on the positions of the active
detector elements and the planarity of the recording surface defined by the surfaces
of the detectors. Nowadays, detectors can be arranged in an orthogonal grid with a
positional accuracy of around 0.2 μηι, a deviation which can normally be ignored in
photogrammetric processing. In contrast, lack of flatness can be in the order of 10 /um.
This should be taken into account in precise analyses, particularly where a wide field
of view is employed.
In digital photogrammetry, the correction of lens distortion is similar to that in analyt-
ical photogrammetry (Section 3.1.3). The pixels must be "shifted" by the correction
values Δξ and Αη (Equation (3.1-2)). This would disturb the pixels from their arrange-
ment in a strictly orthogonal grid. In order to preserve a strictly orthogonal structure
which is convenient for image processing, a resampling is required, as described in
Section 2.2.3. However, it is expedient to adopt a virtual image in the solution, as
explained in Section 4.5.4.
Correction for lack of flatness is done in a similar way and in the following steps:
• convert the Αζ (height) deviations of individual detector elements into radial

shifts Ap in the image plane using the formula Ap — Αζ tan r , where τ is the
off-axis angle of the imaging ray on the image plane
• resolve Ap into component corrections Αξ and Αη by means of Equation (3.1-2)
• continue the process by means of a virtual image (Section 4.5.4)
In cameras which work with video analogue signals and frame grabbers (Section 3.8.7),
pixel positioning is done through line synchronization. Irregularities which may oc-
cur in this process can lead to displacements between image lines (line jitter) and
also within individual lines (pixel jitter). A typical line shift is illustrated in Fig-
ure 3.3-7. The correct image is indicated on the left. The centre diagram shows the
image recorded with displaced lines and pixels; the output using pixels in an orthogo-
nal grid (right) results in a distorted image.
•DHHDD DDHSDD DDHHDD

DBDDSDD • • • a m • B B K D D
•LIMED DUBBED •LIDDED
DDSBDD • • Β • • DDBDDD
correct image lines shifted during distorted image
recording after display in an
orthogonal grid
Figure 3.3-7: Effect of line jitter.33
Finally it should be noted that the electronic signals in analogue videocameras drift
by relatively large amounts during the warm-up phase. In practice, therefore, a corre-
sponding time delay should be observed before use.
Further reading: Luhmann, T., Robson, S., Kyle, S., Harley, I.: Close Range Pho-
togrammetry. Whittles Publishing, 2006. Shortis, M., Beyer, H.: In Atkinson (ed.):
Close Range Photogrammetry and Machine Vision. Whittles Publishing, Caithness,
UK, pp. 106-155, 1996.
3.3.5 Radiometric aspects of CCD cameras

3.3.5.1 Linearity and spectral sensitivity
The electrical signals delivered by CCD sensors are largely proportional to the number
of photons incident on the detectors. Departures from linearity can be held to within
33
Taken from Dähler, J.: Proceedings of the ISPRS Intercommission Conference on "Fast Processing
of Photogrammetric Data", pp. 48-59, Interlaken, 1987.
1 % of the true value 34 but there are very large departures from linearity when the semi-
conductor elements are close to saturation. In fact, very large amounts of radiation
cause an "overflow" into the neighbouring detector elements. The result is blooming,
similar to the effect which also occurs with very bright objects imaged onto film.
The spectral sensitivity (Section 3.2.2.6) of the silicon sensors normally used ranges
from 400 to around 1100 nm, i.e. from visible light into near infra-red, thus consider-
ably wider than that with analogue photography.
3.3.5.2 Colour imaging
Imaging in multiple spectral regions, i.e. colour imaging, is possible in several ways
with CCD cameras. One variant, which is preferred when employing CCD area arrays,
involves the use of a cluster of several cameras. The individual cameras in the cluster
are provided with identical sensors, each sensitive in the range 400 to 1100 nm. How-
ever, different filters which select different parts of the spectral region are attached to
the front of the objective lenses. For example, if normal colour images are required
then there are three cameras in the cluster. The first camera has a blue filter, the second
a green filter and the third a red filter (Figure 3.2-2). If a colour infra-red image is
required then the first camera has a green filter, the second a red filter and the third
an infra-red filter (Figures 3.2-2 and 3.2-10). To display a colour infra-red image, for
example on a colour monitor, the image matrix from the first camera is represented in
blue, the second in green and the third in red (see Figure 3.2-9).
A cluster solution is only realistic for satellite and aerial photography. In these cases
the imaging distances are so large that, despite the small relative displacements of
the cameras, identical images are obtained. At close ranges an alternative solution is
required. One solution makes use of a single camera with a single detector field. The
imaging for the three separate colours takes place sequentially. For each exposure a
different filter from one of the sets mentioned above is positioned in front of the lens.
Between each exposure, the object and the camera's interior and exterior orientation
must remain fixed.
Another solution also employs a single camera with a single detector field. This so-
lution achieves colour separation with a filter mask placed in front of the detectors.
There are various arrangements of filters. Figures 3.3-8 shows two arrangements of
(R)ed, (G)reen and (B)lue filters. For natural objects, the distribution of the green in-
tensities comes close to the brightness distribution. For capturing images with a wide
range of brightness levels, the right-hand arrangement has a certain advantage. Full
image arrays for each colour are then created by interpolation between the pixels of the
individual colours.
There is a very elegant solution for colour images from digital line cameras. The
rays from every object point are divided into three separate and displaced components
by means of a beam splitter. Before each component ray then reaches one of three
34
Dierickx, B.: OEEPE-Publication 37, p. 66,1999. This publication also discusses CMOS technology
which may one day replace CCD technology.
EEÜJEE 3E E i i m i ®
E E E E 1 I HE i E I E
E i E i E 3E e i
E i E E E 3E I E I E
0 Ü 1 0 0 E [IE Ε [Gj 0® [G) ® Hm
E I E E E HE IEIEIHB
Figure 3.3-8: Filter masks for CCD colour imaging
parallel linear CCD arrays, it passes through a colour filter. The type of filter used
is an interference filter which provides a better colour separation than the individual
layers of colour film whose spectral sensitivities have a relatively significant overlap
(see Figure 3.2-10).
All the techniques of colour separation presented here can be applied beyond the three
basic colours. Particular importance is attached to imaging in the four spectral regions
blue, green, red and near infra-red. In contrast, analogue photography can only simul-
taneously record three colour regions.
3.3.5.3 Signal-to-noise ratio
The radiometric sensitivity across the whole spectrum, comparable to the general sensi-
tivity of film (Section 3.2.2.4), is defined for CCD sensors by the signal-to-noise ratio,
SNR:
SNR = — (3.3-9)
σ3
S ... electric signal intensity, as . . . standard deviation of system noise
An SNR value of unity means that the signal S, which contains the relevant informa-
tion, has the same magnitude as the noise a s : in this case the signal disappears amongst
the noise and no relevant information can be extracted from it.
CCD sensors can reach a maximum SNR of approximately 1000 : 1 and sometimes
more. The maximum SNR is the dynamic range DRs of the sensor. In order to quantize
the recorded illumination intensity, a step size interval of 2a s should be chosen 35 , i.e. in
this example the dynamic range of the digital image DRj = DRs/2 = 1000/2 = 500;
a pixel must therefore be represented by 9 Bits (29 = 512).
Noise a s has various sources. Amongst these are dark current noise in which photons
are generated by thermal effects rather than incident light from the object, charge trans-
fer noise due to inefficiencies in transferring charges between neighbouring detectors
during the read-out process, transmission noise in which signals are deformed during
35
Baltsavias, E. in Fritsch/Spiller: Photogrammetric Week '99, Wichmann Verlag, pp. 155-173, 1999.
transmission and before analogue-digital conversion and quantization noise caused by

rounding errors in the analogue-to-digital conversion.
A test card with homogeneous surface properties and illumination is used for the de-
termination of detector noise, a method comparable with the determination of film
granularity (Section 3.2.2.7). The noise a s can be determined from the variation in
grey level. Variations in illumination and surface properties can falsify results. It is
possible to allow for these effects by making a series of exposures with the same ex-
terior orientation and, from the spread of results for individual pixels then deduce the
standard deviation a s .
The signal-to-noise ratio SNR depends on various parameters. In the following dis-
cussion, only the principal parameters for the SNR as a general function / ( ) will be
evaluated:
SNR = /(Αη, At,...) (3.3-10)
Αη ... Detector size (Section 3.3.2)

At ... Available integration or exposure time for recording photons
The parameters of function (3.3-10) interact with one another. The SNR, which can
also be called radiometric resolution, improves under the following conditions:
• with larger Αη, which means that geometric resolution is worse
• with a longer exposure time
Both types of resolution (radiometric and geometric) must be carefully balanced ac-
cording to priorities in the construction of imaging sensors. In photogrammetry, geo-
metric resolution has the higher priority. When resolution is discussed in photogram-
metry, then geometric resolution is implied.
The resolution types mentioned are also relevant to analogue photography. Radiometric
resolution is expressed by the general sensitivity (Section 3.2.2.4) or the granularity
(Section 3.2.2.7). Geometric resolution is given in the form of line pairs/millimetre
(Lp/mm).
3.4 Digitizing analogue images

The explanations above have shown that the geometric resolution of analogue photog-
raphy is significantly higher than the geometric resolution of current opto-electronic
(area array) sensors. For many years to come, therefore, objects will be recorded by
light-sensitive film in metric cameras when executing the most accurate photogrammet-
ric tasks. However, in order to apply the methods of digital photogrammetric analysis,
film can be digitized after development.
Section 3.4 Digitizing analogue images 107
3.4.1 Sampling interval

The sampling interval, which equates to the distance between each pixel in an orthogo-
nal grid, must be compatible with the resolution in the analogue image in order to avoid
loss of information. Section 3.2.2.7 discussed the geometric resolution RT of a "pho-
tographic emulsion" sensor. For current films it varies between 50 and 100Lp/mm. In
a similar way to the sampling theorem, the digitization sampling intervals A(s and Αη3
in both coordinate directions should be chosen as follows 36 :
Αξ3 [mm] = Αηβ [mm] = 0.7/(2Λ) (3.4-1)
Numerical Example. Adequate sampling intervals for the analogue images mentioned
above are between 3.5 and 7μπι (e.g.: Αξβ = Αη3 = 0.7/(2 χ 50) = 0.007 mm =
7 μπι).
The sampling interval should not be confused with the size of the detector elements.
In an ideal situation there is no difference between the detector dimensions and the
sampling interval. This concept lies at the basis of Figure 2.2-3 and many related
diagrams. However, in Section 3.3.3 it was shown that size of the detectors could
certainly be larger than the sampling interval. A larger detecting surface also improves
the radiometric resolution (Section 3.3.5.3).
The sampling interval required on the basis of the geometric resolution of an analogue
image, generates a very large amount of data. To store a pixel generally requires 8 bits
corresponding to 1 byte. At a sampling interval of 7 /im, a 23 cm χ 23 cm metric image
requires a storage capacity of approximately 1 gigabyte, where 1 GB « (230000/7) 2 .
For a colour image it would be 3 GB. This quantity of data only applies to a single
image. In digital photogrammetry, data is required from many images, although only a
few must be simultaneously resident in computer memory. From these considerations
and many others (e.g. Section C 1.1.1.1, Volume 2), when scanning images a sampling
interval between 7 and 30 μπι is chosen in practice.
3.4.2 Grey values and colour values

The discussion will initially be limited to the digitizing of an analogue black-and-white
photograph. The grey values are measured by means of a densitometer or several
densitometers (Section 3.2.2.3). A densitometer transmits a constant stream of light φο
and uses a CCD sensor to measure the component not absorbed by the film. The light
source and CCD sensor are illustrated in Figure 3.4-1 for a flat-bed scanner, a widely
used design. The film carrier is moved with respect to the light source and sensor. The
right hand side of Figure 3.4-1 indicates that a diffuse light source is employed and that
several CCD detectors receive the component of the light which is not absorbed.
36
It is interesting to note the connection between the sampling interval, i.e. pixel size Δ, and the
parameter dots per inch [dpi], employed in the printing industry. Making use of the footnote in Sec-
tion 3.1.5.2, the relationship is: R [dpi] = 25.4 χ 1000/Δ [μιτι], A 7 μπι pixel therefore corresponds to
3629dpi = 25.4 χ 1000/7.
"Taken from Kölbl, O.: OEEPE-Publication 37, pp. 135-150, 1999.
Film
CCD detector
Figure 3.4-1: Flat-bed scanner with diffuse light source.37
The CCD detectors measure the light which is not absorbed and there exists a lin-
ear relationship between this quantity of light and the charges developed in the semi-
conductors (Section 3.3.1). The charges are quantized and converted by the scanner
into grey values g. The logarithms of the grey values g must have a linear relationship
with the densities of the corresponding elements of the film (Section 3.2.2.3).
Figure 3.4-2 shows the relationship between the density D of the exposed and devel-
oped film and the logarithm of the quantized grey values g for a specific scanner. There
is good linearity up to a density of 2.0 but not in darker regions. The steep curve in
very dark regions indicates that here the scanner does not satisfactorily reproduce the
density gradations.
density D
Figure 3.4-2: Density of a reference pattern with respect to grey value g (Scanner:
DSW 300 from LH Systems, pixel size 12.5 /mi) 38
Practical tip: Film negatives often generate high densities. To create the conditions
for optimal digitizing, a contact copy onto transparency film (Section 3.2.2.8) is rec-
ommended. The layers in positive colour film are generally highly saturated. When
digitizing colour films, the radiometric demands on the scanner are therefore greater
38
Taken from Baltsavias, E„ Kaeser, Ch.: OEEPE-Publication 37, pp. 111-134, 1999.
Section 3.4 Digitizing analogue images 109
than when digitizing black-and-white films. Figure 3.4-2, in fact, results from calibra-
tion with a colour film.
Black-and-white images have a density range of 0.1 - 2.0D, positive colour films a
range of 0.3-3.5Z) (source: see footnote ). The density dispersion σο of developed
film lies around ±0.006 in areas with a density of ID (Table 3.2-5 shows the RMSG
which is 1000 times greater than σρ; see also Table Ε 4.2-1, Volume 2). For less dense
regions, say 0.3.D, σο lies around ±0.0043. If a quantization step corresponding to
twice the standard deviation of the density dispersion is chosen for the detector charges
(Section 3.3.5.3), then the following dynamic ranges DRi for grey and colour levels
are obtained:
• black-and-white film: DRj = 2.0/0.0086 ~ 223
• colour film: DRi = 3.5/0.0086 « 407
8 bits (28 = 256) are therefore just sufficient to represent grey values in black-and-
white films. For colour films the 8 bits normally used for image data are clearly too
few.
The interaction between geometric and radiometric resolution, mentioned in Section
3.3.5.3, also applies to scanning. For example, sampling with large pixels delivers a
better radiometric quality, i.e. a better signal-to-noise ratio, than sampling with small
pixels (Table Ε 4.2-1, Volume 2, provides empirical values).
3.4.3 Technical solutions

In addition to flat-bed scanners (Figure 3.4-1), there also exist drum scanners where the
film is mounted onto a cylindrical drum. Drum scanners are widely used in the printing
industry. In photogrammetry, flat-bed scanners take precedence; they generally have a
smaller format than drum scanners.
With regard to the arrangement of light-sensitive detectors, there are three different
principles of construction:
• single detector, which is driven line by line across the film, usually by means of
rotation
• linear array, which is moved in strips across the film. Several thousand detectors
are arranged in a single array
• area array, which records partial images. These partial images are combined
into a complete image by means of a reseau whose crosses have known coordi-
nates. The reseau can be marked on a glass plate which is placed on the film
(Figure 3.4-3). The reseau can also be introduced into the image during expo-
sure. This requires a reseau to be incorporated into the recording film camera
(Section 3.8.4).
Radiometric calibration of a scanner is particularly difficult with linear and area array
scanners. Each detector in the array must have the same sensitivity, otherwise patterns
of stripes (linear arrays) or blocks (area arrays) are produced. However, to a large
extent these patterns can be removed in digital post-processing.
In photogrammetry, the positional accuracy of the pixels resulting from the digitizing
process is of considerable importance. For currently available photogrammetric scan-
ners, the average accuracy of an image coordinate lies around ±1.3 - 2.3 μηι 39 . A
precondition here is that the scanner's calibration certificate provides reference coordi-
nates for the detectors, or at least for a few evenly spaced detectors (further details in
Section Ε 4.2.1, Volume 2). In contrast, repro scanners have significantly larger posi-
tional errors but, in general, provide a better radiometric quality than photogrammetric
scanners.
Among others, photogrammetric scanners are currently offered by the following com-
panies (product names in brackets): Leica Geosystems (DSW), Microsoft-Vexcel (Ul-
traScan), Wehrli (RM) and Intergraph (PhotoScan).
3.5 Digital image enhancement

Digital images, either directly recorded in a metric camera or produced indirectly by
digitizing an analogue image, can be enhanced through the methods of digital image
processing. These improvements are made with regard to the optimal conditions for
both visual interpretation of the images and automated analysis of the images.
Image enhancement is the central theme of this section; digital image analysis (ex-
traction of 2D and 3D data) will be discussed later, particularly in Section 6.8. The
following discussion is limited to black-and-white images (colour images are handled
in Section C 1.3.4, Volume 2) and starts with contrast and brightness enhancement.
39
Baltsavias, E., Kaeser, Ch.: OEEPE-Publication 37, pp. 111-134, 1999.
Section 3.5 Digital image enhancement 111
3.5.1 Contrast and brightness enhancement

In these procedures, the grey values g ^ of the original image matrix G are transformed
to grey values g ^ in the enhanced matrix. The simplest transform function is:
9ij-=cgij+d (3.5-1)
c . . . parameter responsible for contrast, d . . . parameter responsible for brightness

With this transform function, the grey levels of individual pixels are altered indepen-
dently of one another. Viewing all the grey levels together in an image results in a
change of contrast for which parameter c is responsible. Parameter d alters the image
brightness. Figure 3.5-1 shows how contrast and brightness of a digital image can be
altered by the choice of c and d.
Contrast enhancement can also be applied differentially (Figure 3.5-2). Other functions
can also be chosen in place of the linear transform function. Figure 3.5-3 suggests two
general functions.
By making contrast enhancement a function of grey level, the sensitivity of the human
eye can also be taken into account. CCD sensors generate grey levels which are propor-
tional to the number of photons falling on the individual detectors. To the human eye,
grey value sequences appear linear when they follow a logarithmic curve to base 10
(see Section 3.2.2.3 and also Sections 3.3.5 and 3.4.2). Conversion using a logarithmic
transform function is therefore a standard procedure in digital image processing and
extremely efficiently implemented in many systems.
In order to save computing time, transformations such as the one in (3.5-1) are solved
by the use of look-up tables, abbreviated to LUT. To this end, transformation values
gtj are computed for all grey levels gtJ, usually in the range 0 - 2 5 5 , and stored in a
table. When an image is processed it is then only necessary to "look up" the position
gij in the L U T and extract the corresponding grey value g r j ; in this way the modified
grey levels are obtained without computation.
When grey values are manipulated as described, care must be taken to ensure that
the new grey values g ^ continue to lie within the available value and dynamic ranges
(for example from g m \ n = 0 to g max = 255). For values lying outside the ranges the
following conditions are applied:
if 9ij > ömax ("too bright"), then set gl3 = gmax

(3.5-2)
if 9ij < 3mi„ ("too dark"), then set g^ = gmm
By a suitable choice of transformation parameter it is possible to avoid exceeding the
limits of the value range. For example, parameters c and d in function (3.5-1) can be
chosen as follows:
255
c = d = -cgmin (3.5-3)
9max 9mm
The extent to which a human observer judges the quality of an image is dependent
on the frequency distribution of the grey values; in other words, the human eye ex-
pects a "balanced" image containing grey values across the entire range. Figure 3.5-4
g
255
Contrast reduction
c<1,d=0
255
255
Contrast increase
c>1,d=0
255
255
Contrast fixed,
brightness increased
c=1,d>0
255
Grey value complement,

e.g. negative from positive
c=-l,d=255
0 255
Figure 3.5-1: Contrast and brightness enhancement with a linear transform function
Figure 3.5-2: Contrast enhancement with different linear transform functions applied
to different grey level regions
zoo Contrast is:

— increased in dark regions
reduced in bright regions
— reduced in dark regions
increased in bright regions
9
255
Figure 3.5-3: Contrast enhancement dependent on grey level
shows the frequency h(g) with respect to grey value g. This representation is called a
histogram.
An interesting possibility for altering the grey level distribution in an image lies in
modifying its grey level histogram. Although many changes can be implemented in
this way, two types of histogram modification have achieved particular importance.
Both offer the observer a balanced distribution of grey levels. One of the modifications
is histogram equalization and the other is histogram normalization (Figure 3.5-4).
The discussion will first look at histogram equalization and then at histogram normal-
ization.
Figure 3.5-4: Frequency distribution (histogram) of a real image (REAL) and ideal
image (IDEAL-uniformly distributed, IDEAL-normally distributed)
3.5.1.1 Histogram equalization
To a first approximation this represents a balanced image although, from a statistical

point of view, the objective is an equal distribution of grey levels. The procedure is
called histogram equalization because the frequency distribution is "levelled". As is
generally the case with histogram modifications, the required transformation is based
on cumulative frequency distributions, here using Sh{g) for the cumulative frequency
distribution of the real image G being processed and Sh (g) for the ideal modified image
g. For histogram equalization, the cumulative frequency function of the modified image
is a sloping line. Figure 3.5-5 shows the functions for the real and ideal images whose
frequency distributions are shown in Figure 3.5-4. These functions are also known as
cumulative histograms.
Figure 3.5-5 also indicates how to access the look-up table, i.e. the tabulated relation-
ship between real grey level g and ideal grey level g. For any grey level g, the cumula-
tive frequency Sh(g) (T in Figure 3.5-5) indicates how often this grey level and lower
values appear in the real image. By setting this cumulative value Sh(g) equal to the cu-
mulative frequency Sh(g) (<— in Figure 3.5-5), the ideal grey value corresponding to g
is directly obtained on the g axis (J. in Figure 3.5-5). In this way the corresponding ideal
value g is found for every grey level g and the transformation function for histogram
equalization is thereby defined through the look-up table. The histogram equalization
requires no additional outside information; it can therefore run automatically.
3.5.1.2 Histogram normalization
An image with equally distributed grey levels (derived from histogram equalization,
Section 3.5.1.1) is not (yet) visually balanced. The human eye can differentiate medium
grey levels better than bright or dark regions. It is possible to take account of this char-
acteristic by introducing a normal distribution of frequencies as the ideal target func-
tion. In this case the required transformation function is given by a look-up table based
Figure 3.5-5: Cumulative frequency distributions of a real (REAL) and ideal im-
age (IDEAL-uniformly distributed = Histogram equalization), whose frequencies are
given in Figure 3.5-4
Figure 3.5-6: Cumulative frequency distributions of a real (REAL) and ideal im-
age (IDEAL-normally distributed = Histogram normalization), whose frequencies are
given in Figure 3.5-4
on the cumulative frequency distribution Sh {g) of the required normal distribution and
the cumulative frequency S}L(g) of the real image. Figure 3.5-6 shows both these cu-
mulative distributions which relate to the frequency distributions of Figure 3.5-4.
The corresponding values g and g in this look-up table are found by the same proce-
dure as used for histogram equalization. The necessary cumulative frequency distri-
bution Sh(g) for the normal distribution can be found in textbooks on statistics and
least squares adjustment. The user can determine the form of the normal distribution
by defining the standard deviation and average value. For digital images with a grey
level spread of 0 - 2 5 5 , values of ± 6 0 for standard deviation and 127 for the average
have proven themselves in practice. When these values are used internally in software,
histogram normalization can take place fully automatically.
Numerical Example (of histogram normalization). Until now the methods have been
explained on the basis of continuous frequency and cumulative frequency distributions.
Using the discrete distributions generated by real images, the target objective can only
be achieved approximately. The limitations are even more significant when the grey
level range is very small. For reasons of clarity this numerical example will be limited
to an image with only 16 grey levels.
The frequency distribution will be stated in percentages. Columns (1) and (2) in Ta-
ble 3.5-1 provide the starting information. The cumulative distribution Sîg) is cal-
culated from the frequencies h(g) by successive addition. Relative frequencies are
shown in Figure 3.5-7 and cumulative frequencies in Figure 3.5-8. After histogram
normalization a relative frequency h(g)tMg is expected which corresponds to a normal
distribution. Such a normal distribution, in this case with standard deviation of ±3,
is shown in column (4) of Table 3.5-1 and has been drawn in Figure 3.5-7. The cu-
mulative distribution corresponding to this normal distribution is given in column (5)
9 Kg) Sh(g) MfiOtarg Sh{g) targ g 9 m Sh(g)

% % % %
(1) (2) (3) (4) (5) (6) (1) (2) (3)
0 1 1 1 1 0 0 1 1
1 1 2 1 2 1 1 1 2
2 2 4 3 5 2 2 2 4
3 3 7 4 9 3 3 3 7
4 6 13 7 16 4 4 13 20
5 7 20 9 25 5 5 7 27
6 7 27 12 37 6 6 14 41
7 8 35 13 50 7 7 13 54
8 6 41 13 63 8 8 9 63
9 6 47 12 75 9 9 10 73
10 7 54 9 84 10 10 8 81
11 9 63 7 91 11 11 9 90
12 10 73 4 95 12 12 0 90
13 8 81 3 98 13 13 0 90
14 9 90 1 99 14 14 0 90
15 10 100 1 100 15 15 10 100
100 100 100
Table 3.5-1: Histogram normalization Table 3.5-2: Result of

histogram normalization
(1) and (2) given grey levels with relative frequencies
(3) cumulative frequencies calculated from (2)
(4) required relative frequency
(5) cumulative frequencies calculated from (4)
(6) which together with (1) form the look-up table
of Table 3.5-1. Figure 3.5-8 is the corresponding diagram. Following the procedure
sketched in Figure 3.5-6, Figure 3.5-8 shows how, for a grey level g — 9, the corre-
sponding target grey level is found (and in this example g = 7). The entire look-up
table is given in columns (1) and (6) of Table 3.5-1. Using this look-up table the new
image can be created. The result of the histogram normalization, expressed as rela-
tive frequencies h(g) and cumulative frequencies Sh(g)is presented in Table 3.5-2 and
Figures 3.5-9 and 3.5-10.
sh(g) target
h(g) h
(9)targe, i n %
20 ••
01 2 3 4 5 6 7 8 9 101112131415 01
I I2 I3 I4 I5 6I 7I Ί8 9 I 101112131415
I I I I ! I I—- g
Figure 3.5-7: Given relative frequen- Figure 3.5-8: Given cumulative fre-
cies and target relative frequencies quencies and required cumulative fre-
(dotted line) for a histogram normal- quencies (shaded) of a histogram nor-
ization malization, together with the construc-
tion of the look-up table
h(g) h(g)«™«in% s„(g) Sh(g).ame. i n %

100··
20-
50--
01 2 3 4 5 6 7 8 9 101112131415
-t-t
01 2 3 4 5 6 7 8 9 101112131415
Figure 3.5-9: Result of histogram nor- Figure 3.5-10: Result of histogram

malization (relative frequencies) com- normalization (cumulative frequen-
pared with target result cies) compared with target result
Exercise 3.5-1. P e r f o r m a histogram equalization on the data contained in columns (1)

and (2) of Table 3.5-1. (Solution: F o r g i v e n grey levels g = 0 , 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 1 0 ,
1 1 , 1 2 , 1 3 , 1 4 , 1 5 the corresponding grey levels g = 0 , 0 , 0 , 0 , 1 , 2 , 3 , 5 , 6 , 7 , 8 , 9 , 1 1 , 1 2 ,
13,15). Draw the corresponding diagram. (Details of the solution can be found in
Section C 1.3.1.1, Volume 2).
C o m m e n t on the result:
• the objective of histogram normalization is only approximately achieved with

digital images
• histogram normalization of digital images normally involves loss of information.

The original pixels with grey levels g = 4 , 5 are converted to grey level g = 4,
grey levels g — 7 , 8 to g = 6 and grey levels g = 9 , 1 0 to g = 7.
• there are normally gaps in the grey level sequence after histogram normalization
of digital images. There are no pixels with grey levels g = 1 2 , 1 3 , 1 4
In spite of the negative comments, histogram normalization plays a significant role in

practice, particularly when the starting images have a very skewed frequency distribu-
tion. An impressive example f r o m practice is presented further below (Figure 3.5-12).
Because of the disadvantages, which occur with both histogram normalization and
equalization, often only the simpler method of expansion, commonly called stretching,
is employed. Here the given grey level range is mapped onto the available grey level
range by distributing the original values, without grouping, evenly across the available
grey level scale. Naturally this also produces gaps in the grey level sequence. An ex-
pansion is only possible when the frequency distribution of the given image has places
where there are no grey levels, which often occurs at the beginning or end of the dis-
tribution. Isolated grey levels, particularly near the boundaries of the distribution, can
also be grouped into a single grey level in order to create "space" for the expansion.
An expansion method implemented in many image processing systems first calculates
the average value m and standard deviation σ of the grey levels. The following assign-
ments are then made:
• average value m (given image) corresponds to 127 (new image).
• grey level range — 2σ to m (given image) corresponds to 1 to 127 (new image).
• grey level range m to + 2 σ (given image) corresponds to 127 to 255 (new image).
Grey levels outside the range m ± 2σ are ignored. If there are too many in this class,
the factor 2 can be replaced by a larger one.
The following section explains two further contrast and brightness enhancements typi-
cally employed in photogrammetry.
3.5.1.3 Compensation for light fall-off from centre to edge of image
Light fall-off follows a cos™ τ law (n lies between 1.5 and 2.5; τ is the off-axis angle,
see Section 3.1.6). If there is no graduated filter in front of the objective lens at the
time of exposure, light fall-off from image centre to edge is significant (see numerical
example in Section 3.1.6).
A relative contrast and brightness adjustment is discussed next, which derives its trans-
formation parameters directly from the image content. For this purpose a reasonably
large circular patch surrounded by concentric annular rings is defined at the centre of
the image. Within the central area and one of the surrounding rings, several image
regions are located which should have equal grey levels (for example, similar ground
types). Grey levels gR\ng and ^centre for the corresponding regions are then related by
means of a regression line; the result provides the transformation function (3.5-1) for
the annular ring under evaluation.
Figure 3.5-11: Transformation function to compensate for light fall-off from image
centre to image edge
Figure 3.5-11 shows grey levels for 3 corresponding image regions, used to define the
regression line. With the aid of the transformation function, every grey level f/Rmg
within a ring can be converted to a grey level ^centre which is free of the effects of light
fall-off.
This procedure is executed for each annular ring. Parameters c and d of the corre-
sponding transformation function (3.5-1) can be further modified by the physically
based cos™ τ law.
Exercise 3.5-2. The following grey levels for two corresponding image regions at the
image centre and outer ring are found in the original image:
5Ring ^Centre
Open fields on flat ground 25 50
Highway 120 200
Derive the transformation function and the corrected grey level for a pixel at the image
edge with a value of g = 100. (Solution: gcentre = 168.)
3.5.1.4 Histogram normalization with additional contrast enhancement
The second example is related to electronic film copying with contrast control, pre-
sented in Section 3.2.2.8. In particular, the copying devices increase contrast in local
areas. During digital image enhancement it is also the case that such a local improve-
ment in contrast should take place together with a global improvement.
In digital photogrammetry, histogram normalization with additional contrast enhance-
ment is used in place of a copying device providing contrast improvement. Figure
3.5-12a shows a digitized aerial film positive with a very limited contrast range. A his-
togram normalization is first performed. Figure 3.5-12b shows the result which makes
use of a normal distribution with a standard deviation of ± 6 0 and average value of 127.
Figure 3.5-13a shows the frequency distribution h(g) before histogram normalization
and the corresponding distribution h(g) after normalization. The transformation func-
tion Fi for this global contrast adjustment is shown in Figure 3.5-13b.
For additional local contrast enhancement, Kalliany 40 has developed a filter whose
effectiveness has been proven in use at the Vienna University of Technology's Insti-
tute for Photogrammetry and Remote Sensing (Institut für Photogrammetrie und Fern-
erkundung) 41 . Here grey levels i n a 3 x 3 o r 5 x 5 neighbourhood are averaged and the
average value m compared with the grey value g of the central pixel. The difference
is transformed by function F2 (Figure 3.5-13b). The steep rise at the beginning of the
curve has the effect of a contrast enhancement. The remaining part of curve F 2 is not
of significance; it would effect an attenuation in contrast if there were very large differ-
ences between the average value and the grey level of the central pixel. The enhanced
difference, with appropriate sign, is added to the average value and assigned to the
central pixel. Figure 3.5-12c shows part of Figure 3.5-12b which has been enhanced in
this way (result in Figure 3.5-12d).
Both modifications to contrast, global and local, do not need to be executed sequen-
tially; they can be carried out in a single processing step by the following operator:
g = Fx{m + (sign(5 - m)F2\g - m|)) (3.5-4)
g ... grey level of the central pixel in the original image matrix
m ... average value of the grey levels i n a 3 x 3 o r 5 x 5 neighbourhood
Fi ... global transform function from histogram normalization
F2 ... local transform function which can be selected externally
g ... grey level of the result for the central pixel
It is appropriate to indicate the size of the region within which the grey levels are
averaged: in small scale aerial photos with short-range changes in texture, a 3 χ 3
region should be chosen; at medium to large scales a 5 χ 5 region has proven to work.
Figure 3.5-12d was generated using a 5 χ 5 region.
40
Ecker, R„ Kalliany, R., Otepka, G.: In Fritsch/Hobby (Eds.): Photogrammetric Week '93. Wich-
mann Verlag, pp. 143-155, 1993. The examples in this section are taken from this publication.
41
A filter with a similar effect has been developed by Wallis and, amongst other places, published in:
Pratt, W.: Digital Image Processing. A Wiley-Interscience publication, New York, 1991.
a) photo as delivered by photo scanner (see F i g u r e 3.5-13(a) f o r histogram)

b) result of histogram normalization with t r a n s f o r m function F\ in F i g u r e 3.5-13
(b) (see Figure 3.5-13 (a) f o r histogram)
c) part of Figure 3.5-12b, before applying a local contrast operator
d) result of local contrast e n h a n c e m e n t with transform function F i of Fig-
ure 3.5-13(b)
Figure 3.5-12
255
255 0
(a) Histogram h(g) of Figure 3.5-12a (b) Transform function F\ for global contrast
and histogram h(g) after histogram enhancement and function Fi for local con-
normalization (result shown in Fig- trast enhancement
ure 3.5-12b)
Figure 3.5-13
3.5.2 Filtering
The purpose of filtering is to separate the wanted from the unwanted. Undesirable
image content, for example, is the noise from the electronic recording sensors (Sec-
tion 3.3.5.3). However, filtering is also employed in order to enhance indistinct image
information. Human observers, for example appreciate strongly pronounced edges in
an image.
Filtering can take place in the spatial domain (Section 3.5.2.1) or in the frequency
domain (Section 3.5.2.2).
3.5.2.1 Filtering in the spatial domain
A very simple filter is the moving average. In this case the grey levels gij of the new
image G are derived from an averaging of the grey levels g^ in the original image
matrix G. It is formulated as follows:
j η η
~9ii =
(2n + l) 2
Σ Σ 9i+k,j+i η >1 (3.5-5)
Numerical Example (for η = 1). Given matrix G (Figure 3.5-14, left). By application
of Equation (3.5-5) it is possible to obtain, for example, image element g22'
922 = ^ ( 1 0 + 1 4 + 2 + 1 1 + 13 + 10 + 9 + 25 + 8) = 11
G G
I
10 14 2 11 12 I
11 13 10 9 9 ; 11 11 9
1
9 25 8 7 10 1 13 13 11
11 13 14 15 13 1 14 15 14
1
1
12 18 17 19 20
Figure 3.5-14: Moving average
By means of a filter matrix W , Equation (3.5-5) can be alternatively written:

τι η
9ij = Σ 9i-k,j-lWn+\+k,n+l+l (3.5-6)

k=—n l= — n
Equation (3.5-6) describes an operation known as convolution and the filter matrix W
is therefore known as a convolution operator. Equation (3.5-6) can be expressed more
simply as:
G = G *W (3.5-7)
For the moving average, the convolution operator or filter matrix W is:
( w\\ W\2 W\3 \
W21 «722 TÜ23 J = - I 1 1 1 J

J / 1 1 1 \
(3.5-8)
1 1 1
W31 u>32 W33 / V /
The central pixels can be assigned a higher weight in the filter matrix W which creates
a general arithmetic mean from the neighbouring grey levels. For example, such a
matrix known as a binomial filter can be formed with the aid of binomial coefficients:
W - 1 ( 2 4 2 )
When the binomial filter is enlarged, it approximates to the Gaussian filter. This name
expresses the fact that the elements of filter matrix W correspond to a two-dimensional
Gaussian bell curve. It should also be noted that the sum of all elements Wy always
equals unity.
Exercise 3.5-3. Filter Figure 3.5-14, left, using operator (3.5-9). Comment on the grey
levels in the new image compared with the grey levels after convolution by means of a
moving average (Figure 3.5-14, right). (Solution: e.g. 322 = 13.)
Comment: The negative signs in front of k and I in Equation (3.5-6) have the effect of
opposing directions for the indices in filter matrix W and image matrix G. However,
with symmetrical filter matrices such as the moving average, this is of no importance.
Filtering with a moving average results in an attenuation of (high frequency) noise. Ad-
mittedly, (high frequency) image information, i.e. edges, is also attenuated. The mov-
ing average with operator (3.5-8) has reduced the grey level difference in Figure 3.5-14
from 18 ( = 25 - 7) to 6 ( = 15 - 9).
The opposite effect, i.e. an increase in grey level differences, can be achieved with a
Laplacian operator. It is obtained through a double differencing of neighbouring grey
levels:
First difference in ξ direction:
:
9i+\,j — 9i,j (3.5-10a)
First difference in η direction:
dg\
= gi j+l
8η) ' ~ 9i'j (3.5-10b)
Second difference in ξ direction:

= (3
S I . ( I ) , - ( 9 , , , - · · ' ' ' · ' · ' ' · · ' ·5"π)
Second difference in η direction:
Finally by addition of (3.5-11) and (3.5-12):
v2 =
9i,j (ßj^J + ( f ^ ) 9i+i,j + 9i-i,j + 9i,j+i +9i,j-1 ~49i,j (3.5-13)
However, the algorithm (3.5-13) can also be interpreted as a convolution with the fol-
lowing operator:
/0 1 0
W = ( 1 -4 1 I (3.5-14)
VO 1 0
This operator (V 2 ) is called the Laplacian operator. A related operator which also
includes the diagonals of the convolution matrix is:
(3.5-15)
The convolution of the photo in Figure 3.5-15 with this operator produces Figure 3.5-16
as a result. Both the operator (3.5-15) and the Laplacian operator (3.5-14) generate
positive and negative pixel values; prior to visualization they must therefore be shifted
into "positive".
The Laplacian operator, and all others of similar construction, extract the edges in an
image. This property is of great importance for automated image processing and will
be discussed in further detail in Sections 6.8.3.4 and 6.8.7.2. Subtracting the image
of extracted edges from the original image produces an image with enhanced edges
(details can be found in Section C 1.3.3, Volume 2). Figure 3.5-17 shows the result of
this subtraction, i.e. Figure 3.5-15 minus Figure 3.5-16.
In practice the intermediate result of an image of extracted edges can be avoided and
the original image directly convolved with the following operator:
(3.5-16)
If operator (3.5-15) is used in place of the Laplacian operator (3.5-14) in the combina-
tion operator (3.5-16), the following operator for an edge enhanced image is obtained:
(3.5-17)
Note: the elements of operators which enhance edges have a sum equal to unity; ele-
ments of operators which extract edges have a sum equal to zero.
Exercise 3.5-4. Create the convolution of Figure 3.5-14, left, with operator (3.5-16).
Comment on the grey level differences in the new image compared with the grey level
differences after convolution with a moving average (Figure 3.5-14, right).
3.5.2.2 Filtering in the frequency domain
Filtering in the frequency domain starts with a spectral analysis. Here the grey level
image is reduced to its component waves, characterized by amplitude and wavelength
or frequency. The grey levels gij in the original image represent a function in the spatial
domain, dependent on indices i and j or image coordinates ξ and η. By means of a
spectral analysis, achieved through a Fourier transformation, the amplitude spectrum,
dependent on frequency / , is obtained.
Figure 3.5-15: Original photo Figure 3.5-16: Convolution of Figure

3.5-15 with operator (3.5-15)
Figure 3.5-17: Photo with enhanced

edges
The relationship between spatial and frequency domains will be demonstrated with a
simple example. Imagine a (continuous) grey level profile g(ry), consisting of three
wave components g\(η) = 50sin(27r0.0177), <72(»7) = 16.7 sin(27r 0.03 77) and 53(77) =
10 sin(2?r 0.05 η). The three waves are shown in Figure 3.5-18, left. The sum of these
component waves is the grey level profile g(η), represented in the spatial domain of the
cube (Figure 3.5-18, right). The related frequency domain shows the three component
waves with their amplitudes C as a function of frequency / .
Figure 3.5-19 illustrates the filtering in the frequency domain. By means of a Fourier
transform, the original image is converted into its amplitude spectrum C ( f ) . Using
a freely chosen modulation transfer function MTF(f), the amplitudes can be trans-
formed into new amplitudes C (see also Equation (3.1-26)). The result is a filtered
Figure 3.5-19: Filtering in the frequency domain

amplitude spectrum C(f) from which the filtered image can be generated by means of
an inverse Fourier transform. The equation for the inverse Fourier transformation has
already been given (Equation (3.3-2)).
In the filtered image of Figure 3.5-19, the amplitude with the smallest frequency ( / =
0.01) was somewhat attenuated, the amplitude with the mid frequency ( / = 0.03) more
strongly attenuated and the component wave with the highest frequency was elimi-
nated. Since the high frequency parts of an image generally contain the noise, this can
therefore be reduced by using the modulation transfer function in Figure 3.5-19. In this
regard it is usual to refer to a low-pass filter which suppresses the high frequencies and
allows the low frequencies to pass through.
The opposite is a high-pass filter. This can be employed to reverse the effects of a re-
duction in contrast caused by a detector (Section 3.3.2). For this purpose the reciprocal
of the function (3.3-5), i.e. 1 /MTF, must be used for filtering in the frequency domain.
Exercise 3.5-5. For filtering in the frequency domain, sketch the modulation transfer
function which removes the contrast reduction caused by image recording with a detec-
tor. The transfer function causing contrast reduction during image recording is shown
in Figure 3.3-5.
Filtering in the frequency domain can be converted into filtering in the spatial domain
and vice versa. A simple comparison indicates:
• averaging (Equations (3.5-8) and (3.5-9)) correspond to low-pass filtering.

• edge-enhancement operators (3.5-16) and (3.5-17) correspond to high-pass fil-
tering.
Further reading: Albertz, J., Zelianeos, K.: Phia, pp. 161-174, 1990. Gonzalez, R.C.:
Digital Image Processing. Prentice Hall, 2002. Schenk, Τ.: Digital Photogrammetry,
Volume 1. TerraScience, 1999.
3.6 Image pyramids/data compression

Image pyramids enable efficient management of image data (Section 3.6.1) and com-
pression methods reduce the amount of image data (Section 3.6.2).
3.6.1 Image pyramids

Figure 3.6-1 shows an image pyramid. The geometric resolution in the rows and
columns is reduced by a factor of 2 from level to level. The pixels at the next lower res-
olution can be determined in various ways, the simplest method being the elimination
of every second row and column of the current resolution. However, in order to keep
the information loss smaller when changing to a coarser resolution, the grey levels at
the coarser resolution are often derived from the grey levels of the finer resolution by
means of a Gaussian or binomial filter, e.g. using the filter matrix (3.5-9).
Section 3.6 Image pyramids/data compression 129
Figure 3.6-1: Image pyramid with 512 χ 512, 256 χ 256, 128 χ 128 and 64 χ 64 pixels
Exercise 3.6-1. Create two reduced images from Figure 3.5-14, left, one by eliminat-
ing every second row and column and the other by application of the binomial filter
(3.5-9).
Compared with the original image, the storage space required, for all pyramid levels,
increases by a third.
Exercise 3.6-2. Prove this statement with an image pyramid based on an original image
with 512 χ 512 pixels.
This additional storage requirement offers, however, many advantages which will be
evaluated later in further discussions of metric image analysis.
Further reading: Ackermann, F., Hahn, Μ.: in Ebner, Η., Fritsch, D., Heipke, C.: Dig-
ital Photogrammetric Systems, pp. 43-58, Wichmann Verlag, 1991. Li, M.: PE&RS
57, pp. 1039-1047, 1991. Pan, J.-J., Li, S.-T.: IAPR XXIX(B2), pp. 474-478, Wash-
ington, 1992. Kropatsch, W., Bischof, H„ Englert, R.: in Kropatsch, W„ Bischof, H.:
Digital image analysis. Selected techniques and applications, pp. 211-230, Springer,
2000.
3.6.2 Image compression

A reminder about the storage requirement for high resolution, metric images serves as
an introduction to this section. According to Section 3.4.1, the storage requirement
for a 23 cm χ 23 cm metric image with 7 μνη χ 7 μιτι pixel size is around 1 GB. The
processing of a large array of images, which can consist of several hundred, would
necessitate a storage capacity which current hard disks do not offer in medium priced
computers. Data transfer between hard disk and memory, and over a computer network,
would also be too time-consuming.
It is for this reason that compression methods are in demand. Here it is necessary to
distinguish between lossless and lossy compression methods. With a lossless method
the original image can be reconstructed from the compressed image without loss of
information; with lossy methods a certain loss of information occurs during recon-
struction, which may also be called decompression. As would be expected, with lossy
compression methods a significantly higher compression factor can be achieved. The
compression factor is the ratio of the number of bits in the original to the number of
bits in the compressed image.
The method of run-length encoding is a lossless compression. It utilizes the redundancy
between neighbouring pixels. Here, for example row by row, successive pixels with
identical grey levels are grouped together and only the one representative grey level,
together with the repetition number, is stored. Run-length encoding achieves only a
small compression factor with photographs. In small-scale photographs there is almost
no reduction in storage, in large-scale photographs, particularly for close range work,
a compression factor of 3 : 1 can be achieved in the best case.
Lossy compression methods are considerably more interesting. Here a small loss of
information can be accepted in return for high compression. The JPEG compression
method 42 is very well known and can reduce photographic storage to a third without
appreciable information loss. In JPEG compression the digital image is broken down
into tiles of 8 χ 8 pixels. By means of a Discrete Cosine transform (DCT, closely re-
lated to the Fourier transform) there is a transition from the spatial to the frequency
domain within each tile (Section 3.5.2.2). The amplitude spectra are then low-pass
filtered, and saved in compressed form. The degree to which the amplitude spectra are
filtered is controlled by the "quality factor" which also influences the loss which oc-
curs on reverse transformation into the spatial domain. Loss of grey level information
occurs with JPEG compression, as well as losses and shifts of grey level edges, which
is very important from a photogrammetric perspective. Figure 3.6-2 illustrates such
losses and shifts. On the left are edges extracted from the original image (perhaps with
operator (3.5-15)). In the middle are edges from a decompressed JPEG image which
was made from an original image compressed by a factor 10. As a result of image
shifts caused by this procedure, smoothed features appear, as well as artifacts which
develop in particular at the edges of the independently filtered 8 x 8 tiles.
The disadvantages of JPEG compression can be compensated, to some extent, by use of

a wavelet transform instead, allowing a more flexible sub-division of the original image
based on frequency properties 43 . The high quality of wavelet-based compression can
be seen in the extracted edges shown on the right of Figure 3.6-2. Although the same
compression factor as in the JPEG compression has been chosen, there are hardly any
shifts or artifacts compared with the original image. In practice there are a number of
42
Joint Photographic Experts Group, h t t p : / / w w w . j p e g . o r g /
""Literature: Pennebaker, W.B., Mitchell, J.L.: JPEG Still Image Data Compression Standard.
Springer, 1993. Usevitch, B.E.: A Tutorial on Modern Lossy Wavelet Image Compression: Founda-
tions of JPEG 2000. IEEE Signal Processing Magazine, Volume 18(2), pp. 22-35, 2001.
Section 3.7 Aerial cameras and their use in practice 131
from the original image from a decompressed im- from a decompressed im-
age which had previously age which had previously
been compressed by a fac- been compressed by a fac-
tor of 10 using the JPEG tor of 10 by a wavelet-
algorithm based method
Figure 3.6-2: Extracted edges 44
wavelet-based compression methods in use and the JPE group has also announced the
use of such a method.
A totally different approach to compression lies in extracting relevant image informa-
tion rather than storing the grey values. Where images contain small, elliptical object
features as a result, for example, of applying circular or concentric ring targets to object
surfaces, then the relevant information, in this case the image coordinates of the target
centre points, should be extracted as soon as possible, for example using algorithm
(6.8-21). In this way, the image matrix is reduced to the image coordinates of the criti-
cal features. This high reduction method is currently offered by cameras equipped with
the corresponding "intelligence" (microcomputer and software, also known as "smart
cameras").
3.7 Aerial cameras and their use in practice

Before discussing these cameras in detail, a section on flight planning is appropriate.
3.7.1 Flight planning

The main task of aerial surveying is the three-dimensional recording of natural and
man-made features on the ground. It is a requirement of stereophotogrammetry that
every point on the ground is imaged in at least two metric photographs (Section 2.1.3).
This requirement is met when the images within a strip of photographs overlap by
50%. In single image photogrammetry (Section 2.1.4 and, in particular, Chapter 7) it
is sufficient when every point on the ground appears in only one metric image.
'"Taken from Schiewe, J.: PFG 1998(1), pp. 17-25.

Our attentions will now be turned to the stereophotogrammetry for which metric cam-
eras with a square image format are employed45. The area to be mapped is imaged
in a sequence of photographs forming parallel strips which together create a block of
photographs. The geometrical relationships (Formulae (3.7-1)) involved in the flight
planning are illustrated in Figure 3.7-1 for flat ground.
Figure 3.7-1: The geometry of flight planning for flat ground46
In Figure 3.7-1 and in Formulae (3.7-1), symbols have the following meaning:
A . . Distance between flight lines/strips

Β . . Base (between consecutive images)
c . . Principal distance
s . . Image side (to edge!)
h . Flying height above ground
Ζ . Ground height
ν . . Flying speed over ground
L . Length of a strip/block
Q · . Width of block
45
Although current aerial cameras such as the UltraCam or DMC provide rectangular images, in the
following text only square images are considered for reasons of simplicity.
46
Taken from Albertz/Kreiling: Photogrammetric Guide, Wichmann, 1980.
Photo scale number rot = h/c

Image side in ground units S — srrib
Base in photograph b — B/rrn,
Flying height above ground h = crrii,
Absolute flying height Z0 — h + Ζ
Forward overlap (%) I= 100 = ( 1 - ^ ) 100

s V s
A
Side overlap (%) q= -100 = ( 1 - ~ I 100
S
Ground area of one photograph Fb = S 2

= 2
s ml
l
b
Baselength for 1% forward overlap Β = S I 1 —

100,
Distance between strips for A = S —^
10
°; (3.7-1)
% side overlap
Number of models in a strip nm = + lj

47
(length L)
Number of photographs in a nb = nm + 1
strip
Number of strips in a block ns — + 1J

(width Q) 47
Area of a stereoscopic model Fm = (S — B)S
New area for each model in a Fn = AB
block
Β [ml
Time between photographs At [s] = . . > 2.0
υ [m/s]
(see Section 3.7.2)
Comments:
a) every pair of overlapping images in a strip should correspond to the "normal

case" for stereopairs (Section 2.1.5). In practice, this "normal case" is never ex-
actly achieved in aerial surveying and it is necessary to deal with a near "normal
case". Individual images do not have their camera axes exactly vertical but only
near vertical. Assuming there are no extensive navigation aids in use and no sta-
bilization devices for the camera mounting, these departures from ideal amount
to the following in practice:
47
[ J = Largest integer number
• rotation about the longitudinal (fore and aft) axis of the aircraft (roll angle)
±5gon (4.5°),
• rotation about the transverse (wing tip to wing tip) axis (pitch angle) ± 3 gon
(2.7°),
• deviation from nominal heading (yaw angle): ±15 gon (13.5°),
• variation in flying height ±2%,
• deviation from flight path around ±200 m.
b) even for an ideal flight plan, a forward overlap I — 50% and a side overlap
q = 0% would not be sufficient. To process the entire block of photographs
(see references to aerotriangulation in Chapter 5), neighbouring strips, as well
as stereomodels within a strip, must overlap to some extent. For this reason,
and also because of the departures from an ideal imaging configuration listed in
paragraph (a), a forward overlap I = 60% and a side overlap q = 25 - 30% are
planned.
For single image photogrammetry, equal forward and side overlaps are chosen
with I = q = 25- 30% (Section 2.1.4 and, in particular, Chapter 7).
c) Figure 3.7-1 and Formulae (3.7-1) assume flat ground. Where there are large
height differences, the recommended forward and side overlaps should be de-
signed for the highest ground. At lower heights the overlaps are then larger. For
flight planning over mountainous terrain, good knowledge of height variations is
necessary.
Exercise 3.7-1. Take a topographic map, 1 : 50000, of a mountainous region and draw
in the edges of a strip flown in a direction that crosses the ridges and valleys roughly
at right angles. Smallest photo scale 1 : 30000, c = 150 mm. Then draw the axis of a
second strip, parallel to the first, which has a minimum side overlap of 30%. Draw, in
a different colour, the edges of this second strip and finally measure the maximum side
overlap of the two strips.
Exercise 3.7-2. What overlap results when photographs are flown along a course down
the centre of a U-shaped valley when the forward overlap at the edge of the strip is
to be 60%? Flying height above the valley floor is 1200 m, the highest points to be
photographed lie 400 m above the valley floor, and the principal distance is 152 mm.
(Solution: 73.3%.)
d) the required flying height and photo scale depend on the required resolution on
the ground (Sections 3.1.5, 3.2.2.7, 3.3.2) and the required accuracy of the pho-
togrammetric end product (Sections 4.6 and 6.7).
e) for the various applications in photogrammetry, metric cameras with different

objective lenses are required, i.e. lenses with different fields of view or principal
distances. The lenses vary from normal to super wide angle and their parame-
ters are listed in Table 3.7-1. The standard format for aerial film cameras is
23 cm χ 23 cm. The base-to-height ratios, stereomodel areas and flying height
Angle
Normal Intermediate Wide Super wide
Field of view: Diagonal 62gon 85gon lOOgon 140gon
(56°) (77°) (90°) (126°)
Field of view: Image side 47gon 64gon 83gon 117gon
(42°) (58°) (75°) (105°)
Principal Distance c [cm] 30 21 15
Ratio of Principal Distance 1 :1 2 :3 1 :2 1 :4

to image diagonal
Base-to-height ratio for 1:3 1:2 2:3 1:1

60% forward overlap
Stereomodel area (in A) yl/4 A/2 2.9 A

(h = const.)
Flying height above ground 2z 1.5z 0.6 ζ

(in z), area = const,
scale = const.
Single image photogrammetry
High mountain regions
Interpretation
Orthophotos
Cities
Applications and 5verview flights
Height accuracy
evaluation criteria Lower flight costs
Stereophotogrammetry
(Table 2.1 -1 and Section 4.6)
Table 3.7-1: Most common objective lenses used in aerial surveying (standard image
format 23 cm χ 23 cm, h = flying height above ground)
above ground shown in the table have been derived with the aid of this format,
the various principal distances and the relationships in (3.7-1). The lower part
of the table indicates the applications in which cameras with longer principal
distances and those with shorter principal distances tend to be used.
Exercise 3.7-3. Check the fields of view given in Table 3.7-1 using the image format
and the corresponding principal distances.
f) the depth-of-field problem (Section 3.1.4) does not exist with aerial photographs
due to the relatively large flying height. Aerial metric cameras are generally fo-
cused at 1000 m, i.e. according to Equation (3.1-4) flying heights between 500 m
and infinity lie within the acceptable depth of field.
Exercise 3.7-4. An aircraft flies with a metric camera at an absolute flying height of
3000 m. The ground height varies between 500 m and 1200 m. State the greatest image
blur (diameter of cricle of confusion) for exposures with a principal distance of 15 cm
and an aperture of 11. (Solution: 1.2 μπι.)
g) if defined map sheet edges must be taken into account within a strip, there are
two possible solutions: use of aimed single images or dense image sequences.
For flights requiring aimed single images (also called a point flight) very good
navigational aids are required in order to fly to, and "hit", the selected target
area. GPS (Global Positioning System) significantly eases the problem of a point
flight. Flights which adopt a dense image sequence, from which the best placed
images can be selected, require a forward overlap of around 90%.
In addition to the flight parameters mentioned above, particular attention shoud be

given to flying speed υ when flight planning with digital line cameras. To ensure that
individual image lines can be sequenced together without gaps, the following condition
must be met:
ν = Axfs (3.7-2)
ν ... flying speed
Δχ ... detector dimension projected onto the ground (Equation (3.3-1))
fs ... scan frequency.
Numerical Example. A digital line camera with a scan frequency of 750 Hz (750 lines
are read out per second), a pixel size of 6 /im and a focal length principal distance)
of 80 mm, should be flown at the following speed when at a flying height of 2 km:
V = 7 5 0 = 112
^ Μ Γ - 5 m / s - 400km/h
In practice a speed of around 300 km/h would be set in order to deal with any pitch an-
gle changes (paragraph a) above) and to achieve some overlap of pixels in the direction
of flight (see Section 3.4.1). A digital line camera (Figure 3.3-3) provides records for
single image photogrammetry. In contrast, the 3-line camera provides images suitable
for stereophotogrammetry. The elements of a 3-line camera can be chosen so that, for
example, the combination of the a and b line correspond to a wide angle view and the
combination of the a and c lines, or c and b lines, correspond to a normal angle of view
(Figure 3.3-3, right).
As part of flight planning, a navigation plan is produced. This can either be an enlarged
aerial photograph or, better, a good topographic map. It contains:
• the area of interest without any rounding off, i.e. the region which absolutely
must be recorded, normally stereoscopically
• obstacles in the flight path

• restricted areas which on no account must be overflown (e.g. army training

grounds, foreign countries)
The region marked on the navigation plan must be sufficiently large to accomodate a 5
to 10 km extension at both ends of the strips which allows for navigation manoeuvres.
At the end of the strip an incremented numbering, absolute flying height and heading
(in sexagesimal degrees) is added. An additional sheet contains the following infor-
mation, for each strip if required: project identification, purpose, date; rrib, c, Z0,1, q;
minimum film requirement, film type; organizational details such as agreed targeting
of points (control points, new points, etc.), proximity of international borders, flights
only in overcast weather, etc.
Standards exist for quality control of survey flights, e.g. DIN standards: Schwebel, R.:
PFG 2001(1), pp. 3 9 ^ 4 .
3.7.2 Metric aerial cameras

Large format, aerial film cameras, which are presently most commonly used in prac-
tice, will be discussed in more detail in the first Section 3.7.2.1. Digital aerial cam-
eras, which record the illumination intensity electronically using a CCD area array,
are then discussed in Section 3.7.2.2. Finally, digital 3-line cameras are presented in
Section 3.7.2.3.
3.7.2.1 Large format, metric film cameras
Figure 3.7-2 is a schematic diagram of the most important components of an aerial

film camera. They are designed for serial photography in which a sequence of related
Eyepiece
Film magazine Film-transport indicator
Vacuum
Overlap control
Pressure platen
Viewfinder Unperforated rollfilm,
telescope
24cm wide
Coupling for
drive ana vacuum Lens cone and
camera frame
Control instrument
and switches
pîE Drive unit with control
electronics
c 28 Volt DC
Lens
J Camera mount
' —
\ d Earthing cable
Vibration dampers
Aircraft floor
Shutter Filter Aperture Port cover
Figure 3.7-2: Schematic diagram of an aerial film camera

metric images is recorded. (German information sources may refer to these as "Reihen-
messkameras" (RMK), i.e. serial metric cameras.) A film cassette contains around 400
exposures. Above the port is the camera mount which, in the simplest versions, can
be levelled manually by footscrews and has rubber or steel spring fixings for vibration
damping. The mount carries the drive unit which can be rotated about a vertical axis
and which contains all the elements required for operating the imaging cycle. The lens
cone, containing the required lens type, is inserted in the drive unit. The film magazine
is placed on the focal-plane frame of the lens cone and contains the pressure platen and
the mechanisms for creating the vacuum and pressing the platen against the focal-plane
frame.
After every exposure command, the imaging cycle runs automatically as follows:
• expose.
• in modern cameras, shift the image plane during exposure (image motion com-
pensation, to be discussed in more detail in Section 3.7.4).
• raise the pressure platen and release the vacuum.
• transport the film and advance the photo number counter.
• apply the vacuum again (also known as pneumatic flattening).
• press the platen against the photo frame plane.
The camera is now ready for another exposure and awaits the next exposure command.
The operator can generate this manually by pressing a button for "aimed single pho-
tographs" or it can be generated automatically for serial photographs. The imaging
cycle requires about 1.6 to 2.0 s, a time which defines the shortest possible interval
between photographs.
In the viewfinder the camera operator sees the ground moving past. Combined with
the viewfinder telescope is the overlap control which sets the required forward overlap
when the set of moving curved lines appear to move at the same speed as the ground. If
the lines move faster than the image of the ground, the operator must slow them down
manually, or accelerate them if they are moving too slowly. When the lines move at the
same speed as the image of the ground, the photographs will be released at the proper
intervals to generate the required forward overlap.
The drift angle is also adjusted via the viewfinder. A ground point imaged on the central
line must appear to move along this line if the camera is correctly oriented. If it moves
off at an angle a to this line, the camera must be rotated by this angle a and the aircraft
heading correspondingly adjusted (Figure 3.7-4) to ensure the correct ground track.
Principal distance, to ensure

use of the correct image frame
Image frame
Side overlap 20%
Side overlap 30%
Warning mark for aimed photographs
Forward overlap 60% (beginning of base)
Level bubble
Flight direction
Moving lines of the overlap regulator
Direction of motion of the moving lines
Figure 3.7-3: Viewfinder image (Wild RC10)
The following information can also be seen in the viewfinder (Figure 3.7-3):
• the image frame for the selected objective lens
• the base end points, for example for 60% forward overlap
• the side overlap, for example for 20% and 30%
• horizontal control by means of an image of the level bubble
• warning mark for aimed single photographs
The requirement for a photogrammetric image to be close to a central projection en-

sures the need for a central shutter instead of a focal-plane shutter and a very short
exposure time ( 1 / 2 5 0 - 1/1000s). Figure 3.7-5 shows a central shutter used by the
manufacturer Zeiss.
Shorter exposure times require a compensating increase of light. Therefore, during
exposure time the shutter needs to be fully open as long as possible. The condi-
tions are indicated in Figure 3.7-6. The required exposure time is determined by an
exposure meter (Section 3.2.2.3), taking into account to the sensitivity of the photo-
graphic material and the prior choice of aperture. The reverse procedure is sometimes
required, i.e. the exposure time is specified and, from the exposure reading the cam-
era system determines the appropriate aperture. The moment of exposure is the mean
time: tm = 0.5 (ίο +1\)· If precise positioning and orientation sensors are built into the
Required track Course to be set
\
a)
κ=0 κ= -a 0
a) If the aircraft flies along its nom- b) The course is corrected by the an-
inal course, a side wind will drive it gle κ = - a . The camera is rotated
along a different course or track over the by κ = + a . The true ground track of
ground. The drift angle is determined the aircraft now lies in the required di-
by checking the apparent motion of a rection and a selected ground point will
selected point on the ground. appear to move along the central line.
Figure 3.7-4: Principle of drift correction
aircraft (Section 3.7.3), then the moment of exposure tm must be synchronized with
these. In a metric camera, the flash exposure of the fiducial marks must be made at the
moment of exposure tm.
An analogue metric image not only contains the image itself in the 23 cm χ 23 cm
format but an auxiliary image displaying extensive additional data (Figure 3.7-7) which
can be supplied in analogue or digital form. The auxiliary image or, more accurately,
the data it contains, can be divided into two groups. One contains strictly necessary
information, the other contains optional data.
Strictly necessary auxiliary information:
• the number of symmetrically distributed fiducial marks should be 8 (previously

4 were common) in order to support sufficient redundancy for the transformation
of fiducial marks. Identifiers for fiducial marks, such as numbers, are necessary
in order to identify them automatically in the digitized images.
Figure 3.7-5: Example of a rotary shutter (Zeiss Oberkochen), left closed, right open.
A is the drive. The lamellae 1 rotate faster than lamella 2. The sword lamella 3,
controlled by the overlap regulator, is swung aside relatively slowly by the eccentric
4 while lamella 2 is not yet in the open position. Lamella 2 opens while lamellae 1
are still closed, before opening very rapidly. The speed of rotation of the lamellae
determines the exposure time. The closing procedure is the same in reverse.
Aperture opening
100%
t] Exposure time
Figure 3.7-6: The shutter efficiency is an optimum when the cross-hatched area is as
small as possible, i.e. the shutter is fully open as long as possible. The shutter must
therefore be opened and closed as fast as possible.
1) eight fiducial marks (with number) 10) data of exterior orientation

2) ninth, asymmetric fiducial mark 11) overlap
3) brief name of project (data pad) 12) photo scale
4) date (data pad) 13) circular level
5) time 14) coarse height indicator
6) photo number 15) exposure time, aperture
7) camera (lens) number 16) image motion (length)
8) principal distance 17) grey wedge
9) magazine number (data pad)
Figure 3.7-7: Auxiliary data imaged with a photograph; left: digital, right: analogue
miniplots of instruments
• if the fiducial marks do not have identifiers then a ninth, asymmetrically placed
fiducial mark, is required in order to identify automatically the fiducial marks in
a digitized image, as well as to determine automatically the orientation of the
image 48 (laterally correct, mirror reversed, rotated). If the additional mark is
missing then the level bubble or centre of the clock (Figure 3.7-7) can be used
instead.
• the principal distance for re-creating the interior orientation and a possible deter-
mination of scale
• the date, camera number and film cassette or pressure platen number in order to
have a link to both the flight plan and the calibration certificate
• project name and incrementing image number for managing larger image blocks
48
Ellenbeck, K.H., Waldhäusl, P.: BuL 52, pp. 70-71, 1984.
Optional auxiliary information:
• grey wedge or colour chart for contrast control and colour adjustment during
photographic post-processing and for brightness and contrast enhancement, as
well as colour adjustment of digitized metric images
• current exposure data: aperture, exposure time and image motion compensation
(Section 3.7.4)
• clock to record moment of exposure so that the agreed (midday) period for flying
can be checked and, for example, the direction of north in the image can be
determined from time of day and direction of shadows
• information relevant to exterior orientation:
- level bubble to detect a poorly levelled aircraft. Centrifugal forces will also,
of course, affect the bubble.
- Coarse height measurement of absolute flying height to an accuracy of

around ± 5 0 m. Combined with a coarse ground height, this enables an
average image scale to be derived.
- GPS position and IMU orientation which are only available in expensive
multi-sensor systems (Section 3.7.3). Since post-processing is generally
required to provide good quality values, they maybe exposed on film later
on.
Leica Geosystems (formerly Wild, LH Systems) currently offers the RC30 camera
system (Figure 3.7-8). Technical details include: Interchangeable lens cones with
principal distances of 8.8, 15 and 30cm, largest aperture of / / 4 , 100Lp/mm reso-
lution, exposure times between 1/1000 and 1/100s. For further details: h t t p : / /
w w w . l e i c a - g e o s y s t e m s . c o m / c o r p o r a t e / e n / n d e f / l g s _ 5 7 6 3 2 . htm
(November 2006).
Intergraph (formerly: Z/I Imaging, Carl Zeiss Jena, Carl Zeiss Oberkochen) currently
offers the RMK TOP camera system (Figure 3.7-9). RMK TOP15 (15 cm princi-
pal distance, aperture between / / 4 and / / 2 2 ) , RMK30 (30 cm principal distance,
aperture between f/5.6 and / / 2 2 ) , Objective lens distortion < 3 /xrn, rotary shut-
ter (Figure 3.7-5), exposure times between 1/500 and l / 5 0 s , two different circu-
lar graduated filters (Section 3.1.6) with a transmission in the centre of 35% and
60%, 8 numbered fiducial marks, shortest interval between photographs 1.5 s. Further
details: h t t p : / / w w w . i n t e r g r a p h . c o m / r m k t o p a s c s / d e f a u l t . a s p (No-
vember 2006).
Both camera systems can be combined with GPS and IMU (Section 3.7.3). In addition
there are components for image motion compensation (Section 3.7.4).
Figure 3.7-8: RC30 with viewfinder

(Figure 3.7-3) from Leica Geosystems Figure 3.7-9: RMK TOP from Intergraph
3.7.2.2 Digital cameras with CCD area sensors
There are now aerial cameras with CCD area sensors which significantly fulfil the
quality requirements demanded by photogrammetry. One problem is the number of
pixels which are currently available in arrays up to 7K χ 4K; aerial photogrammetry
requires a multiple of this (Section 3.3.1). A second problem is the read-out time: this
must lie within the shortest interval between photographs, between 1 and 2 s.
With its DMC (Digital Mapping Camera), Intergraph has solved these problems by
mounting four CCD cameras in the camera housing in place of a single CCD camera,
each with 7K χ 4K pixels and synchronous exposure. The four images overlap only
to a limited extent (Figure 3.7-10, top). During post-processing a single metric image
is created with around 13500 (normal to flight path) χ 8000 (along flight path) pixels.
The pixel size is 12 μτη and field of view is 82gon (74°) (normal to flight path) and
49gon (45°) (along flight path). In other words, the DMC corresponds to a wide angle
camera normal to the flight path and to a normal angle camera along the flight path
(Table 3.7-1). Image recording uses 12 bits per pixel.
Exercise 3.7-5. The following data are redundant: principal distance, field of view,
pixel size and number of pixels. Consider to what extent the DMC data are free of
contradiction.
In addition to the four panchromatic camera heads, the DMC has three cameras for
the spectral regions blue, green and red, and a fourth camera for near infra-red (Sec-
tion 3.3.5.2). These last four cameras have a resolution which is worse by a factor of
around 4 compared with the panchromatic camera, i.e. approximately the same area is
recorded by a multi-spectral image (3000 χ 2000 pixel) as is covered by 4 images in
*· flight path
Figure 3.7-10: The DMC from Intergraph (top) and UltraCam-D from Microsoft-
Vexcel (bottom)
panchromatic mode (Figure 3.7-10, top).

Like the RMK TOP (Figure 3.7-9), the DMC can be coupled with a GPS and IMU
(Section 3.7.3). The DMC likewise has a facility for image motion compensation (Sec-
tion 3.7.4.2).
Further reading and information: Hinz, Α., Dörstel, C., Heier, Η.: IAPRS ΧΧΧΙΙΙ(Β2),
pp. 164-167, 2000. h t t p : / /www. i n t e r g r a p h . c o m / r m k t o p a s c s / d e f a u l t .
a s p (November 2006).
Microsoft-Vexcel has also presented a digital photogrammetric camera, named the
UltraCam-D, with very similar technical features to the DMC. In addition it has the
following specifications:
• pixel size on the ground at a flying height of 300 m: 3 cm (at this extremely
low altitude the system requires 50 pixel for image motion compensation, Sec-
tion 3.7.4.2)
• minimum interval between images: 0.75 s
• geometric accuracy in image plane: < ± 2 μιη
A prominent difference between the UltraCam-D and D M C is that the four panchro-
matic CCD cameras are arranged in a row along the direction of flight and are not
synchronously exposed. Exposures are delayed such that the perspective centres of
the four images are approximately at the same location in the ground coordinate sys-
tem ("syntopic exposure"). Each camera has one, two or four CCD area arrays in the
image plane, arranged such that the final composite image is built up of 3 χ 3 tiles
(Figure 3.7-10, bottom). A further difference between the cameras is that the objec-
tive lenses of the four panchromatic cameras in the UltraCam-D each cover the entire
image format of 11500 (normal to flight path) χ 7500 (along flight path) pixels. In
contrast, each lens in the DMC covers only the format of its corresponding image tile
of 7K χ 4K pixels (Figure 3.7-10, top). The second camera in the UltraCam-D cluster,
with 4 CCD area arrays (one in each corner), defines the geometric extent of the final
composite image, created by post-processing from the nine image tiles each with 4K
χ 2.7K pixels (Figure 3.7-10, bottom).
Further reading and information: Leberl, F., Gruber, Μ., Ponticelli, Μ., Bernoegger,
S., Perko, R.: ASPRS Meeting, Anchorage, Alaska, 2003. www. v e x c e l . com (No-
vember 2006).
3.7.2.3 Digital 3-line cameras
The concept of the digital 3-line camera (Figure 3.3-3, right) is due to Hofmann 49 .
Digital 3-line cameras were initially designed for space missions 50 and were then
adapted for aerial use, in some cases requiring significant additional development work.
The HRSC-A (High Resolution Stereo Camera-Airborne) from the German Aerospace
Centre (Deutsches Zentrum für Luft- und Raumfahrt (DLR)) is an adopted model for
airborne sensing and has been successfully used for photogrammetric sensing 51 .
The following text will deal in more detail with the ADS (Airborne Digital Sensor)
from Leica Geosystems. In contrast to other multispectral line cameras, the ADS40
uses a sensor arrangement called tetrachroid which splits up the incoming light beam
into the four spectral regions of blue, green, red and near infra-red. By this principle,
the exact co-registration of all colour ranges from one object detail can be guaranteed.
There is no need for any colour fusion or resolution merging for generating true colour
and colour infra-red images. In addition to the tetrachroid, three panchromatic line
sensors (covering the range from blue to red) are responsible for the forward (40°),
backward (16°) and nadir looking data acquisition. Two models are available: the first
model has one nadir looking tetrachroid sensor, which is mainly intended for colour
orthophoto production. The second model has one nadir looking and one backwards
looking tetrachroid for more universal applications including colour stereo compila-
tion. The nadir looking panchromatic line is complemented by a second line staggered
by half a pixel. This reduces the sampling interval which leads to a more reliable re-
construction of object frequencies. All sensor lines consist of 12000 individual sensor
elements.
The combination with GPS and IMU is a valuable addition to digital cameras with
CCD area arrays, but is not strictly necessary. However, with digital line cameras such
as the ADS, such a combination is extremely important (Section 3.7.3). ADS with
image motion compensation will be considered in Section 3.7.4.3.
49
Hofmann, O.: BuL54(3),pp. 105-121, 1986.
50
Wewel, F., Schölten, F., Neukum, G., Albertz, J.: PFG 1998(6), pp. 337-348. Reulke, R„ Scheele,
Μ.: PFG 1998(3), pp. 157-163. Sandau, R„ Eckhardt, Α.: IAPRS XXXI(Bl), pp. 170-175, 1996.
5
'Neukum, G.: in Fritsch/Spiller: Photogrammetric Week '99, Wichmann Verlag, pp. 83-88, 1999.
Further reading and information: Flicker, P., Sandau, R., Walker, S.: OEEPE-Pub-
lication No. 37, pp. 81-90, 1999. Sandau, R. et al.: IAPR 33(B1), pp. 258-265,
2000. http://www.leica-geosystems.com/corporate/en/ndef/
l g s . 5 7 627 . h t m (November 2006).
3.7.3 Satellite positioning and inertial systems

Use of the well established GPS (Global Positioning System) 52 for navigation purposes
during flying missions will be discussed first. For this purpose a modest accuracy is
sufficient. In the subsequent sections the use of satellite and inertial systems for exterior
orientation control will be presented.
3.7.3.1 Use of GPS during photogrammetric flying missions and image exposure
With regard to its original concept, GPS can be advantageously used for navigating
photogrammetric flying missions. An additional application for either aerial film cam-
eras or CCD area array cameras makes use of the GPS to release the shutter for individ-
ual exposures. For this purpose the aircraft's position must be continuously monitored
in real time by the GPS and compared with the planned positions of the perspective
centres given in the coordinate system of the GPS system currently in use. On the ba-
sis of this comparison, individual exposures can be made under computer control. Such
an application of GPS can provide aimed single images (Section 3.7.1). The overlap
regulator (Section 3.7.2) is then largely redundant.
For this application, GPS is used in kinematic mode with the C/A (Coarse/Acquisition)
Code (wavelength = 293.1 m). In this mode, the distance of the GPS receiver from the
satellite is determined by electronic distance measurement using the transmission time
t of the signal and its known speed of propagation c = 299762 km/s. The C/A code
is repeated after one millisecond (ms) and unambiguous range measurement with this
code is possible up to 300 km.
The method of trilateration applied to the ranges measured simultaneously from the
aircraft receiver to three GPS satellites, enables the calculation of receiver coordinates
in WGS84 (World Geodetic System 1984), in which GPS satellite parameters are cur-
rently provided. Since the receiver clock cannot be exactly synchronized with the
satellite clocks, an unknown time difference At must be taken into account. It is there-
fore necessary to measure range to at least four GPS satellites simultaneously. (Ranges
which incorporate the unknown time difference are known as pseudo ranges.)
The position of individual imaging locations, initially provided in the geocentric, Carte-
sian coordinate system of WGS84, can be converted into geographical coordinates φ
and Λ and ellipsoid height h, for example using the formulae given in Section Β 5.3.1,
Volume 2. In modern aerial survey systems, the geographical coordinates φ and λ and
52
In addition to the American GPS there is also the comparable Russian GLONASS. GALILEO is a
third satellite positioning system currently being deployed by the European Union. A general term for
such systems is GNSS (Global Navigation Satellite Systems). In this textbook "GPS" is also used as a
synonym for these other systems.
ellipsoid height h are recorded during the flight onto the individual metric images (see
Figure 3.7-7). Accuracy is at the C/A level, i.e. some tens of metres. The accuracy of
GPS-supported navigation of the flying mission lies around ±50 m 53 . If it is assumed
that, on average, an acceptable deviation between planned and actual imaging location
is ± 5 mm in the image, then this accuracy is sufficient for flying missions which deliver
an image scale smaller than 1 : 10000 (— 5/50000). More accurate navigation is only
required for flights delivering very large image scales.
The navigational accuracy of GPS in kinematic mode can be raised to a few metres
when differential GPS (DGPS) is employed. In this case two receivers are required:
one receiver remains at a (fixed) reference station whose coordiantes are known; the
other receiver is located in the aircraft. Both receivers are positioned absolutely us-
ing the C/A code. At the (fixed) reference station corrections can be calculated and
transmitted by telemetry in real time to the receiver in the aircraft. The function of a
fixed reference station is being increasingly taken over by a network of permanently
installed GPS stations.
3.7.3.2 Accurate determination of exterior orientation elements by GPS and

IMU
The coordinates of the imaging locations can be determined by GPS and the angles
which define the attitude of a given image can be determined by IMU (inertial mea-
surement unit). Nowadays with DGPS a positional accuracy of around a decimetre can
be achieved "on the fly". Instead of time-of-flight measurement with the C/A code, a
phase comparison technique using the relatively short wavelengths of both GPS carrier
waves is used (LI = 0.1905 m, L2 = 0.2445 m). With phase measurement only the
residual part of a single wavelength is determined and some multiple of the full wave-
length (0.1905 m or 0.2445 m) must be added to this in order to calculate the actual
range.
The GPS literature list at the end of this section provides several methods to eliminate
the ambiguity. After solving for the ambiguity, the GPS receiver counts the increase or
decrease of integer multiples of the wavelength compared with a starting value. The
measured distances between a GPS receiver and a GPS transmitter are strongly dis-
torted by atmospheric influences. These error sources have a similar effect on neigh-
bouring receivers such that the differences between the receivers are largely free of
the influences of the (systematic) errors. The differences between the original distance
equations provide the components of a vector between both receivers. Provided one
receiver is located at a point with known coordinates (the reference station) then the
vector, applied to this station, provides the coordinates of the second receiver which is
located in the aircraft.
Cycle Slips and the Multipath effect can, however, interrupt the almost continuous
determination of the vector from the (stationary) reference station to the GPS antenna
in the aircraft. Bridging techniques are required which will be discussed below.
"Becker, R„ Barriere, J.: PE&RS 59, pp. 1659-1665, 1993. Arnold, H., Schroth, R.: in
Fritsch/Hobbie (Eds.): Photogrammetric Week '93, pp. 53-62, Wichmann Verlag, Karlsruhe, 1993.
Exact positioning of the GPS antennae in the aircraft takes place approximately every
second. Depending on aircraft speed, this corresponds to separations in position of
around 50 to 100 m. These nodal points can be connected by a spatial curve 54 . A
spatial flight path supported by GPS points is illustrated in Figure 3.7-11. Time acts as a
parameter along this curve. If the individual exposures follow this time scale, normally
GPS time (t m in Section 3.7.2), then by interpolation it is possible to determine the
coordinates of the antenna tips Ai at the times when the images are exposed. This
synchronization is possible within about 1 ms, i.e. at an aircraft speed of 250 km/h the
error is some 7 cm.
Figure 3.7-11: Photo strips with GPS positioning (Af. interpolated positions of an-
tenna tips, Poi: perspective centres)
For the transfer to the coordinates of the perspective centres Poi the angular attitude
of the sensor, e.g. the aerial film camera, is required at the moment of exposure (Fig-
ure 3.7-11). These three orientation parameters are also of considerable interest in
completing all six elements of exterior orientation for every image.
However, GPS only provides the positions of the imaging locations Ai and, from the
known time intervals between exposures, the speed along the flight path of the GPS
antenna mounted in the aircraft. Other measurement systems, which provide the posi-
tion as well as angular orientation of the sensors are (I)nertial (N)avigation (S)ystems
(INS) 55 . These make use of the inertia of a mass in relation to its acceleration ä. Ac-
cording to Newton's second law of motion, force / is the product of mass m and
acceleration ä, i.e. / = mä. (For clarification: a = distance, da/dt = ά = ν = speed,
dv/dt = ä = acceleration.)
Figure 3.7-12 shows one of the measurement principles. A force, acting along the axis
of a cylinder to which a known mass is connected through an elastic spring, causes a
54
Lancaster, P., Salkauskas, K.: Curve and Surface Fitting. Academic Press Ltd., 1986.
55
We distinguish between INS and IMU. While the INS is a complete navigation system, the IMU is
just one of its components which measures acceleration and attitude changes.
f Spring A
vJUUU ΛUOülUb
Mass /
lllllllllllllllllllllllllllllllllll
Read-out
Figure 3.7-12: Deflection method of measuring acceleration
change in position of the mass and a resultant deflection on a scale also attached to the
cylinder. This deflection is a measure of the corresponding acceleration ä of the mass.
From the acceleration determined in this way, speed is obtained by integration over
time; a further integration over time provides the distance or, effectively, position.
In order to record the movement of a mobile vehicle, in this case an aircraft, it is

necessary to have three accelerometers, arranged orthogonally to one another (Fig-
ure 3.7-13). Three accelerometers are insufficient to completely determine the move-
ment of a vehicle and directional information is also required. This is determined
continuously by three gyroscopes (gyros), one mounted on each axis of rotation (Fig-
ure 3.7-13). Nowadays optical gyroscopes are mostly used. An optical gyroscope with
a simple constructional principle is illustrated in Figure 3.7-14.
Two contra-rotating monochromatic beams of light are introduced into a circular op-
tical waveguide. If the device has not rotated then, after the beams have traversed the
waveguide, they arrive simultaneously at their starting point, i.e. without any phase
difference. However, if the circular waveguide rotates with angular velocity ά, then
the beam travelling in the direction of rotation has a greater distance to cover to reach
the starting point which has now moved ahead (Figure 3.7-14, left); for the beam trav-
elling in the opposite direction the path is shortened (Figure 3.7-14, right). When the
beams are brought together after completing the circular path, a relative phase shift Αφ
is detected from which the angular velocity can be determined as follows: 56
(3.7-3)
Co ... speed of light in vacuum: 299792 km/s

Λ . . . wavelength of light beam, e.g. 0.6 μπι
F ... circular area enclosed by the light guide (fibre optic) which can be
multiplied by a factor η through the use of η windings of the fiber
56
Taken from: Schwarz, W.: Schriftenreihe des Deutschen Vereins für Vermessungswesen, Volume 22,
pp. 54-97, 1996. This publication provides a comprehensive presentation of kinematic sensors.
Yb
Figure 3.7-13: Body coordinate axes for a strapdown inertial navigation system
Rotation in Rotation in
same direction opposite direction
Figure 3.7-14: Optical gyroscope for an intertial system
By integration of the angular velocity ά, for example over the time difference (£, — tg),
the angle of rotation a can be determined (Figure 3.7-14).
Table 3.7-2 shows the three currently available accuracy classifications for inertial nav-
igation systems57. The mid accuracy class is appropriate for determining the exterior
57
Schwarz. K.-P.: in Fritsch, D„ Hobbie, D.: Photogrammetric Week'95, pp. 139-153, 1995. The
stated tilt accuracies relate to roll and pitch; yaw angle (heading) is a factor 3 to 5 less accurate. At the
Photogrammetric Week 2001 the following accuracies, achieved during practical tests, were reported:
GPS positioning between ± 5 and ± 3 0 cm, IMU orientation for roll and pitch angle between ± 3 and
± 5 mgon (10" and 16", resp.) and for the heading angle ± 8 mgon (26").
orientation of aerial photographs and is reasonably priced. Over short time periods it
provides positional accuracy at the decimeter level. An aircraft's tilt can only be deter-
mined by an IMU. Over a period of one minute, an accuracy of 5 mgon (15") can be
achieved which is sufficient for standard requirements. At a flying height of 2000 m,
this corresponds to a planimetric error of 15 cm on the ground.
Time interval System accuracy

high medium low
Position
Is 0.01-0.02 m 0.03-0.1m 0.3-0.5 m
1 min 0.3-0.5 m 0.5-3 m 30-50m
lh 300-500 m 1 - 3 k m 200-300 km
Tilt
Is < 1" 10"-15" 30"-2'
1 min 1" - 2 " 15"-20" 10'-20'
lh 10"-30" 30"-3' 1° - 3 °
Price (US$) « 1000000 » 100000 » 10000
Table 3.7-2: Three accuracy classes for inertial navigation systems
As can be seen from Table 3.7-2, an IMU's positional accuracy is only acceptable
over short time periods. A Position and Orientation System (POS) therefore combines
GPS with IMU. GPS supplies position and speed and supports the IMU. The IMU
provides direction and tilt and densities the relatively large separations between GPS
measurements which, as already mentioned, are between 50 m and 100 m at a recording
interval of 1 s. In contrast, the IMU's recording interval is around 0.01 s (100 Hz). Short
interruptions in GPS recording (cycle slips, etc.) can also be bridged with the IMU.
Figure 3.7-15 shows the arrangement of the most important sensors in a POS, together
with an imaging sensor. Great care must be taken to determine the relative positions
of the three illustrated sensor modules. Details of a corresponding calibration can be
found in specialist literature58. The relationships are relatively simple if all the sensors
are fixed to the aircraft:59
• from the GPS coordinates of the antenna tips Ai, the perspective centres Pot can
be calculated along the spatial directions delivered by the IMU (Figure 3.7-11).
• if the coordinate axes of the imaging sensor and IMU are strictly parallel then
the rotation angles of the IMU directly correspond to the rotation angles of the
image sensor60.
58
E.g. Shin, E„ El-Sheimy, N.: ZfV 127(1), pp. 4 1 ^ 9 , 2002. Heipke, Ch. et al.: OEEPE Publication
No 43, 2002.
59
If the inertial system is fixed to the aircraft or vehicle, it is known as a strapdown system, as shown
in Figure 3.7-13.
60
This type of configuration cannot generally be made to a sufficient accuracy. Deviations from the
Figure 3.7-15: Imaging sensor together with GPS and IMU in a Position and Orienta-
tion System (POS)
The relationships are considerably more complicated when there is continuous banking
to compensate for the drift angle and, on the basis of the IMU gyroscope, the imaging
sensor platform is continually tilted, in other words stabilized. In this case the rota-
tions must be recorded with timing information in order to establish the relationships
between all three sensor groups during post-processing. This procedure is greatly sim-
plified if at least the imaging sensor and IMU are fixed to the same tilting platform so
that their relative positions remain unchanged throughout the entire flying mission.
The POS data are valuable for processing aerial images taken with either film or CCD
area array cameras but their importance is not so great as when processing images
taken with digital line cameras. POS data are also essential for airborne laser scanners,
discussed in Section 8.1. This will be discussed in more detail in Section 3.7.3.3.
A position and orientation system is marketed by Applanix under the name POS/AV.
Further reading and information: Lithopoulos, E.: in Fritsch/Spiller: Photogrammetric
Week '99, Wichmann Verlag, pp. 53-57, 1999. www. a p p l a n i x . com. The German
company IGI offers a comparable system (e-mail: i n f o @ i g i - c c n s . com).
3.7.3.3 Gyro-stabilized platforms and particular features of line cameras and

laser scanners
On flying missions with a digital line camera (Section 3.3.1) and/or a laser scanner
(Section 8.1), aircraft navigation along the planned route will also, of course, employ
GPS in a similar way to a flying mission with film or CCD area array cameras. This
usage was discussed in Section 3.7.3.1.
ideal are known as misalignments. The transformation of IMU rotations into rotations of the imaging
sensor or into the coordinate system of the image processing software generally involves the successive
multiplication of several rotation matrices.
Gyroscopes provide rotation angles as a function of time. Using this information, an

image sensor's platform can be continually stabilized. Compensation can therefore be
made for the considerable changes in roll, pitch and yaw angles (Section 3.7.1) dur-
ing a flight. Such gyro-stabilized platforms are available for the large format, aerial
film camears mentioned at the end of Section 3.7.2.1 and the CCD area array cameras
discussed in Section 3.7.2.2. However, the mechanical and electronic complexity as-
sociated with a gyro-stabilized platform does not, in general, justify their use for such
cameras. Roll, pitch and yaw angles can either be extracted from POS information
during photogrammetric processing or, if a POS was not employed, by means of aero-
triangulation (Chapter 5). This is possible because with these sensors the recording of
a single image effectively takes place at a single point in time 61 .
The situation is different with dynamic image recording, as is the case when using a
single or 3-line camera or a laser scanner. Images taken without the use of a gyro-
stabilized platform show large distortions as a result of the flight dynamics, and these
are much worse with line cameras than with laser scanners. Certainly such images can
be processed with the aid of POS information. However, conditions for processing are
significantly improved if line cameras, and possibly also laser scanners, are mounted
on gyro-stabilized platforms. The IMU should also be mounted on this platform. An
accurate IMU need then only record small "residual" tilts, which is a considerable
advantage in the design and construction of such a device.
However, the constant rotations of the gyro-stabilized platform continually alter the
relationship between the GPS measurements and the imaging sensor or laser scanner,
as can be appreciated in Figure 3.7-15. In order to provide a continuous "eccentricity
vector" between GPS antenna and the reference point of the imaging sensor, continuous
recording of the gyro-stabilized platform rotations is necessary. A separate gyroscope
system is required to stabilize the mounting platform and this must be:
• fixed to the aircraft
• able to measure large rotations
• less accurate than the gyroscopes in the IMU which is fixed to the imaging sensor
In summary it can be said that the exterior orientation of a line camera alters from im-
aged line to imaged line and that therefore the direct recording of the six orientation el-
ements by GPS and IMU is obligatory. For laser scanners, the exterior orientation even
alters for each range measurement. With film and CCD area array cameras the POS
information is, in contrast, only an interesting option, as will be seen in Section 4.1. A
gyro-stabilized platform is here not strictly necessary; however, it improves the qual-
ity of the image and makes processing easier, particularly for dynamically recorded
images.
6l
More exactly: during exposure there should be no large shifts or rotations of the camera. To some
very reasonable extent, a gyro-stabilized platform improves image quality, particularly for longer expo-
sure times (Section 3.7.4.1).
Further reading for Section 3.7.3 (selected examples only): Hofmann-Wellenhof, Β.,
Lichtenegger, Η., Collins, J.: GPS—Theory and Practice. Springer, 1997. Hutton, J.,
Lithopoulos, E.: PFG 1998(6), pp. 363-370.
3.7.4 Image motion and its compensation

The blur which an optical system produces in an image arises from the following
sources:
• from the spread in depth of a three-dimensional object, which cannot be sharply

focused onto a single image plane (Section 3.1.4)
• from the wave nature of light, which results in diffraction at the aperture stop
(Section 3.1.5.1)
• from the resolving power of the optics (Section 3.1.5.2)
• from the resolution of the emulsion in film cameras (Section 3.2.2.7)
• from the size of the opto-electronic sensor elements and their geometric arrange-
ment in digital cameras (Section 3.3.2) and film scanners (Section 3.4.1)
Further blurring of the image occurs as a result of image motion of the object or camera
during the time the shutter is open. The following discussion concentrates on aerial
photogrammetry.
The theoretical image motion u t h of an object in the image plane is (Figure 3.7-16):
c vt _ , „
uth = vt- = — (3.7-4a)
η rrib
10 3 υί
= — (3.7-4b)
i.orrib
ν = velocity in [km/h] u = image motion in [mm]

t = exposure time in [s] = image scale number
During the exposure time t the image moves over the image plane and the energy in
the illumination is distributed. Only in places where sufficient light energy falls on
the chemical or electronic sensor will the image become visible. A single object point
becomes a "line" in the image. If the chemical or electronic sensors could react to the
smallest possible level of illumination energy, then the length of the "lines" would be
given by Equation (3.7-4a). In fact, the sensors react only to a certain minimum energy,
so that the "lines" in the image are shortened. (When a camera is subject to vibration,
particularly in a helicopter, double images can appear.) On the basis of current film
Image plane
Object Vt
Figure 3.7-16: Theoretical image motion
sensitivities used in photogrammetry, the "shortened" image movement u amounts to

around half the theoretical image movement:
140vt ν
u « 0.5uth = = 140ci- (3.7-5)
mb h
c in [mm] υ in [km/h] t in [s]

h in [m] u in [/im]
Numerical Example (to estimate the image motion).

from linear motion from rotation
Velocity ν 300 km/h 2gon/s = 170 km/h
Exposure time t l/300s 1/300s
Principal distance c 150 mm 150 mm
Flying height h 1500 m 1500 m
Image scale 1 : m^ 1 :10000 1 : 10000
Theoretical image motion uth 28 μπι 16 μπι
Image motion u visible in practice 14 μιη 8 //m
The blurring of the image caused by image motion significantly reduces the resolu-
tion. Meier 62 shows that a theoretical image motion of 1.5 times the reciprocal of the
resolution is just tolerable in practice.
u -1 (3.7-6)
th,ma\ < 1-5R
62
Meier, H.K.: BuL, pp. 65-77, 1960.
Numerical Example (of the maximum permissible exposure time for which the image
motion remains tolerably small).
Given: R = 80Lp/mm
Equation (3.7-6): Uth = 19/im
Equation (3.7-5): u « 10 μτη
Given: V = 300 km/h, c = 150 mm, h
=> mb = 10000
Equation (3.7-5): tmax = 1/440 s « 1/500 s
Whether an exposure time of 1 /500 s is acceptable depends upon the overall sensitivity
of the film to be used and the object brightness (season, time of day, weather, light
intensity...).
Exercise 3.7-6. What is the largest acceptable image scale for a flying speed of
300 km/h and an exposure time of 1/300 s (R = 40Lp/mm)? (Solution: 1 : rri,m-m —
1 : 7400.)
Compensation for image motion is particularly required at low flying heights and large
image scales. There are a number of solutions.
3.7.4.1 Compensation of image motion in aerial film cameras
In modern aerial cameras, the image motion due to the aircraft's forward movement can
be compensated during exposure by a computer controlled displacement of the film's
pressure plate with a velocity of v' = uth/t = vc/h (see Equation (3.7-4a)). This is
known as Forward Motion Compensation (FMC).
This type of motion compensation eliminates the most significant part of the image
blurring, but it does not eliminate the effects of:
• the three rotations of the aircraft; gyro-stabilized camera mounts are necessary
to compensate for this effect (see also the ends of Sections 3.7.3.2 and 3.7.3.3).
• the variations of flying speed, unless the actual flying speed ν for image motion
compensation is taken on line from the overlap control (Figure 3.7-3), or the
navigation system.
• differences of ground heights. To compensate for these, the image motion com-
pensation must be at least partially separated from the overlap control system.
The overlap must be controlled so as to ensure overlap in the highest parts of the
ground (see Section 3.7.1c); the image motion compensation, on the other hand,
must be set to achieve sharp imaging for the most important parts of the ground
(the valleys, up to the average ground height).
Exercise 3.7-7. With the data of the numerical example above, compute the remaining
image motions: for the three rotations of the aircraft of about 2gon/s, for height dif-
ferences of ±500 m and for variations of flying speed of ± 2 0 km/h. (Solution in [μπι]:
uu=uv = 13, uK — 9, uAh = 7, uAv = 1.)
Exercise 3.7-8. What is the loss in resolution of the film camera introduced in the
numerical example of Section 3.2.2.7 (RT = 67Lp/mm, which according to Equa-
tion (3.2-11) derives from R0 — 9 0 L p / m m and RP = 100Lp/mm), if a theoretical
image motion of u t h = 28 μπι (see first numerical example of this section) is taken
into account? (Solution: From Equation (3.7-6), which only has an empirical basis:
-RimageMotion = 5 4 L p / m m ; applying this to an expanded version of Equation (3.2-11):
RT — 4 2 L p / m m ; there is therefore a reduction in resolution from 6 7 L p / m m to
42Lp/mm.)
Exercise 3.7-9. What is the deterioration in the modulation transfer function MTF(f)
of the full system introduced in Figure 3.2-11 when the above image motion u t h =
28 μπι is taken into account?
Tip: As modulation transfer function for MTF(f)imageMotion use the sine function
(3.3-5) with values MTF(0)imageMotion = 1 and MTF(54 L p / m m ) I m a g e M o t i o n = 0. To
evaluate Equation (3.2-12) use MTF(f)0 and MTF(f) P from Figure 3.2-11.
Another solution for image motion compensation is to "nod" the camera by a small
angle very quickly during exposure.
Exercise 3.7-10. Calculate the angular velocity of nodding (pitch angle velocity) in
order to compensate for the image motion of 28 μπι in the above numerical example.
(Solution: 3.5gon/s = 3.15°/s).
3.7.4.2 Image motion compensation for digital cameras with CCD area arrays
For as long as the camera shutter is open, an image of the ground moves continuously
across the focal plane in the direction of flight at a speed of w = vc/h. At the moment
the shutter opens to take a photograph (t = 0, say) an image of a particular narrow
strip of ground at right angles to the direction of flight will start to be imaged on the
sensor elements of row η; in general the exposure will be insufficient. At time (t + Δί),
where Δί can be calculated from w and the pixel dimensions, an image of the same
strip of ground will fall on, and start to be recorded on, row (n + 1); again the exposure
will be insufficient. Immediately before this, however, the charge packets from row η
were moved to row (n + 1). This process can be repeated several times, row by row, at
intervals of Δί; the image of that particular strip of ground is therefore integrated over
a number of rows of sensor elements, resulting in an image with an adequate signal-to-
noise ratio. Finally, readout of all rows is initiated; that is to say, readout of the whole
array. In this way, during the period when the shutter is open, the whole image has
been shifted electronically to keep pace with the moving image of the ground, rather
than being shifted mechanically as in analogue cameras.
This method of FMC is called Time Delay Integration. Very high demands are made of
the readout process when the CCD arrays are large since, within exposure times which
are a fraction of a second, all the pixels in the array must be read several times (see
also Section 3.3.1). The limitations on FMC for film cameras mentioned at the end of
Section 3.7.4.2 above also apply to motion compensation with TDI.
The DMC from Intergraph and the UltraCam-D from Microsoft-Vexcel (Section 3.7.2.2)
incorporate TDI for motion compensation.
3.7.4.3 Image motion compensation for digital line cameras
A compensation for image motion, caused by forward movement of the aircraft, would
only be possible with TDI in digital line cameras if every linear array was composed of
several parallel lines of detectors. The ADS from Leica Geosystems has not made use
of this possibility. In consideration of the very high read-out rate of 800 Hz, achieved
with very sensitive detectors, motion compensation at mid to high altitudes is not re-
quired.
3.7.5 Effective illumination in aerial photography
In Section 3.2.2.8 the exposure Η was defined 63 as the illuminance Ε falling on the
emulsion, or opto-electronic detector, times the exposure time t. But what is Ε for an
aerial photograph? Of the total radiation from the sun reaching the atmosphere about
one-third is immediately reflected back into space by the atmosphere. The remainder
is subject to extinction (absorption plus scattering) in the atmosphere, caused by air
molecules and aerosols (suspended particles from 0.01 to 100 μνα in diameter such as
dust, water droplets in haze, fog and clouds, and ice crystals). Short wavelengths are
scattered much more strongly than long wavelengths, hence the blue sky we see. The
remainder of the sun's direct radiation, Sd, (emitted at zenith angle Zi, Figure 3.7-17),
together with the skylight, Ss, form the global radiation which finally illuminates the
ground. The two components of this total incident light on an object vary in proportion
from Sd : Ss — 3 : 1 at a solar altitude of 50° and with a cloudless sky, to 1 : 30 in
light haze.
Photographs can be taken without difficulty under a sky completely covered by a high
layer of cloud. The greater degree of scattering lightens the shadows of objects but also
reduces the contrast (see Section 3.1.5), thus reducing measuring accuracy. In order to
increase the contrast various yellow filters are used according to the haze conditions.
An object on the ground absorbs a part of the total incident light; the remainder is
diffusely reflected from the natural and artificial rough surfaces forming the ground,
some at a zenith angle ZR towards the camera. This reflected component mixes with the
scattered light which does not reach the surface (haze light), and is further scattered by
the atmosphere and aerosols. The strong contrast reduction caused by the total amount
of blue scattered light is minimized by placing a yellow filter in front of the objective.
The 1 2 - 1 5 individual lens elements are also coated with an anti-reflex coating.
63
In this textbook on photogrammetry we use photometric units. In remote sensing, radiometric units
are common. In place of illuminance Ε the irradiadiance is used; in place of luminous intensity I the
radiant intensity is used, etc.
Sun (radiant illuminance I) Illuminance Ε
Figure 3.7-17: Raypath from sun to image plane
The result of the superimposition of haze light on the light reflected from the object is
that contrast decreases with increasing height. Films with greater gradation (7 = 1.4-
1.8) are therefore used for small-scale photographs from greater flying heights.
The non-absorbed, reflected part of the white (sun) light is an important information
carrier. Studies of the spectral reflectance of various types of ground objects show:
a) that the reflected light covers strongly differentiated parts of the spectrum, so
permitting object colours to be differentiated (Figure 3.7-18, upper left);
b) that even at a uniform spectral distribution, the reflected light can vary greatly in
intensity;
c) that the reflectance of vegetation is almost constant in visible light (in the green
part of the spectrum somewhat more than in the blue and red), but that significant
differences in intensity occur in the near infra-red (Figure 3.7-18, lower part) and
for this reason colour infra-red films (Section 3.2.2.5b) or black-and-white infra-
red films (Section 3.2.3) are used for classifying vegetation types.
The object brightness, now defined as the integrated intensity over all wavelengths of
the reflected visible light, can lead to extreme differences of density in the photograph
(Table 3.7-3). The scattered light (Figure 3.7-17) causes, however, a significant reduc-
tion of the overall contrast.
Sand
Concrete Limestone
Clay soil Pine trees
_ Sand
— . rrr Snow
Brown soil 2 o η ~ Gneiss
λ [nm] λ [nm]
400 500 600 700 800 900 1000 400 500 600 700 800 900 1000
Figure 3.7-18: Spectral reflectance of various ground objects as a percentage of the

incident light
Numerical Example. The ratio of object brightnesses of a tarred road and new snow,
from Table 3.7-3, is 8 : 80 = 1 : 10. If we add a haze brightness of 20, the ratio
above is reduced to 28 : 100 = 1 : 3.6. For the same haze brightness, the contrast of
a coniferous forest to new snow is reduced from 3 : 80 = 1 : 27 to 23 : 100 = 1 : 4.3
and the contrast of a coniferous forest to a tarred road from 3 : 8 = 1 : 2.7 to 23 : 28 =
1 : 1.2! (Contrast is defined here as a ratio, see Section 3.1.5.3.)
Coniferous forest 1. . . 3
Water 3
Tarred road 8
Dry meadow 7.. . 14
Wet sand, deciduous forest 18
Yellow, dry sand 31
Light, dry concrete 35
Old snow 42.. .70
New snow 80.. .85
Table 3.7-3: Total reflectance as a percentage of object illuminance
3.7.6 Survey aircraft

Naturally, it is possible to convert almost any type of aircraft for aerial photography
by cutting a port for the camera in the bottom of the fuselage. The selection of a
type must satisfy certain performance characteristics, however. Relatively slow aircraft
are preferred for large-scale photographs (currently permitted minimum flying height
is around 300 m above open ground and around 1 km over densely populated areas).
However, for rapid flight to and from the area of interest and to the required flying
height a fast speed64 and rate of climb are desirable. Since the flights to and from
the area of interest are expensive, it is desirable to complete the whole job in one
flight and therefore the maximum possible flight endurance is demanded. Projects in
developing countries often require the use of short landing strips, i.e the use of STOL
aircrafts (short take-off and landing). Oxygen masks are needed above a height of
4000 m. A pressurized cabin is much more comfortable, if only from the point of
view of temperature. The camera port must then be sealed by an expensive plate of
optical glass which must cause no additional distortion, but which must be capable
of withstanding the stresses caused by pressure and temperature differences (AT can
easily reach 50 Kelvin). An alternative solution here is a remotely operated camera.
The navigation instrument with overlap regulator is then placed inside the cabin over
a smaller sealed port while the camera is outside the pressurized part of the fuselage
above an open port.
Camera ports in the centre of the width of the floor have under them an axially symmet-
ric air cushion which, seen as an optical component, cause less asymmetric distortion
by refraction than do ports at the side of the floor. It must be possible to close the port
at take-off and landing in order to prevent mechanical damage to the port glass or the
camera.
High-wing aircraft give the camera operator a good view downwards, an important aid
in navigation during the approach flight, particularly for large-scale photographs. An
intercom system is essential for communication between crew members. Navigation,
including exposure release, is increasingly done with GPS and for large scale work
with DGPS (Section 3.7.3.1).
Aerial-photography flights in central Europe must take full advantage of every day of
good weather. IFR65 (blind flying) equipment is therefore essential for take-off and
landing through ground mist. The time available for photography is limited not only
by the weather, however, but also by such factors as shadow lengths, which should
usually not exceed 3.5 times the object heights. Flights under full, high cloud cover
are possible in summer; they may even be desirable, for example for photographs of
cities, so as to reduce the loss of detail in shadows. The resulting photographs have
relatively weak contrast, however, and the height accuracy may suffer as a result of
loss of detail contrast66. Sometimes a flight may be expressly required before trees are
in leaf, although stereoscopic acuity in the maze of shadows may be so poor that the
forest floor cannot be measured. These factors, taken together, produce a maximum
flying time, in central Europe, of about 300 flying hours with about 120 flights, per
64
Airspeed is still given in knots (nautical miles per hour). The conversion formula is: lknot =
0.515 m / s = 1.852 km/h.
65
IFR = Instrument Flight Rules
66
Waldhausl, P.: Results of the Vienna Test of OEEPE Commission C. Institute for Applied Geodesy,
Frankfurt, pp. 13-41, 1986.
Section 3.8 Terrestrial metric cameras and their application 163
year and machine. The result is that each hour of photographic flight is relatively
expensive 67 .
3.8 Terrestrial metric cameras and their application

The photographs for terrestrial photogrammetry are usually taken with the cameras
in fixed positions. Photographs at large distances, camera to object, are only used in
special cases, for example for topographic surveys by expeditions and for glaciological
research. In general, the objects are much closer and the camera must therefore be
focused on finite distances and the depth of field (Section 3.1.4) has to be considered.
The most important applications in close-range photogrammetry are:
• architecture and civil engineering (surveys of old and new buildings, documen-
tation of buildings and damage to buildings, etc.)
• computer Aided Facility Management (internal building surveys, etc.)
• conservation of monuments (preserving cultural assets by photographic records,

which can be used at any time for restoration or reconstruction of the monuments)
• archaeology (documentation and surveys of excavations)
• biophotogrammetry (measurements of living creatures)
• forensic photogrammetry (scenes of crimes, most often used for documentation

and reconstruction of the details of traffic accidents)
• industrial photogrammetry (automobile design, production control, assembly

control)
• mobile mapping (surveys along traffic routes)
• computer-controlled vehicle navigation (robotics)
• support and control for medical operations
The "normal case" of terrestrial photogrammetry will be outlined in the following Sec-
tion 3.8.1, after which terrestrial cameras will be presented. This presentation starts
with cameras based on chemical sensors. These fulfill the rigorous conditions of met-
ric cameras but in the concluding discussion the strict demands of photogrammetry
are relaxed and the discussion moves to off-the-shelf (amateur) cameras. Electronic
sensors then take the place of chemical sensors.
67
A review of the more common survey aircraft can be found in Albertz/Kreiling: Photogrammetric
Guide, 4 th ed., Wichmann, Karlsruhe, pp. 112-117, 1989.
3.8.1 "Normal case" of terrestrial photogrammetry

The "normal case" of stereophotogrammetry was presented in Section 2.1.5. In the
"normal case", both camera axes are perpendicular to the base and parallel to one
another. In Section 2.1.5 the object coordinate system is configured for aerial pho-
togrammetry, i.e. the direction of photography is the opposite direction of the Ζ axis
(Figure 2.1-13). In terrestrial photogrammetry, the direction of photography is the di-
rection of the y axis (Figure 3.8-1) 68 .
Figure 3.8-1: The horizontal "normal case"
The horizontal "normal case" is derived from the "normal case" of aerial photogram-
metry (Figure 2.1-13) by substituting 69 ωι = ω2 = lOOgon = 90°. Equations (2.1-32)
and (2.1-34) become, for the "normal case" of terrestrial photogrammetry:
X = y— — B-
c Ρξ
y = B— (3.8-1)
Pi
m Dvi m
ζ = y— = Β — = y—
C Ρξ c
The XYZ coordinates have been replaced by xyz coordinates. This expresses the fact that the
object coordinate system in the normal configuration is a local xyz system which can be arbitrarily
oriented in 3D space, for example with vertical base Β or with tilted camera axes.
69
In order to avoid singularities in the conversion of rotation matrix elements into rotation angles,
the angles a, ν and κ are often used in terrestrial photogrammetry instead of the angles ω, ψ, κ (Sec-
tion Β 3.4.2, Volume 2). Here α (primary) is the azimuth of the camera axis, and ν (secondary) is the
angle between the camera axis and the vertical (nadir distance). In terrestrial photogrammetry the rotation
matrix Β (3.4-6) from Volume 2 is therefore used instead of the rotation matrix of Equation (2.1-13).
= γ (~TTlbß(TPii)) + ( m t
) 2
y2 (3.8-2)
= mb—aPi = — σΡί
/ Tf\ y \^
σ
ζ = y \~^mb~ßaP() +
(mb^v)2
3.8.2 Stereometric cameras

To reconstruct spatial objects at close ranges, it can be a great convenience to obtain the
required two metric images simultaneously using a single camera system known as a
stereometric camera. Such a stereometric camera is illustrated in Figure 3.8-2. In most
systems the base is fixed and very accurately known. A common base length is 120 cm
and these devices are often known as fixed base cameras. When using a fixed base,
which has a metric camera at each end and is configured for the "normal case", the 12
exterior orientation elements of both cameras are known with respect to the local xyz
coordinate system. With this type of photogrammetry, no control points are required, a
considerable advantage in practical use. Control points, or other control elements such
as measured lengths, are recommended only for checking purposes. However, control
Central control unit

Orientation device for aperture, exposure time etc.
Figure 3.8-2: Schematic diagram of a stereometric camera

points are necessary if the photogrammetric results must be supplied in a global XYZ
system rather than a local xyz system.
Despite the great advantages of stereometric cameras, they play a secondary role in
practice. The fixed base has the effect that the average error in the direction of the
camera axes, i.e. in the y direction, increases with the square of the distance (Equa-
tion (3.8-2b)). Stereometric cameras should therefore be available with different base
lengths; the most flexible solution is a rail along which a variable base can be individ-
ually set.
Numerical Example. The accuracy behaviour of stereometric cameras will be ex-

amined on the basis of data from Zeiss and Wild cameras. These stereometric cam-
eras were constructed with base lengths of 120 cm and 40 cm. Evaluation of For-
mula (3.8-2b) results in the following accuracies (c = 60 mm, σρ£ = ±7 μπι):
Base 120 Base 40

y Η cTy [mm] y [in] σ υ [mm]
3 0.9 2 1.2
6 3.5 3 2.6
10 10 6 11
20 39 10 29
Table 3.8-1: Depth accuracy of stereometric cameras with a base of 120 cm and 40 cm
(c = 60 mm, σρ ( = ±7 μπι)
Exercise 3.8-1. The Zeiss and Wild stereometric cameras are fixed focus cameras. The
Zeiss camera with a base of 120 cm is focused at 9 m and the 40 cm base is focused at
4 m. For these cameras calculate the circles of confusion for the distances given in
Table 3.8-1 (aperture setting / / l l ) . For both camera systems also calculate the depth
of field for a maximum circle of confusion of 40 μπι diameter and give recommenda-
tions for the use of the Zeiss stereometric camera with respect to expected accuracies
(Table 3.8-1).
Exercise 3.8-2. Give the 12 elements of exterior orientation for a stereometric cam-
era with respect to the local xyz system. (Solution: Corresponds to details given on
Figure 2.1-13.)
3.8.3 Independent metric cameras

Independent metric cameras are preferable to stereometric cameras when the base
should be increased in order to achieve optimal accuracy at increased object ranges.
Figure 3.8-3 shows a universal independent metric camera. To determine elements of
exterior orientation there are levelling bubbles and horizontal and vertical angle set-
tings. By means of forced centering the camera can be replaced by another measuring
instrument such as a theodolite. These components for determing the elements of ex-
terior orientation are, however, used less and less (Section 3.8.8).
4 &
Figure 3.8-3: Schematic diagram of a terrestrial metric camera
The camera illustrated in Figure 3.8-3 has a rectangular format which can be set in
either landscape or portrait orientation. To further accommodate a particular imag-
ing situation, for example photographing tall buildings, Wild cameras incorporated an
offset principal point. Three typical configurations of the image field are shown in
Figure 3.8-4.
The technical data for the Wild P31 Universal Terrestrial Camera, a system comprising
three interchangeable precision metric cameras, are summarized in Table 3.8-2. Optical
distortion is smaller than 4 //m. Glass plates or cut-film in special holders can be used
(see also Table 3.2-1). Calibrated intermediate focusing rings control the depth of field
(Section 3.1.4 and in particular Figure 3.1-14). The optical resolving power of the P31
has already been given in Section 3.1.5, in particular with Figure 3.1-19.
The smaller version of the P31 is the Wild P32. It is designed to be mounted on a
theodolite telescope. Principal distance c = 64 mm. The P32 can also use glass plates
(64 mm χ 89 mm) or roll film. The principal point is offset by 10 mm. It has a fixed
focus distance of 25 m (see numerical example b in Section 3.1.4).
Jenoptik Jena has manufactured the universal terrestrial camera UMK with a format of
13 cm χ 18 cm. There are five different objective lenses from narrow angle to super
wide angle. The UMK's construction corresponds largely to Figure 3.8-3. Calibrated
focusing controls the depth of field (Section 3.1.4, and in particular Figure 3.1-15).
f
i»3
Portrait- (
format, main
view inward Η
I
as
1
Figure 3.8-4: The various image-field arrangements for Wild cameras with offset prin-
cipal point
Field of view Super wide angle Wide angle Normal angle

Principal distance c [mm] 45 100 202
Aperture stops 5 , 6 . . . 22 8 . . . 22 8 . . . 22
Exposure times [s] ß , 1 , . . . , 1/500
Usable format to edges of
image70 [mm] 92 χ 118 84 χ 117 ( 9 0 - 83) χ 118
Principal point
displacement [mm] 0 19 12
Standard focusing [ m] 7 25 35
Intermediate rings for
other distances no yes yes
Plates (P) or film (F) Ρ P,F P,F
Table 3.8-2: Technical data for Wild P31
70
The object extent which can be measured can be seen in Figure 3.8-9.
3.8.4 Semi-metric cameras

Semi-metric cameras occupy a position between metric and amateur cameras. Fig-
ure 3.8-5 contrasts a semi-metric camera with a metric camera. The key component of
a semi-metric camera is the reseau. This is a (stable) glass plate on which an array of
crosses at intervals of several millimetres is marked. The reseau is exposed on every
image. The reseau crosses have known coordinates for their centres which are deter-
mined prior to mounting the plate in the camera. Measured values for these points are
acquired when individual images taken with the camera are processed. A transforma-
tion71 onto reference coordinates significantly eliminates distortion and unflatness in
the film. In a sense, this transformation provides a metric image in the plane of the
stable glass plate on which the reseau is marked.
Camera body Camera body
Roseau points
Fiducial m a r k s — « ι
system
/
Film-
Shutter Shutter
Image f r a m e — ι
Reseau
Image plane
a) Metric camera b) Partial-metric camera
Figure 3.8-5: Metric camera (left) and semi-metric camera (right)72
The following comments can be made with regard to calibration and the elements of
interior orientation:
• the optical distortion of semi-metric cameras is usually significantly greater than

in metric cameras (Figure 3.1-9 shows a typical distortion curve for a semi-metric
camera). Numerical correction of optical distortion (Section 3.1.3) is therefore
absolutely necessary for high accuracy applications.
71
The bilinear transformation (Section 3.2.1.2) is a suitable transformation for each grid square. Since
the 8 unknowns of the bilinear transformation are derived without any overdetermination from the 4
reseau crosses in the corners of the square, a prior check of the data using a similarity or affine trans-
formation is recommended (see Table 3.2-3). An interesting alternative is the derivation of a closed
interpolation function for the entire image by simultaneous use of all measured reseau points. An ap-
propriate method is based on interpolation by least squares, which can also largely eliminate (filter out)
random measurement errors (Section Β 9.5.1.3, Volume 2).
72
Taken from Luhmann, T., Robson, S., Kyle, S., Harley, I.: Close Range Photogrammetry. Whit-
tles Publishing, 2006. This textbook discusses semi-metric cameras in some detail, since semi-metric
cameras play a significant role in close range measurement.
• semi-metric cameras can be focused to almost any distance. Calibration data

exists for marked focal positions. The location of the perspective centre over
time is not as stable as in a metric camera (indicated in Figure 3.8-5); the position
of the perspective centre is therefore only approximately known.
• semi-metric cameras are often calibrated using test fields. In applications de-
manding high accuracy there would also be an on-the-job calibration (Sections
3.1.2 and Ε 3.4, Volume 2).
The image format of semi-metric cameras varies from small format up to full metric
camera format. Medium format cameras (6 cm χ 6 cm) are most widely represented.
Commercially available lenses vary from super wide angle to telephoto.
Semi-metric cameras using the reseau principle are available from Geodesign, Hassel-
blad, Leica, Linhof, Pentax and Rollei. A list is available in Luhmann, T., Robson, S.,
Kyle, S., Harley, I.: Close Range Photogrammetry. Whittles Publishing, 2006 (see Ta-
ble 3.8, p. 143). Wester-Ebbinghaus (Ph.Rec. 76, pp. 603-608, 1990) has made major
contributions to the development of semi-metric cameras.
3.8.5 Amateur cameras

Amateur cameras can only be used in photogrammetry to a limited extent. The sim-
plicity of the photography is offset by the complexity of analysis and processing. This
is discussed in Volume 2 of this series of textbooks. As regards calibration, only on-
the-job calibration is a realistic option (Section 3.1.2 and Ε 3.3, Volume 2). The par-
ticular detail of defining the image coordinate system, in which the (initially unknown)
perspective centre is determined during on-the-job calibration, is discussed in the fol-
lowing section (Figure 3.8-6).
3.8.6 Terminology and classification

The classification of photogrammetric cameras into terrestrial and aerial cameras de-
rives from practical experience. From a methodological point of view, however, the
following classification is relevant:
• metric cameras, specially developed for photogrammetry.

• semi-metric cameras, not originally intended for measuring purposes, but which
have been developed for photogrammetry even though not all the elements of
interior orientation may be stable.
• non-metric cameras, which can be used for low-accuracy photogrammetric ap-
plications, particularly when supported by modern, non-specialist software pack-
ages.
Cameras can also be classified according to the internal features used to define interior
orientation:
• fiducial-mark cameras, specially developed for photogrammetry, with 4, 8 or

more fiducial marks.
• reseau cameras, which have a glass plate supporting a regular grid of crosses
placed in front of the image plane. This grid can be used to detect, in particular,
film deformation and lack of flatness.
• frame cameras, which have no fiducial marks, but sharply imaged edges. The
"indirect" fiducial marks, formed by the corners of the frame, replace the fiducial
marks and are located by intersecting lines derived from measurements of points
on the frame edges (see Figure 3.8-6).
Figure 3.8-6: Typical 35 mm film. The frame sides are measured at the points shown
and the corners determined exactly by intersection of the lines.
• false-frame cameras, e.g. Polaroid Land Camera, in which the image is bounded
not by a fixed frame in the camera, but by a paper frame which is then removed.
The interior orientation cannot be reconstructed and such photographs can hardly
be used for photogrammetry. A similar situation occurs when only an enlarged
part of an image is available.
Exercise 3.8-3. An object is to be reconstructed using two images. Consider the num-
ber of unknown elements of interior orientation when using the following cameras:
• metric camera (Solution: no unknowns.)
• calibrated and stable reseau camera (Solution: no unknowns.)
• uncalibrated but stable reseau camera (Solution: 3 unknowns and additional

polynomial coefficients for the optical distortion)
• partially stable amateur camera or frame camera with the same focus setting for
both images (Solution: 3 unknowns and additional polynomial coefficients, both
for the optical distortion and for the film deformation.)
• amateur or false-frame camera (e.g. two different image details) (Solution: 6

unknowns and additional polynomial coefficients, both for the optical distortion
and for the film deformation.)
3.8.7 CCD cameras

In close range photogrammetry film cameras have been more or less completely re-
placed by digital cameras. The lower geometric resolution of digital cameras can be re-
placed, if necessary, by choosing a shorter imaging distance to the object. Of particular
interest is the data flow in a complete photogrammetric system when data acquisition
is made not with a film camera but with a digital camera. In analytical photogramme-
try, considering the need to develop the film, there is inevitably an interruption in the
process; this is an off-line technique. In contrast, digital photogrammetry can offer an
on-line process from data acquisition through to the presentation of results. By means
of this real-time photogrammetry new application areas can be opened up (robotics,
computer-controlled medical operations, etc.).
CCD cameras used in close range application normally have a single CCD area array
sensor (Section 3.3.1) covering the entire image format. In order to increase the geo-
metric resolution there are digital cameras which sequentially sample the image in the
image plane, using the same principle already discussed when digitizing photographs
(Section 3.4, in particular Figure 3.4-3). When using a digital scanning camera there
must be no appreciable movement between camera and object during the scan. In
addition, the real-time property mentioned above is compromised.
Only a few manufacturers have been able to provide small numbers of CCD cameras
which, based on their geometric stability, can be classed as digital metric cameras,
(e.g. Figure 3.8-7 73 ). Integral arrays of 4K χ 4K detectors (and higher) are already
commercially available.
In addition to their medium format film cameras, many manufacturers offer as an al-
ternative digital semi-metric cameras. A CCD area array sensor is used in place of the
film magazine (e.g. Figure 3.8-8).
WWW addresses for digital metric cameras: www. r o l l e i . d e , www. i m e t r i c .
com, www . g e o d e t i c . com.
For moderate accuracy demands there are inexpensive standard CCD video cameras on
offer which follow television standards.
Video cameras were used in the past as electronic imaging devices in real time pho-
togrammetry. They operate according to diverse standards (for instance, CCIR/PAL
in Europe and EIA/NTSC in the US), which are quite similar with differences mainly
in the frame frequency (25 vs. 30 image frames per second) and the number of lines
per image (625 vs. 525 lines). It is important to mention that they use an analogue
technique. If digital images are required, additional frame grabbers with analogue-
to-digital converters have to be employed. Due to the analogue nature of the signal,
various perturbing influences can be observed and high quality imaging with an ac-
ceptable geometric stability for photogrammetric purposes needs very careful set-ups.
Motion pictures have never had significant importance in photogrammetry and, where
73
The photogrammetric close range systems, to which these cameras belong, reduce the images to the
coordinates of targeted points at a very early stage (Section 3.6.2).
Figure 3.8-7: INCA 3 digital metric Figure 3.8-8: AIC integral from Rollei
camera from Geodetic Systems, Inc.
(GSI)
they have been employed, video movie cameras have been "misused" as still image
cameras.
The predecessors of today's digital cameras were often called "still video cameras".
The new generation of digital consumer cameras excels the former still video technol-
ogy concerning stability, resolution, convenience and price. Cameras with a sensor
array of up to 10 Megapixels are becoming standard, some professional models have
more impressive specifications. Still, most of these cameras have all the shortcomings
of a typical amateur camera: unstable zoom lenses, careless mounting of the sensor
plate, automatic settings which can rarely be adjusted manually and often cannot be
reset to certain predefined values or cannot even be fixed. Conventional digital cam-
eras are also not suited to certain industrial applications. Some manufactures offer
special cameras for very specific applications (e.g. DALSA (www. d a l s a . com), JAI
(www. j a i . com), TVI (www. t v i v i s i o n . com)). From the photogrammetric point
of view they also belong to the category "amateur cameras".
Further reading for Section 3.8.7: Luhmann, T., Robson, S., Kyle, S., Harley, I.: Close
Range Photogrammetry. Whittles Publishing, 2006. Shortis and Beyer: In Atkinson
(ed.): Close Range Photogrammetry and Machine Vision, Whittles Publishing, UK,
pp. 106-155, 1996.
3.8.8 Planning and execution of terrestrial photogrammetry

Guide values for the maximum possible object distance or smallest possible image
scale can be derived from the required accuracy of coordinates which, on the ob-
ject, are parallel to the image plane. This is done by inverting the Formulae (3.8-2a)
and (3.8-2c), restricted to the second term (see also Equation (2.1-35)). Using For-
mula (3.8-2b), the smallest possible base can then be derived from the required accu-
racy in depth (a relevant example is solved in Exercise 2.1-12). On this basis it can be
decided if a stereometric camera with a fixed base can be used or if images must be
taken with an independent camera.
After clarifying the issue of accuracy it is then necessary to decide from where the
object can be imaged most advantageously for photogrammetric purposes, i.e. the area
within which camera stations can be located and where line-of-sight restrictions must
be avoided. In this procedure the field of view, or more accurately the focal length and
image format, can be decided. Large object distances demand long focal lengths, short
ranges not infrequently require super wide angles. The depth of field is also important
for planning the optimal object distances and it is necessary to stay within the depth-of-
field limits of sn and 57 (see Section 3.1.4). A variety of cameras and objective lenses
is valuable in making these decisions.
Finally, the object extent acquired by the images must be taken into account when
planning terrestrial photogrammetry. The width of an object acquired by a stereopair
of images can be seen in Figure 3.8-9. The relationship between base B, object range
y, object width ΟIV, principal distance c and usable image width s (from edge to edge)
is illustrated in Figure 3.8-9 and expressed in Equation (3.8-3) which allows for a 5 mm
safety margin at the image edges. (For CCD cameras 100 pixels should be chosen as a
safety margin.)
1 0 m m
*- = (3.8-3)
c y
Considerations relevant to the optimal baselength, which is a very critical parameter in

planning, can be summarized as follows:
• the required accuracy of depth measurement συ (Equation (3.8-2b)),
• the need to view into object recesses, taking into account that spatial location of
an object point requires it to be visible from at least two camera stations,
• recording the object with as few images as possible.
The greater the required accuracy of the result, then the greater the base/distance ratio
B/y must be; i.e. for a given y the base must be larger. When using independent
metric cameras the base/distance ratio B/y can be held constant by scaling the base
in proportion to the distance. The range error συ then is directly proportional to y,
according to Equation (3.8-2b). However, with fixed base stereometric cameras the
base cannot be changed; συ therefore varies in proportion to y2. If a photogrammetric
error analysis shows that a fixed base stereometric camera will not give the required
accuracy in a particular case, an independent camera with freely selectable baselength
must be used instead.
Working with independent metric cameras is similar to aerial photogrammetry. Cam-

era stations are not exactly determined and the camera axes are not exactly oriented
Figure 3.8-9: Planning the horizontal "normal case"
but aligned approximately according to the "normal case". The elements of exterior
orientation are therefore derived from control points and tie points74.
This has a number of advantages:
• photography is accelerated.
• systematic errors in interior orientation, as well as exterior orientation during

photography and processing, are largely detected by the control points.
• photography can be made from unstable platforms such as ladders or lifting plat-
forms, for zenith views the camera can be laid on the ground pointing upwards,
or it can simply be used hand-held. The most important issue is often to ensure
good and accurate ray intersections at the object.
Figure 3.8-10 is intended to encourage some solutions.

Where an object to be recorded is not roughly flat, such as a building facade, but has
some significant extent in all directions, such as a complete building, then a configura-
tion of multiple images surrounding the object is required. Four images are sufficient
if they are directed at the object corners; in this configuration every object point would
appear in at least two images. In contrast, with four images each looking perpendicular
to the object sides a three-dimensional measurement of the object is not possible.
74
When determining control points inside buildings, in narrow streets, on balconies which overhang
one another etc., a software package is required which does not separate out plan and height determination
but works directly in 3D space (Waldhäusl, P.: Vermessungswesen und Raumordnung, pp. 128-139,
1979).
1) horizontal and/or tilted images from the ground, (approximately) horizontal base
2) approximately orthogonal images with an approximately horizontal base, taken
from a lifting platform
3) stereoimagery with an approximately vertical base, orthogonal to the object
fa9ade
Figure 3.8-10: Various camera configurations
If an all-round configuration of eight images is created, as shown in Figure 3.8-11, the

following advantages are obtained:
• the diagonal images contribute to a very stable network.
• every object point is recorded in at least three images; this produces results of
high accuracy and reliability.
• the perpendicular images are little distorted; they are very suitable for producing
orthophotos.
The imaging configuration shown in Figure 3.8-11 has one disadvantage: there are no
image pairs which correspond approximately to the "normal case". The photogrammet-
ric operators, with their stereoscopic vision, would use an approximately "normal case"
(Section 6.1.2) for stereoscopic processing, for example of a facade with many three-
dimensional details. With due regard for the various processing options, the imaging
configuration of Figure 3.8-11 should be modified or extended such that, in place of
just one perpendicular image there are two, with an appropriate base. Simple tools exist
to aid the approximate alignment of the camera axes perpendicular to the object plane
(e.g. a regular grid in the viewfinder, aiming at the camera's reflection in a window or
polished surface of the fagade).
Figure 3.8-11: Recording an object with all-round imaging 75
This section on planning, which has a continuation in Section 5.7, will be concluded
with comments on the application area known as Mobile Mapping. This type of data
acquisition makes use of a road vehicle equipped with the following sensors (Fig-
ure 3.8-12):
• GPS receiver which determines position with reference to a specific GPS refer-
ence station
• inertial measurement unit (IMU) which determines the angular attitude of the
camera, as in aerial survey use (Section 3.7.3.2), and which also supports GPS
positioning
• odometer which determines distance travelled by counting the rotations of the

wheel
• barometer which provides a coarse measurement of height
• digital stereometric camera or single digital camera
• laser scanner which provides polar (spherical) coordinates to points on the road
surface (Chapter 8)
75
Other object forms and imaging configurations have been analysed with respect to accuracy and
reliability and are discussed in Section Β 4.5, Volume 2.
GPS •·'•
· •·*
reference station
Figure 3.8-12: Mobile Mapping System76
Where a single metric camera is used, the base is defined between successive images
and is roughly oriented in the direction of travel when the road is straight. Analysis
of the images is very difficult because corresponding object elements are imaged in
successive images at very different scales.
For sampled object points, such mobile mapping systems achieve an accuracy of around
±30 cm at a speed of 60km/h 7 7 .
Further reading on mobile mapping: Novak, K.: ZPF 59, pp. 112-120, 1991. Benning,
W„ Aussems, T.: ZfV 123, pp. 202-209, 1998. Gajdamowicz, K., Öhman, D„ Rise,
K.: Journal of the Swedish Society for Photogr. and RS, Nr 2002(1), pp. 103-112,
2002. He, G., Novak, K., Feng, W.: IAPRS 30, Commission II, pp. 480-486, 1992,
and Commission V, pp. 139-145, 1992.
Exercise 3.8-4. An object with an extent of 2 m χ 3 m is to be photographed by a

Rollei Scanning Camera RSC and the format is to be filled with the object's image.
The number of pixels in a full image is 4200 χ 6250 and in the scanning array's partial
image 512 χ 512. What is the resolution of the full image? (Solution: width of line
pair = 0.67 mm.) What would be the resolution of the partial image only? (Solution:
maximum width of line pair = 8.2 mm.) Values should be calculated with respect to
the object surface.
76
Provided by the ikv working group at the Institut für Geodäsie der Universität der Bundeswehr,
Munich.
77
Schwarz, K.P., El-Sheimy, N.: IAPRS XXXI(B3), pp. 774-785, Vienna, 1996.
Exercise 3.8-5. Repeat the exercise above for a CCIR video camera. (Solution: maxi-
mum width of line pair = 6.7 mm.)
Exercise 3.8-6. Repeat the exercise above for a Wild P32 (Section 3.8.3) with a total
resolving power (film and optics) of 75 Lp/mm. (Solution: maximum width of line
pair = 0.45 mm.)
Exercise 3.8-7. An 18 m wide and 12 m high, almost flat fagade is to be photographed

at height of 8 m from the full width balcony of the building opposite. Stereophotogra-
phy is required which will deliver an accuracy of σχ = σζ = ±0.5 cm in the plane of
the faijade and a y — ± 1 cm normal to the facade. The building opposite is 20 m away.
Accuracy within the image can be assumed to be ±10μπι. Is the Wild P32 sufficient
or must the P31 be used (for camera data see Section 3.8.3)? (Solution: The P32 is
sufficient.) What baselength should be used? (Solution: 6.25 m.) What is the image
scale? (Solution: 1 : 312.) How many photographs are required? (Solution: 2.) What
is the base/distance ratio? (Solution: 1 : 3.2.) How many times must an image be
enlarged to obtain an orthophoto at a scale of 1 : 50? (Solution: 6.3 times.)
Chapter 4
Orientation procedures and some

methods of stereoprocessing
This chapter deals with the reconstruction of shape and position of an object point by
point from two images. In much of Chapter 4, the alignment of the two photographs
approximates to the "normal case". In the case of aerial survey one speaks of near
vertical photographs. The base is generally chosen so that the two pictures overlap each
other by about 60% (Figure 4.0-1). Such pictures can be observed stereoscopically, or
"in three dimensions" (Section 6.1.2), when one sees a stereoscopic model, sometimes
also called an optical model, of the photographed object. In several passages of this
chapter we also deal with the general case where the photographs depart from the
"normal case".
Photograph 1 Photograph 2
Figure 4.0-1: The "model area" (overlap) in two nearly vertical photographs
Stereoprocessing in the narrow sense, that is, the creation of three-dimensional digital
models of the photographed objects, is preceded by a quite extensive preparation phase,
which is generally called orientation. This preparatory phase is crucially dependent on
whether or not the elements of exterior orientation of the two photographs are known.
In principle, photogrammetric processing leads to the computational re-establishment
of the geometric relationship of the cameras to the global coordinate system at the time
of exposure and, then, to the digital reconstruction of the object from corresponding
image points or image elements.
The orientation of the photographs is considerably simplified if the interior orientation

of the pictures is known—as, in general, is assumed in this chapter. It is also presumed
that we have photographs taken with a metric camera on film or with CCD sensors
(Section 3.7.2.2) or by means of three-line digital cameras (Section 3.7.2.3).
Section 4.1 With known exterior orientation 181
4.1 With known exterior orientation

Here we distinguish between the cases of two metric photographs (film or CCD sen-
sor, Section 4.1.1) and that of metric images taken with a three-line sensor camera
(Section 4.1.2).
4.1.1 Two overlapping metric photographs

The elements of exterior orientation are available when the pictures were taken:
• with terrestrial stereometric cameras (though only with respect to a local object
coordinate system, Section 3.8.2), or
• with a terrestrial metric camera when the position and attitude of the camera have
been accurately determined (Section 3.8.8), or
• with aerial survey cameras of which the position and attitude have been accu-
rately established in post-processing of the corresponding GPS/IMU recordings
from the photo flight (Section 3.7.3.2).
When using GPS and IMU, one speaks of direct georeferencing. GPS and IMU allow
direct transformation into a terrestrial coordinate system, normally the national coor-
dinate system. One speaks of indirect georeferencing when the elements of exterior
orientation are determined in a roundabout way using control points.
The photogrammetric processing of a stereopair with known exterior orientation starts
with the measured image coordinates ξι, η\ and ξ2, η2 of the corresponding points
Pi and P2 (Figure 4.0-1). The following relationships (4.1-1), derived from Equa-
tions (2.1-20), define intersection of rays in three dimensions; they may be used to find
the object coordinates Χ, Υ, Ζ of the point Ρ :
X = Xoi + { Z - Z 0 i)k x i
Image 1
Y = Ybi + ( Z - Z 0 l ) k y l
(4.1-1)
X = X02 + (Z - Z(yi)kX2
Image 2
Y = Y02 + (Z - Z02)ky2
The quantities k are derived from the elements of the interior and exterior orientation
together with the four measured image coordinates. Since the coordinates X(H, YQ% and
Zoi of the perspective centres are known, there are four linear equations in the three
unknown object coordinates Χ , Y and Z.
Provided the Z-coordinate axis is roughly parallel to the direction of photography (as in
the case of aerial images), then from the first and the third of the four Equations (4.1-1)
we have:
X02 — Zo2kx2 + Zo\kx\ — X01
182 Chapter 4 Orientation procedures and some methods of stereoprocessing
The X coordinate may be found from the first or the third equation, while Y can be
found from both the second and the fourth equations (providing a check). The two
slightly different Y values are then averaged.
For the "normal case" of photogrammetry (which, as we know, is arrived at directly
when using stereometric cameras) the relationships (4.1-1) become especially simple,
as one can see with the help of Equations (2.1-32) or (3.8-1).
The equation systems (4.1-1) and (4.1-2) can also be used during the processing of
images which depart greatly from the "normal case". The solution for intersection out-
lined above is, however, not rigorous in the case of such a general camera arrangement.
The redundancy cannot really be used by an averaging of the different Y values, but
must be dealt with using least squares estimation from the original measurements, that
is the image coordinates ξι, ή\ and £2, % (Appendix 4.1-1). In this estimation prob-
lem we have four (non-linear) observation equations in three unknowns. Using the
collinearity equations (2.1-19) we derive four linearized observation equations:
=
äx) dx +
{sr <"-+(§§) dZ - ((,-(«)
"Vi (νι-νϊ)
0 (4.1-3)
V
Ü2 = dY
>
"Vi dY-
The partial derivatives ()° are evaluated from the relationships in Appendix 2.1-3 us-
ing approximate values for the unknown coordinates. ξ® 2 and rfj 2 are computed image
coordinates derived from Equations (2.1-19) using the approximate values for the un-
known object coordinates and the known elements of interior and exterior orientation.
Approximate values for the unknown object coordinates can be determined by means
of relationships (4.1-1) and (4.1-2).
Numerical Example. In the exact "normal case" the approximate solution and the
rigorous solution result in identical answers (why?); nevertheless we wish to compute
such a case by means of least squares estimation because the computational effort of
the least squares solution will can be kept small for the "normal case". Β = 1.20 m,
c = 64.20 mm.
Image coordinates:
ζ [mm] η [mm]
Image 1 3.624 34.202
Image 2 -14.697 34.196
Approximate values from (2.1-32): -Z° = 4.205 m, Y° = 2.240 m, X° = 0.237 m.

The design matrix A (Appendix 4.1-1) is developed using the partial derivatives which
are found in Exercise 2.1-18 (Appendix 2.1-3). The image coordinates η® 2 (4.1-3)
Section 4.1 With known exterior orientation 183
to be calculated for the vector 1 of observations (Appendix 4.1-1) come from the (sim-
ple) central perspective equations (2.1-28) to (2.1-31) of the exact "normal case":
( 15.268 0 0.861 \ ( 0.00559 \

0 15.268 8.133 0.00271
A = 1 imml
15.268 0 -3.496 0.00564
γ 0 15.268 8.133) \ -0.00329 )
After setting up and solution of the normal equations and calculating the accuracy one
finally obtains (Appendix 2.1-3):
0.23700 \ 0.00037 0.23737 m '

2.24000 + -0.00001 2.23999 m
-4.20500 -0.00001 -4.20501 m
±0.23 mm
±0.75 mm
±1.36 mm
The solution of a three-dimensional intersection using least squares has the great ad-
vantage that, as a by-product of the adjustment computation, the accuracy of the new
points is also found. The accuracies calculated in the above numerical example are too
optimistic because the elements of exterior orientation were taken to be error-free. In
Volume 2 (Sections Β 3.5.7 and Β 3.5.8), extended adjustment formulae are dealt with
for the case of imperfect elements of interior and exterior orientation.
Publication No. 43 of the EuroSDR (European Spatial Data Research, former
OEEPE—Organisation Europeenne d'Etudes Photogrammetriques Experimentales)
concerned an international test of direct georeferencing. The results achieved were
roughly in accord with those given in Section 3.7.3.2. See also Cramer, M.: GIS
6/2002, pp. 3Ί-Α2.
In principle, with known exterior orientation one requires no control points. In many
cases, nevertheless, control points are incorporated in the processing; on the one hand,
this provides checks and, on the other hand, improves the known elements of exterior
orientation.
4.1.2 Metric images with a three-line sensor camera

When taking images with a three-line sensor camera, a GPS/IMU recording must be
made simultaneously, from which at every point in time the elements of exterior ori-
entation are available with a relatively higher accuracy (Sections 3.7.3.2 and 3.7.3.3).
When one identifies a point P, either manually or automatically, then the three time-
stamps t\,t2 and ti are known from the indices of the three rows; these are the times at
which the point Ρ was imaged in the three lines one after the other (Figure 4.1-1). Cor-
responding to these three time-stamps, the elements of exterior orientation available
from the GPS/IMU records are taken:
*o(«2), Yo(t2), Z0(t2),üj(t2), <p(t2), n{t2) (4.1-4)

Χ*(ί 3 ), Yo(f3), ω(ί 3 ), ν»(ί3), κ(ί 3 )
As a result, the prerequisites exist for a three-dimensional intersection, using three di-
rections, to determine the X Y Z coordinates of the point P. To do so, the equation
system (4.1-1) or (4.1-3) is extended to a third picture; that is, there are six observa-
tion equations for the three unknowns. In calculating the quantities k, the following
calibrated values for the ξ coordinates are used:
• for kx\, kyi the distance a,

• for kx2, ky2 the distance 0 and, (4.1-5)
• for kx3, kys the distance - b.
The η coordinates of the three image points Pi, P2 and P3 differ only slightly from
each other.
02(t2)
Figure 4.1-1: Object point Ρ in the three images of a three-line sensor camera
In processing the three scan lines, if the elements of exterior orientation obtained from
the GPS and IMU are available only with limited accuracy and are to be improved,
control points are introduced in the procedure. This matter is discussed in Section 5.5.
4.2 With unknown exterior orientation

In this case, using control points, we must find the following twelve elements of orien-
tation for two images, whether on film or on CCD chip:
Picture 1: Χ 0 ι, *οι, Z0\, ω{, ψ\, κ ι

Picture 2: Χ02, Υ02, Ζο2, ω2, φ2, κ2
The different methods of solution may be classified in three groups:

Section 4.2 With unknown exterior orientation 185
• separate orientation of the two images (see Section 4.2.1)
• combined, single-stage orientation of the two images (see Section 4.2.2)
• two-step combined orientation of a pair of photographs (see Section 4.2.3)
4.2.1 Separate orientation of the two images

Here it is assumed that for each image at least three control points are known (Fig-
ure 4.2-1). For each image at least six equations (2.1-19) can be written in the six
unknowns, which are underlined below:
& = /(£o, ε,Χο,Υ^), Ζο,ω, κ, X i , Y i , Z i )

t = 1,2,3 (4.2-1)
Vi = /{νο,ο,Χβ,Υβ,^,ω,φ,κ,Χί,Υί,Ζι)
After linearizing them using approximate values the six equations can be solved (iter-
atively, Newton's method) for the six unknowns. The procedure is known resection in
three dimensions.
Overdetermined resection in three dimensions is solved using the method of least

squares estimation (Appendix 4.1-1). Each point Pi provides two linearized obser-
vation equations for the observations & and ήί (instead of writing out the partial deriv-
atives ( w h i c h are calculated using the approximate values for the unknowns, we
use the quantities a and b introduced in Appendix 2.1-3):
vî = a2<iXo + a^dYo + a^dZ^ + α^άω + α^άψ + αηάκ — — ξ®)
ν η ί = b2dXo + b^dYo + ^ciZo + b$duj + b(,άφ + b j d n — (ήί — η
ξ® and η® are the image coordinates computed from Equations (2.1-19) using the known
elements of interior orientation, the known control point coordinates Xj,Yj,Zj and the
approximate values for the unknown elements of exterior orientation. For the first
iteration, in the case of near-vertical photographs taken in a West-East flying direction,
one can use the following approximations: = φ° = κ° = 0. Values introduced
from a GPS-supported navigation system in the aircraft normally provide sufficiently
accurate first approximations for yo°, ZQ (Section 3.7.3.1).
Numerical Example. We are given image coordinates and ground coordinates of the
following points:
ξ [mm] 77 [mm] Y[m] Z[m]

1 - 8 6 . 1 5 - 6 8 . 9 9 36589.41 25273.32 2195.17
2 - 5 3 . 4 0 82.21 37631.08 31324.51 728.69
3 - 1 4 . 7 8 - 7 6 . 6 3 39100.97 24934.98 2386.50
4 10.46 64.43 40426.54 30319.81 757.31
The elements of interior orientation are: c = 153.24 mm, ξο — Vo = 0.

Since we are concerned here with a near-vertical photograph, we can use simple pro-
portion as for a truly vertical photograph to find approximate starting values for the
camera position. From two control points Pi and P j one gets:
7° —-
Z { c { X j - X i ) + ( - j Z j - ( i i Z i ) H t i j - &
0
— X i - t ^ Z f t - Z f b / c
yO _
Y i - m ( Z l - Z \ ) l c
The two control points Pi and P j should lie at approximately equal ground heights
Zi and furthermore their ξ-coordinates should not be too close together. One would
therefore use either points Pi and Pj or points P2 and P4.
Solution:
X0 = 39795.45 m / 0.99771 0.06753 0.00399 \

Y0 = 27476.46 m R= -0.06753 0.99772 -0.00211
Z0 = 7572.69m \ - 0 . 0 0 4 1 2 0.00184 0 . 9 9 9 9 9 /
Calculation and presentation of the accuracies are omitted.

The computation of resection is frequently required in photogrammetry. For this reason

there are many solutions; efforts have been concentrated on linear equations which
work without initial approximations. A successful linear solution exists for four control
points as fully described in Volume 2 (Section Β 4.1.2). It is not, however, a least
squares adjustment so the accuracy is limited. If the highest accuracy is demanded one
will regard the results of this method as approximations and then complete a precise
computation with the observation equations (4.2-1).
On many flying missions nowadays the position of the camera station is already very
accurately determined by GPS (Section 3.7.3.2). The attitude of the photograph, how-
ever, is not always recorded with IMU. In such a case the missing orientation angles
can be found using the equation system (4.2-2), without the terms for dXo, dYo and
dZo.
An alternative for finding the elements of orientation is the Direct Linear Transfor-
mation (DLT). The DLT establishes the relationship between the two-dimensional im-
age coordinates and the three-dimensional object coordinates by means of a projective
transformation, with which we are already acquainted for the case of plane objects
(Sections 2.1.4 and 2.2.3b) and which for three-dimensional objects may be extended
in the following form (an elegant derivation using homogeneous coordinates is to be
found in Appendix 4.1-1):
aiX + a2Y + ατ,Ζ + a,4

~~ c\X + c2Y + c3Z + 1
(4.2-3)
_ biX + b2Y + b3Z + b4
V
~~ ciX + c2Y + c3Z + 1
These transformation equations are linear with respect to the transformation parameters
ο», bi? Cj (Section 2.1.4 contains a numerical example for a plane object; the computa-
tional steps outlined there are no different from those for a three-dimensional object).
Each control points leads to two equations. Therefore, for the determination of the
eleven transformation parameters a*, bt, a a minimum of six control points is needed
(more accurately, for the sixth control point only one of the two image coordinates
needs to be measured). The fact that there are eleven transformation parameters is sur-
prising. In Section 2.1.3 we established that a central projection of three-dimensional
space is defined by nine independent parameters (three for the interior orientation and
six for the exterior orientation). The additional two parameters in the DLT can be
interpreted as an extension of the interior orientation to allow a scale difference be-
tween the two image coordinate axes and non-orthogonality of these axes. Projective
photogrammetry, therefore, usually discards a Cartesian coordinate system.
The parameters of the DLT cannot be interpreted physically. While the eleven para-
meters aj, bi, Ci can indeed be successfully expressed in terms of the parameters of
interior and exterior orientation, accounting for a known interior orientation within a
computation is difficult. Hence, the DLT is most suitable for the processing of non-
metric images, such as amateur photographs, video images and so on (Section 3.8.6).
Nonetheless, DLT is also used for the evaluation of metric images. The results of the
DLT, which are easy to obtain on account of the linearity of the DLT equations, are used
as approximations for a subsequent adjustment based on central perspective relations.
Valuable information concerning the interior orientation can be extracted from this rig-
orous adjustment, its accuracy considerably enhanced. In Volume 2, Section Β 4.7.1,
the DLT is dealt with in detail.
Separate orientation of the two images, for example by means of resection, making use
of the relationships (4.2-2), has disadvantages which can be formulated as follows with
reference to Figure 4.2-1:
• no use is made of the information that the homologous rays intersect each other
(as, for example, in the new object point Pa,).
• one needs at least three full control points (with known Χ , Υ, Ζ) in the stere-
opair as opposed to the procedures described below; for these later methods the
minimum requirement is just two plane control points (with known X, Y) and
three height control points (with known Z).
4.2.2 Combined, single-stage orientation of the two images

The image coordinates of the control points and some new points are measured (Fig-
ure 4.2-1). For each control point there are four equations (2.1-19) in the twelve un-
knowns:
ζα — / ( £ 0 ; ci2Loi >^01 <4^1> Ψγΐ !

Image 1
Vn = Χ^^Υ^γ,^,ω^,ψ^, Κχ,Χί,Υί, Zi)
(4.2-4)
C :
ζί2 = , Ä 0 2 > ^ 0 2 ' 2 i
Image 2
c X Y Z
Vi2 = f(m, ,2Lo2,Yü2'^2'til2,ip2,K2i i, i, i)
For each new point there are in fact three further unknowns (with double underlining
below), but also four additional equations as follows (2.1-19):
ζιΐ = / ( £ 0 , C, X o ! , Yj0l, Zq! , ωχ, ψχ, , Χ±, Υ±, Ζ±)

Image 1
Vn = f(vo, c, Xjqi , Zoi, Zjoi , ω λ , ψ_χ, κχ, X±, Ζ±)
(4.2-5)
ζί2 = /(&, C,Xo2,£>2>^> to,
Image 2
First, it is necessary to linearize the Equations (4.2-4) and (4.2-5) using suitable first
approximations (Appendix 2.1-3); these systems of equations are usually highly over-
determined. Subsequently the linearized equations are to be solved by means of least
squares estimation (Appendix 4.1-1). The results of the computation are the twelve
elements of exterior orientation and the object coordinates Χ , Υ, Ζ of the new points.
For this combined, two-image resection1 in the following example, a tally of the obser-
vations and unknowns involved may be given:
Given: 2 metric images with
3 control points and
3 new points
Observations: 12 coordinates, 1st image ·
nd • ^
12 coordinates, 2 image
=> 24 observations in all Δ ·
Unknowns: 6 rotations of the two images
6 coordinates Xo, Yo, Zq of the two perspective centres
9 coordinates Χ , Υ, Ζ of the three new points
=>· 21 unknowns in all
Redundancy: 24 - 21 = 3
Exercise 4.2-1. Revise the above summations assuming that both images have the
same interior orientation, but that it is unknown. Give your critical opinion of this
exercise especially from the viewpoint of comprehensive checking of the computation.
(Answer: In this case, there are 24 observations and an equal number of unknowns.
That is to say, the system of equations is soluble but gross errors cannot be discovered.
Note that for this special, degenerate, case of least squares one still needs the partial
derivatives αϊ and b\ of Appendix 2.1-3.)
Since arbitrarily many control points and new points may be introduced in this method
and since the adjustment is based on the indirect observation equations (4.2-4) and
(4.2-5) (which equations establish the direct connection between the unknowns and the
original observations, the image coordinates), this method of orientation is the most
accurate. It is a one-step solution, in contrast to the following procedure which splits
the orientation into two steps and ignores the correlations between the results of the
first step in solving the second. A disadvantage of the one step solution is that addi-
tional operations are unavoidable in order to find approximate values, as described in
Section 4.2.1.
4.2.3 Two-step combined orientation of a pair of images

The orientation procedure shown in Figure 4.2-2 works in two steps. In the first step
a stereomodel in an arbitrary three-dimensional xyz coordinate system is created from
the two photographs. In the second step this model is transformed into the X Y Z co-
ordinate system. The procedure is most easily understood if we begin with the second
step:
'There is a single, special word in German for this procedure: "Doppelbildeinschaltung", literally
"double image insertion". As is frequently the case, there is no word in English for this German word.
One might call it "two image resection" but, more frequently, one would simply say that the problem was
solved using the "bundle method" (Section 5.3).
V>
o-v
Figure 4.2-2: Two-step orientation procedure

Second step: The relation between the model coordinates x, y, ζ and the object coor-
dinates Χ, Υ, Ζ can be expressed by the equations (see also Figure 4.2-2):
in which:
Xu, Yu, Zu ... are the object coordinates of the origin of the xyz system
m . . . is the scale number of the xyz system
R . . . is the matrix of the three-dimensional rotation of the xyz system into
the XYZ system defined in terms of the three rotations Ω, Φ, Κ
See Equations (2.1-11) and (2.1 -13) or Appendix 2.1-1.
The seven parameters, Xu, Yu, Zu, Ω, Φ, K, m, are called the elements of absolute
orientation. Equation (4.2-6) represents a three-dimensional similarity transformation
(2.1-18).
At least seven equations are required for the computation of the seven elements. Equa-
tion (4.2-6) gives us:
• three equations for a full control point (Χ, Υ, Ζ all known)
• two equations for a horizontal, or plan, control point (Χ, Y both known)
• one equation for a height control point (Z known).
The absolute orientation requires at least two plan control points and three (non-col-
linear, in plan) height control points; alternatively, it requires two full control points
and a height control point not collinear in plan with the full control points. Further
details of absolute orientation, especially in the over-determined case, are dealt with in
Section 4.4.
Exercise 4.2-2. On the basis of the matrix (2.1-13) write out the individual equations
for a full control point, for a plan control point and for a height control point.
First step: When the photographs were taken, each individual point on the object
could be thought of as having given rise to two rays, one towards each camera, and
hence to two image points. If the positions and orientations of the images with respect
to each other are correctly restored, all such pairs of homologous rays will once again
intersect, in points which define the surface of the model in the xyz system.
Since seven of the twelve unknown elements of exterior orientation can be determined
by the absolute orientation, five unknowns remain for the first step. Therefore it is
reasonable to assume that a necessary condition for relative orientation is that the ho-
mologous rays from at least five well distributed points intersect and that if this is
achieved, all other pairs of homologous rays will intersect, thus forming a complete
photogrammetric model2, see Section 4.3.
The procedure required to achieve the above state is called relative orientation, since
only the relative positions and orientations of the two bundles of rays are determined.
There is no reference to the X Y Z coordinate system and no control points are required
for relative orientation.
The condition for correct relative orientation, the intersection of each of five pairs of
homologous rays in a model point, may be formulated using the scalar triple product
of the three vectors b, p ) t , ρ2% (Figure 4.2-3 and Equation (4.2-7a)). The scalar triple
product, which is composed of the scalar product of one of three vectors with the vector
product of the other two vectors, gives the volume of the parallelepiped of which three
concurrent edges have the lengths and directions of the three vectors concerned. If the
three vectors are coplanar their scalar triple product is zero, because the vector product
of two of them is then perpendicular to the third and thus their scalar product vanishes.
Equation (4.2-7b), which is frequently called the coplanarity condition and which is
the condition that homologous rays intersect at each of five points, expresses the basic
condition for relative orientation.
Figure 4.2-3: Condition for intersection, or coplanarity
bT(pHxp2i)=0 i — 1,... ,5 (4.2-7a)

The scalar triple product can also be written using the following determinant:
bx Pli,x P2i,x
D = by P\i,y P2i,y — 0 i= 1,. (4.2-7b)
bz P\i,z P2i,z
2
This two stage process brings to mind the definition of photogrammetry given at the very start of
Section 1.1. In relative orientation a model is created which reproduces the form of the object; absolute
orientation determines the scale, position and orientation of the object with respect to the ground coordi-
nate system. Instead of absolute orientation one could use the term georeferencing.
Section 4.3 Relative orientation 193
Sections 4.3 and 4.4 are devoted, respectively, to relative orientation and to absolute
orientation.
4.3 Relative orientation

4.3.1 Relative orientation of near-vertical photographs
The coplanarity condition (4.2-7) is considerably simplified if it can be assumed that
the photographs are almost vertical. It is first necessary to set out the relationship
between the image coordinates and the model coordinates under the assumption of
near vertical photographs. In this case in the relationships (2.1-20), instead of the
object coordinates Χ, Υ, Z, the model coordinates x, y, ζ are introduced and, instead
of the orientation elements, increments to the orientation elements. The quantities used
can be seen in Figure 4.3-1:
o,
ω — άω,
φ = dtp,
κ = du,
ξο — Vo — χ ο\ = 2/oi = 0, (4.3-1)
p.(*,y,z) x&2 = bx,
2/02 = dby
Z02 = zoi + dbz
h = zqi — z.
Figure 4.3-1: Orientation elements

in the case of near-vertical images
For small rotations the matrix R (2.1-13) simplifies to:
(4.3-2)
For photograph 1, therefore, from Equation (2.1-20):
ξι — η\άκ\ — οάψ\
χ —
= (-h)
ξ\άψι + ηιάω\ - c
Division of both numerator and denominator by — c gives:
— — + —άκι + άψ\
* = — - —
11 Η, ^Α
άψ\ αωι
c c
Expanding the expression on the right-hand side in terms of the series 1/(1 + x) =
1 - χ + x 2 — . . . and ignoring second and higher order products of small quantities
gives:
, , χ 6 - ηιάκι - cdipi
X = (—h) — ;
-ξιάφι + η\άω\ - c
= i-h) (—— + %άψ\ - ^^-(koi + —άκι + άψχ

\ c cz cz c
An expression for the y coordinate may be found in a similar way:
y= , , χ ξιάκι— +ηι
; — Λ +cduι
-ξ\αψ\ + η\άω\ - c
Rearrangement then= (—h)

gives f——
\us the — ηί·
^άωχ relationship
η following + nL — η— άκ\ the
between — model
άω\ coordinates
x, y, h (h instead of z), the image coordinates of a point P, the orientation elements
and the principal distance for picture 1 :
x\ = h ( ^ — ( 1 + ^'Γ ) άφι + ^γ-άωχ — ^άκλ

(4.3-3)
C \ c2 J & ' c
cz to O2, a\ similar
Taking into account a shift cof origin c z process
/ gives
c corresponding rela-
tionships for picture 2:
X2 = bx + (h + dbz) ( -- -- ( (l 1 + Az , άψ
. 2. + —dw2 -
\ rc
V \V rc / r-i
c? rc
( 4 3 4 )
- - "
y2 = dby + {h + dbz) ( ^ - ^ψάφ2 + ( 1 + \ ) άω2 +

VC C V C J c
Setting y\ equal to j/2 gives the condition for intersection of two rays, (4.2-7), in the
case of near-vertical photographs:
0 = dby + Idbz + + ψώρ, - (1 + 1 ) dun
_ _ ϊψ ά ψ 2 + Λ + ή ) ^ + kdK2
In Section 2.1.5 we introduced the definition of y-parallax, ρη — η ι — %· We may

therefore rewrite the condition for intersection as:
Pr, =ydby + ψ-dbz + ^^-άφι - (c + — ) άω\ - ξ\άκ\ - άψ2

h h c \ c j c
(4.3-5)
/ 2\
+ ( c+T^-)dw2 + ξ2άκ2
Equation (4.3-5) expresses the η parallax, in the case of near-vertical photographs, as

a function of eight orientation elements. Consideration of the effects of small physical
translations and rotations of the photographs shows that not all eight elements can,
however, be computed from measured η parallaxes. Equal but opposite small rotations,
dn ι and du2, of the left and right photographs about their (near-vertical) axes will have
the same relative effect on the η parallax as a small change, dby, in the y position of the
right-hand photograph.
Clearly, not all three of dn\, dni and dby may be computed simultaneously from mea-
sured //-parallaxes; it is possible to compute the following pairs of elements from mea-
sured ^-parallaxes: dn\, άκ2\ άκ\, dby\ dK2, dby. Likewise, either but not both of άψ\
and db 2 can be found since they have the same relative effect; and either but not both
of άω\ and (Ιω2 can be found. Three of the elements of orientation included in Equa-
tion (4.3-5) must be excluded from the computation; if, for instance, both άω\ and άω2
were to have been included, the matrix of observation equations, corresponding to that
found in the following numerical example, would have been singular or almost singu-
lar. The implication in Section 4.2.3 that only five elements of orientation were to be
found in relative orientation is confirmed.
Of the possible choices of five orientation elements for the computation of numerical
relative orientation of two near-vertical photographs, the following two variants are
usual:
Relative orientation using rotations only. The photographs are rotated only, their
positions meanwhile remaining unchanged; this is known as independent relative ori-
entation and in the US as the "swing-swing method".
ρη = —ξ\άκ\ + $,2dK2 + ——dipi — + ^c + — ^ di02 (4.3-6)
For approximately vertical photographs the image coordinate η\ is approximately equal

to 772; the coefficients of duj\ and άω2 are almost identical, apart from their signs. One
can choose to determine either άω\ or du;2, but not both; άω2 has been chosen in Equa-
tion (4.3-6).
Relative orientation using only elements of the second photograph. One photo-
graph remains fixed while the other is translated and rotated; this is frequently known
as dependent relative orientation and often as the "one-projector method".
ρη — ydbv + ^dbz - + ( c + — ) duj2 + £ 2 ^ 2 (4.3-7)

h h c \ c J
It can be seen that, of the six elements of orientation of the right-hand photograph,
only five occur in Equation (4.3-5). For near vertical photographs, within the limits
of approximation applied in this section, translation b^ of the right-hand photograph
does not influence the 77-parallaxes. This may be seen as further confirmation that the
measurement of five 77-parallaxes, or the intersection of five pairs of homologous rays,

suffices for relative orientation.
If more than five η-parallaxes have been measured, a least squares solution may be
used (Appendix 4.1-1), in which case Equation (4.3-6) or (4.3-7), augmented by the
residual v Pn is regarded as observation equation with the substitution of the residual
v Pv as the constant term.
Numerical Example (of relative orientation of independent images). We are given the
image coordinates (in [mm]) of eight points and the principal distance c = 152.67 mm.
Point 6 »7i 6 m Ρη=ι71 - m

1 93.176 5.890 6.072 5.176 0.714
2 -27.403 6.672 - 1 1 2 . 8 4 2 1.121 5.551
3 83.951 107.422 -4.872 105.029 2.393
4 -11.659 101.544 - 9 9 . 2 9 8 95.206 6.338
5 110.326 -97.800 34.333 -99.522 1.722
6 -12.653 -87.645 -96.127 -93.761 6.166
7 37.872 40.969 - 4 8 . 3 0 6 37.862 3.107
8 41.503 -37.085 -42.191 -40.138 3.053
We wish to compute the elements of relative orientation and their accuracies. From
Equation (4.3-6) the observation equations are of the form:
"pv = ξινι
+ iidki + ^^-άφι - + ( c + ^ ) di02 - ρη
c c
The observation equations in matrix form are (Appendix 4.1-1):
/ -93 6 4 0 153 ^ ( 0.714 \

27 - 1 1 3 - 1 1 153 5.551
-84 -5 59 3 225
( dk \ x
2.393
12 - 9 9 - 8 62 212
dkj 6.338
-110 34 - 7 1 22 218
dtp 1 1.722
13 - 9 6 7 -59 210
άφ 2 6.116
-38 -48 10 12 162
V düJ2 J 3.107
-42 -42 -10 -11 163 j \ 3.053 /
which we may write as:

ν = Ax — 1
The normal equations have the form (Appendix 4.1-1):
A T A x = Α τ 1; χ = (ATA)"1AT1
The solution of the normal equations gives:
άκι = 1.73 gon = 1°33' dn2 = - 0 . 8 2 gon = -44'

άψι = - 0 . 3 4 gon = - 1 8 ' άψ2 = 0.05 gon = 3'
du2 = 1.40 gon = 1°16'
The weight coefficient matrix is (Appendix 4.1-1): Q = ( A T A ) 1
/ 70 63 05 00 29 Ν
63 62 06 00 28
05 06 13 03 02
00 00 03 13 00
^29 28 02 00 13 j
Standard deviation of an observation ρη:
η . . . number of observation equations ( = 8)

u ... number of unknowns ( = 5)
Standard deviation of an element of orientation: = σρη ^Jqkk
a Ki = 0.009VaÜÖÖ7Ö = ±15 mgon (49")

σΚ2 = 0.009^0.00062 = ±14 mgon (45")
σφι = σψ2 = σωι = 0.009νΌ.00013 = ± 7 mgon (23")
While the procedure put forward in this Section 4.3.1 is exceedingly instructive, it is
inadequate, especially for error analysis, when the photographs have large tilts. For
photographs with arbitrarily large tilts, relative orientation and error analysis are dealt
with in Sections 4.3.2 and 4.3.6.
Exercise 4.3-1. The rotation matrix (4.3-2) is orthogonal only to a first order of accu-
racy. If the matrix is to meet the conditions for orthogonality with errors no larger than
10~6, how large can the angles ω, φ and κ be? (Answer: 0.6 gon (32')).
Exercise 4.3-2. There is an obvious and very strong correlation between some of the
unknowns of relative orientation as can be seen from the matrix of weight coefficients,
Q, given above. Among other things this leads to the fact that accuracies of functions
of the orientation elements should be found by applying the general rules of error
propagation. How large is the standard error of the difference Ακ — κ\ — «2? (Answer:
±4.4 mgon = ±14")
4.3.2 Relative orientation and model formation using highly tilted

photographs
We restrict ourselves initially to relative orientation using rotations only. In deriving
linearized equations for the case of highly tilted photographs we cannot, as we did in
the preceding section (see Equations (4.3-1) and (4.3-2)), adopt the unit matrix for a
first approximation to the attitude of the pictures; we must use general rotation matri-
ces. Figure 4.3-2 represents the relative orientation of tilted pictures. The three vectors
b, pi and P2 must be coplanar for a minimum of five points. We wish first to consider
relative orientation using rotations only and we set ω\ = 0. The Ci axis is thus at right
angles to the y axis of the model coordinate system (Figure 4.3-2).
Figure 4.3-2: Relative orientation of tilted photographs using rotations only
With respect to their separate image coordinate systems the two vectors pi and p2 have
the following components:
Pi P2 = I V2 (4.3-8)
We transform these vectors into a common coordinate system, the xyz model coordi-
nate system, using the two rotation matrices R i and R2.
pi = Ri(wi = 0, φι,κι) P2 = R-2(W 2 ,^2,«2) (4.3-9)
Each of the two rotation matrices R i and R2 can be written as the product of one
matrix R ° rigorously computed but based on approximate rotation angles and a matrix
riR representing incremental rotations (corresponding to that of Equation (4.3-2)):
6
PI - d R i R ? [ 771 P2 = riR 2 R^ I 772 (4.3-10)
The third vector shown in Equation (4.3-10) is the base vector b; it lies in the direc-
tion of the χ axis of the model with an arbitrarily chosen unit length. From Equa-
tion (4.2-7b) coplanarity of the three vectors p i , p 2 and b, corresponding to point PL,
is expressed by Equation (4.3-11) in which, as in Figure 4.3-2, the index i is omitted
1 Pi,χ P2,x
D = 0 Pl,y P2,y = P\,yP2,z ~P\,zP2,v — 0 (4.3-11)
0 Pi,ζ P2,z
From the coplanarity equation (4.3-11) we may derive the following equation in the
increments to the five unknowns for relative orientation using rotations only:
/ 3 D λ0 / d D \ ° ( d D \ ° J
The partial derivatives ()° are evaluated as in Appendix 4.3-1 using the approximate
values for the unknown angles. The determinant D° is evaluated as follows:
a) working out the rotation matrices R , and using Equation (2.1-13) and the
approximate values for the unknown rotation angles ψ\, κ°ν ψ\ and κ\
b) computing the components of the vectors p^ and p® from Equation (4.3-9) us-
ing the measured image coordinates ξι,ηι and ξ2,η2 of homologous points, the
principal distance c and the rotation matrices R® and R!
c) from Equation (4.3-11) evaluating the determinant D° — yp\ z — p2 yP°\ z
Following the first iteration which results in increments, άφ®, άκ®, άω®, άφ\ and άκ^, re-
spectively, to the approximate values, the steps a), b) and c) above are repeated and the
partial derivatives ()° of Equation (4.3-12) are computed anew as in Appendix 4.3-1.
After stopping the iterations we have the elements of relative orientation, so that for-
mation of the model can proceed, using intersection (Section 4.1) with the values:
Zoi = 2/01 = Zo\ = ω \ = 2/02 = Z02 = 0, x02 = 1.
Numerical Example (Continuation of the Numerical Example of Section 4.3.1). The
results obtained in that example were ω\ — 0, ψ\ — - 0 . 3 4 gon = - 1 8 ' , κ,\ —
1.73 gon = 1°33', ω2 = 1.40 gon = 1°16', ψ2 = 0.05 gon = 3', κ2 = - 0 . 8 2 gon =
- 4 4 ' . These are used as approximate values for the second iteration, applying Equa-
tion (4.3-12). After some iterations we arrive at the following results which do indeed
depart considerably from the simplified solution of Section 4.3.1:
ωι = 0 ω2 = 1.387 gon = P14'54"
ψ\ = - 0 . 4 5 5 gon = - 2 4 ' 3 4 " φ 2 = - 0 . 0 9 6 gon = -5Ί1"
κι = 1.708gon = 1°32Ί4" κ 2 = - 0 . 8 3 8 g o n = - 4 5 Ί 5 "
Model formation is performed using these final values.
Exercise 4.3-3. Compute the xyz model coordinates of points 1 to 8 of the numerical
example in Section 4.3.1 and of both perspective centres, using the elements of relative
orientation from the above numerical example and adopting the value for the base
bx = 100 mm. Solution (from Equations (4.1-3)):
Point χ [mm] y [mm] ζ [mm]

1 107.236 9.563 -173.269
2 -30.721 6.888 -177.348
3 96.141 128.340 -178.034
4 -15.472 117.838 -177.894
5 140.093 -116.509 -186.622
6 -10.627 -101.529 -176.316
7 44.222 49.029 -178.058
8 50.827 -41.946 -177.727
Oi 0.000 0.000 0.000
O2 100.000 0.000 0.000
For the dependent relative orientation (using elements of the right-hand photograph
only) the three vectors required for the coplanarity equation are as follows (Ri = I):
Pi = I P2 = R-2(W2,<P2,K2) b=i (4.3-13)
The coplanarity equations (4.2-7b) are set up using these three vectors; subsequently
the observation equations (corresponding to Equation (4.3-12)) and the partial deriva-
tives (as in Appendix 4.3-1) have to be evaluated, and so on.
This section should be concluded with a note of criticism. In the strict sense of ad-
justment by the method of least squares, the corrections (often called residuals) should
apply to the original observations, in this case to the image coordinates ξι, η\ and ξι, r/2,
and should not, as in the Equation system (4.3-12), relate to the "pseudo-observation"
the function 3 D. This imprecision can be remedied in one of two ways: one is by
changing to the Gauss-Helmert model of adjustment by least squares and the other is
by adapting the combined, single-stage orientation of Section 4.2.2 to relative orienta-
tion.
4.3.2.1 Gauss-Helmert model of relative orientation
The vectors pi and are functions of the observations ξι, η\ and £2, V2 in (4.3-9).
In the coplanarity equation the residuals υ and the unknown elements of orientation
appear. A strict relative orientation leads to the Gauss-Helmert solution, also known
as the general case of least squares estimation.
( f + ( f O ° * v + ( f ) \ 2 + ( ι + ( « ) V . ( 4 3 1 4 )
3
"Residuals" when applied to a quantity other than the original measurements are also called "alge-
braic residuals".
The partial derivatives for the unknowns and for the determinant D° have been fully
discussed in connection with Equation (4.3-12). These derivatives can be derived from
the relationships (4.3-9) and (4.3-11):
3D dD
•qT — 7*1,21 P2,ζ - 7*1,31 P2j3/ 7j— = 7*l,22P2,z ~ n,32P2,i/
„ ™ (4.3-15)
dD dD
S
9ξT2 = 7*2,31 Pi,y - 7*2,21 Pi,ζ δη2 = ^"2,32 Pi ~ 7*2,22 Pi,2
The indices in r 1,32, for example, refer to the element in the third row and the sec-
ond column of rotation matrix R4. Details of the solution of the estimation prob-
lem (4.3-14) are to be found in the adjustment literature4. The partial derivatives
of (4.3-15) have to be re-calculated from iteration to iteration, especially when one
has begun from very poor approximations for the rotation matrices and Ri>.
4.3.2.2 A combined, single-stage relative orientation
The single-stage solution for the 12 elements of external orientation of two pictures was
described in Section 4.2.2; for relative orientation this takes on a special form. Since
there are no control points but just new points, Equations (4.2-5), only, are applicable,
i.e. two homologous points leading to four observation equations. The unknowns are
the model coordinates, Xi, yt, ziy for each pair of homologous points together with the
five unknowns of dependent relative orientation, άφ\, dn\, du>2, άψ2 and άκ2- Of the
original 12 unknowns in Equations (4.2-5), the following are known: xoi = yoi =
zoi = 2/02 = -202 = ωι = 0, £02 = 1 (compare Figures 4.2-1 and 4.3-1).
Literature relevant to Sections 4.3.1 and 4.3.2: Rinner, K.: Phia, pp. 41-54, 1942.
Schut, G.: Phia XIV, pp. 16-32, 1957/58. Thompson, E.H.: Phia 23, pp. 67-75, 1968.
Blais, J.A.R.: Can. Surv. 26, pp. 71-76, 1972. Stefanovic, P.: ITC-J (1973), pp. 417-
448. Molnar, L.: Geow. Mitt, der TU Wien, vol. 14, 1978. Mikhail, Ε., Bethel, J.,
McClone, C.: Introduction to Modern Photogrammetry. John Wiley, 2001.
4.3.3 Alternative formulation of relative orientation

Since the orientation of two overlapping pictures is a central problem in photogramme-
try and in computer vision, it is hardly surprising that many different formulations have
been developed. In Section 4.3.2 we learnt of a method typically used in photogram-
metry, a technology which normally works with metric photographs. The formulation
presented in what follows is more frequently associated with computer vision, in which
the use of images without known interior orientation predominates. This alternative
formulation also offers very interesting insights into the relative orientation of metric
images.
4
See, for example, Mikhail, E., Observations And Least Squares. IEP-A Dun-Donnelley, New York,
1962.
We start with the basic equation (4.2-7a) of relative orientation with three coplanar
vectors b, pi and p 2 . The value of a scalar triple product is independent of the ordering
of its terms. We choose the following form for the coplanarity condition:
p|(bxp2)=0 (4.3-16)
It can be shown that the cross product b χ p2 can be written as in Equation (4.3-17) in
which the elements of the skew-symmetric matrix Β come from the vector b:
b χ pa = B p 2 = (4.3-17)
Exercise 4.3-4. Expand the cross product b χ and show that the result is the same
as that found from the matrix multiplication of (4.3-17).
All skew-symmetric matrices are singular; in particular detB = 0, as is easily shown:

d e t B = 0 (b2x) + bz(-bxby) + by(bxbz) = 0.
Here we limit ourselves to the case of independent relative orientation (Figure 4.3-2).
Taking the base vector as b T = (1,0,0) the coplanarity equation (4.3-16), as expressed
in the form of (4.3-17) becomes:
p T b P 2 = pT I Ρ2 — 0 (4.3-18)
The two vectors pi and p2 (see Equations (4.3-9)) are derived from the image coordi-
nates ξι, ηι and ξ 2 , η2, the principal distance c and the rotation matrices Ri and R2,
which describe the attitudes of the two pictures with respect to the model coordinate
system. The matrices C\ and C2, describing the interior orientation of the two images,
may be introduced into Equations (4.3-9), generalized by the inclusion of the coordi-
nates of the principal point, ξο and ηο, and usually identical for metric photographs:
ξι ~ξο,\ \ / 1 0 —£0,1
PI = R] I 771 - 770,1 = R] 0 1 -770,1
0 - ci / V 0 0 -C!
= R i Q (6,771,1)T
(4.3-19)
/ 6 - £0,2 \ / 1 0 £0,2
P2 = R2 \ m - m,2 J = R2 0 1 -770,2
\ 0-C2 / \ 0 0 -c2
T
= R2C 2 (6,r?2,l)
With this representation, the coplanarity equation (4.3-18) reads as follows:
(£i,7 ? 1 ,1)CTRTBR 2 C 2 I 7)2 1 - 0 (4.3-20)

The matrix F is known as the fundamental matrix of relative orientation or the cor-
relation matrix. Provided that we are dealing with metric images, the fundamental
matrix F can be computed using the parameters of relative orientation, that is to say,
the two rotation matrices R i and R2 (Section 4.3.2), the known matrix Β (4.3-18) and
the known matrix C of the interior orientation. Although the relationship (4.3-20) de-
scribes a relationship between the image coordinates ξι, η\ and ξι, Vi of corresponding
points in two images, it is not possible, given image coordinates in one image, to find
the corresponding image coordinates in the other image. This becomes clear from the
following expansion of the coplanarity equation (4.3-20):
£ \ & f u + i \ m f n + i \ f n + & m h \ + m m h 2 + m h ? > + & h \ + m h 2 + h'} = 0 (4.3-21)
With known matrix elements fik and, for example, both image coordinates ξι and 771,
one has a single linear equation in the two unknowns £2 and ηι· Adopting arbitrary
coordinates 772 one can, therefore, compute matching coordinates £2·
As a consequence, we know that corresponding to a point in one picture there is a

straight line in the other picture on which the homologous point must lie. (This result
is pursued further in Section 6.8.5.5, as illustrated in Figure 6.8-9.).
An interesting application of the coplanarity equation (4.3-21) lies in the relative orien-
tation of images for which the interior orientation is unknown. Since Equations (4.3-19)
contain the coordinates of the principal point (ξο,ηο) it is possible for the image coor-
dinates £1,771 and £2^2 of homologous points to be referred to an arbitrary coordinate
system. Division by one of the 9 elements fik of the matrix of Equation (4.3-21) re-
duces their number to 8. As the divisor one should select the largest possible element,
which may be a different element fik in different cases. We choose thus the nine
elements become /ifc=/jfc//33· In this way, for each pair of homologous points in two
pictures, we get the following linear equation for a least squares estimation (Appen-
dix 4.1-1):
ν = ξ \ ξ 2 f u + ξ l η 2 f n + ξlf\3+ξ2ηlf2\+mη2f22+ηlf23+ξ2f3\+η2h2 = - 1 (4.3-22)
Since Equation (4.3-22) has eight unknowns fik and since each pair of points provides
only one such equation, it would seem that this kind of relative orientation requires
eight homologous points. The solution using Equations (4.3-22) does not involve the
elements of interior orientation, even when they are known, but has more than five un-
knowns. It was shown in the postscript to Exercise 4.3-4 that the matrix Β is singular;
therefore the F matrix, which contains Β as a factor, is also singular and, thus, not
all eight of the fik are independent; in fact the number of independent unknowns is
seven. The two additional unknowns, beyond the five of conventional relative orienta-
tion, represent general (unknown) parameters of interior orientation. The question of
the conversion of the elements fik resulting from a relative orientation using the above
Equations (4.3-22) into the parameters of conventional relative orientation will not be
considered here. It should certainly be mentioned here that, if the interior orientation
is known but not is employed for relative orientation using Equations (4.3-22), it is
possible to derive the five standard elements of relative orientation from the values fik•
In conclusion, it remains for the link with projective geometry to be presented. The
coplanarity equation (4.3-20) contains eight matrix elements fik:
(fn hi fn\ ίξι\

hi hi hi \\m =0 (4.3-23)
V/31 hi 1 / V 1 /
This representation may be compared with the projective relationship between an ob-
ject plane and an image plane, especially when it is expressed using homogeneous
coordinates (see Equation (2.2-1-1) of Appendix 2.2-1). The two image planes in a
relative orientation are projectively related to each other, though not "point to point"
but "point to straight line", as already mentioned above.
How the unfamiliar elements of relative orientation f ^ may be introduced into pho-
togrammetric processing will be discovered in Section 6.8.5.5.
Exercise 4.3-5. In Section 4.3.1 the image coordinates £ι, η\ and ηι of eight cor-
responding points in two metric images are given. Determine the relative orientation
using Equations (4.3-22).
Solution:
fn fn /i3 \ / - 0 . 0 0 0 0 1 8 0.000284 -0.005009 \

hi hi h-i = -0.000302 0.000071 -0.576408
hi hi 1 / V -0.017098 0.550187 1 )
Exercise 4.3-6. This solution, however, takes no account of the fact that the determi-
nant det F must vanish. If this additional constraint is introduced in the least squares
estimation, one is dealing with an adjustment by the method of least squares with a
constraint equation 5 .
fn fn fn\ ( 0.000001 -0.000040 -0.008565 \

hi hi fi3 = 0.000031 0.000039 -0.261240
hi hi 1 J \ - 0 . 0 0 2 0 4 4 0.263777 1 J
The additional constraint equation markedly influences the result in this example.
Exercise 4.3-7. Write the F matrix for the "normal case" of photogrammetry (Sec-
tion 2.1.6) and derive the coplanarity equation from it.
Solution:
/0 0 0\
If unit baselength is chosen (||b|| = 1),B = 0 0 - 1
Voi Oj
5
The solution may be found, for example, in Mikhail, E., Observations And Least Squares. IEP-A
Dun-Donnelley, New York, 1962.
/0 00\ /ξ2\
(4.3-20): (ξι,ηι, 1) 0 0 c ) l η2 = -βη2 + οηι=ηί-η2=0
\ 0 - c 0 j \ l )
The coplanarity equation for the "normal case" of a stereopair leads to the statement
that there are no //-parallaxes, as has already been stated in Section 2.1.6. Further
note: Because fe = 0 in the exact "normal case", relative orientation by means of
Equation (4.3-22) will fail.
Literature: Brandstätter, G.: Mitt, der TU Graz, Folge 87, 2000. Haggren, Η., Niini,
I.: The Photogr. Journal of Finland 12, pp. 22-33, 1990. Hartley, R„ Zisserman,
Α.: Multiple View Geometry in Computer Vision, Cambridge University Press, 2000.
Mayer, H.: Journal of the Swedish Society for Photogrammetry and Remote Sensing
2002(1), pp. 129-141, 2002. Förstner, W.: IAPR 33(B3/1), pp. 297-304, Amsterdam,
2000, and PFG (2000), pp. 163-176. Ressl, C.: IAPR 34(3A), pp. 277-282, Graz,
2002.
4.3.4 Relative orientation of near-vertical photographs by

y-parallaxes
Following the excursion into an alternative relative orientation, in particular for non-
metric photographs, we turn again to relative orientation of metric images. In the
following we assume approximately vertical photographs, just as in Section 4.3.1 in
which the relative orientation was solved in terms of ^-parallaxes, that is differences
in image coordinates (η\ - η2). For analogue photogrammetric instruments (which
receive no attention in this new edition of the textbook) a relative orientation using
y-parallaxes was commonly used, the y-parallaxes being measured in the stereomodel.
Although orientation with y-parallaxes is no longer of any practical significance, the
formulae are well suited for assessment of accuracy and for consideration of "critical
surfaces".
When carrying out relative orientation in an analogue instrument, any existing x-

parallax should be removed before the observation of y-parallax; x-parallax is of no
significance in relative orientation (see Figures 4.3-3 and 4.3-4).
For a relative orientation using model parallaxes py rather than image parallaxes ρη, the
relationships (4.3-6) and (4.3-7) must be transformed such that model-related values x,
y and py replace the values ξ, η and ρη which refer to images. The following relations
are valid for near-vertical photographs (see Figure 4.3-1):
h h h χ h, h, ,, „ „^
py « -Ρη, -m, « -6 ~ -6 + b (4.3-24)
c c c c c
Substituting these results in Equation (4.3-6) gives:
py = -χάκι +(x- b)dn 2 + ^άψι - ^ ~ ^ άψ2 + h( 1 + ^Λ άω2 (4.3-25)

J r
\y-parallax
χ-parallax
\ /
N
,<
Figure 4.3-3: .r-parallax after the re- Figure 4.3-4: «/-parallax after removal
moval of y-parallax in a relatively ori- of the z-parallax in a model before rel-
ented model ative orientation
Exercise 4.3-8. Derive corresponding equations for dependent relative orientation.
The elements of relative orientation can be estimated by least squares using observed
y-parallaxes and the associated model coordinates x, y and h (instead of z) in Equa-
tion (4.3-25). If the ^/-parallaxes are measured in specially chosen positions, explicit
expressions can be derived for the required five unknowns. With mountainous or rela-
tively flat ground, however, different procedures are necessary.
4.3.4.1 Mountainous country (after Jerie)
In this procedure one chooses six corresponding points, at which to measure the y-
parallax, and which are constrained to lie (Figure 4.3-5)
• on two parallel lines in the model, separated in the χ direction by a distance b

and
• on three parallel lines, separated in the η direction in the image planes by dis-
tances δ (y direction in the model)
Under these conditions the following ratio remains constant for point i where
i = 3 , . . . , 6.
ψ = *= r (4.3-26)
hi c
Let
R = 1 + r2 (4.3-27)
Defining the constant R as in (4.3-27) and introducing it in (4.3-25) we get the follow-
ing observation equations for independent relative orientation:
άκι d,K 2 άψι άψ2 άω2

V\ 0 -b 0 0 hi ~Py,
v2 -b 0 0 0 h2 ~Py2
v3 0 0 br h3R -Pyi (4.3-28)
v4 -b 0 br 0 h4R ~Py4
V5 0 -b 0 —br h5R
V6 -b 0 —br 0 h6R ~Pyt
Forming the normal equations gives (Appendix 4.1-1):

άκ ι dK2 άψ\ άψ2 du> 2

2
3b —6/12 — bh4R — bh(,R KPyi +Py4 +Py6)
3 b2 —bh\ - bh?,R - bh5R KPy ι +Pyi +Py5)
lb2r2 brhAR — brheR -br{Py4 Pye)
26V 2 brh$R — brh$R -br{pm -Py5)
h\ + h\+ -hipm - h2pyi-
2 2 2 2 2
R (h + h + h + h ) h$Rpm — hARpyA—
h5RPy5 - h6RPye
The solution of the normal equations is:
άω2 — - ((-2hi + h-iR + h5R)(2pyi - pm ~py5) +

( - 2 / i 2 + hAR + h6R)(2py2 - pyi - py6)) /
2
( ( - 2 hi + h3R + h5R) + ( - 2 / i 2 + hAR + h6R)2)
d
1 R +
<P2 = ^(Pyi ~Pys) +
ά
1 R (4.3-29)
Ψ\ = 2fr(Py< ~Pye) + +
dn 2 = - ^ (pm + py3 + pys) + ^ (/ii + h^R + h5R)du> 2
du ι = - ^ (pyi + pm + py6) + ^ (/i2 + hAR + h6R)du}2
Explicit expressions for the orientation elements of dependent relative orientation can
be derived in a similar manner 6 .
Numerical Example (of Equation (4.3-29)).
Given: Principal distance c = 152.64 mm
Model base 6 — 170.00 mm
Point separation in image δ — 80.00 mm
Observed: Parallaxes in the model py and vertical distances h from the six orientation
points in the model to the perspective centres:
Ρ h [mm] py [mm]
1 245 0.2
2 232 -0.70
3 225 0.14
4 225 -0.99
5 240 0.82
6 250 0.70
6
Jerie, H.G.: Phia (1953/54), pp. 22-30.
Substitution in expressions (4.3-29) gives the following values for the elements of ori-
entation result:
dm = 0.82 gon = 44' dn2 = 0.55 gon = 30'
άψχ = -0.53 gon = - 2 9 ' άφ2 = -0.20 gon = - 1 1 '
άω 2 0.43 gon = 23'
4.3.4.2 Flat ground (after Hallert)7
For approximately flat ground one can assume that h = const. The six orientation
points (also known as Gruber points) then lie on a rectangular grid both in the model
and in the images (Figure 4.3-6 shows the image space).
In a similar manner to that for Equations (4.3-28) and (4.3-29) one can form the obser-
vation equations corresponding to the six orientation points, the normal equations and
their solutions in explicit form:
άκι =
(pm (6c2+4<j2)+Pm (6c2+852)
~(Pv3 +Pm)(3<? + 2<52) - ( Ρ * + + P * ) ( 3 c 2 - 2δ2))
dKi =
" JL· (pri] ( 6 2° 2 ( 26 c 2 + 4 s 2 } 2
~(Pm + Pvs)(3c - 2<5 ) - (pVi + +p7?6)(3c + 2δ2)) ( 4 .3. 3 o)
άψι = Ρ η
~ ^
d(
P2 = 2^2 (Pm ~Ρη5)
=
^ 4J2 (~2Pm - 2
Pm + Pn3 + Pv* + Pvs +
7
Hallert, B.: Über die Herstellung photogrammetrischer Pläne. Diss. Stockholm 1944, 118 p.
Exercise 4.3-9. Derive Equations (4.3-30) by establishing the observation equations,

the normal equations and their solution.
Exercise 4.3-10. In (4.3-29) substitute a constant value for h(= c), the value b in place
of δ (Figure 4.3-6) and ρη in place of py. Then check the agreement between (4.3-30)
and (4.3-29).
4.3.5 Critical surfaces in relative orientation

It is appropriate to ask whether relative orientation procedures lead to a unique and sta-
ble result for every imaginable form of object surface and location of exposure stations.
We can most simply answer this question by considering the relationships (4.3-29).
A critical situation can arise with respect to the elements of orientation κι,φχ and
ψ2 if the base b is zero or very small, since the denominator in the relevant expressions
then vanishes or almost vanishes. If the base is very small or, to put it another way, the
forward overlap is very large, the two profiles on which the orientation points lie move
very close together. It is not surprising then that no solution is possible using Jerie's
method. Although such stereomodels are to be avoided on account of their very poor
accuracy (in the first of Equations (2.1-34) the base Β appears in the denominator) the
relative orientation can be solved by applying, for example, the Formula (4.3-6) using
corresponding points lying outside the profiles beneath the perspective centres.
If there is a very small value of r or of δ, the elements ψ\ and ψ2 are indeterminate

when using the relationships (4.3-29). In this case all the orientation points lie more or
less beneath the base (Figure 4.3-5). Such a situation arises in practice if in either the
upper or lower margins of the pictures no corresponding points can be found (see also
Section 4.3.6). This may happen because of the presence of water surfaces, forested
areas, clouds, and so on.
Regarding the orientation element ω 2 it is necessary to consider whether the denomi-
nator in the first Equation of the system (4.3-29) can become, or approach, zero. The
form of the denominator leads one to suspect that a critical situation may occur if the
profiles 3, 1, 5 and 4, 2, 6 (Figure 4.3-5) have the same shape, especially if they are
symmetrical about points 1 and 2. Thus we arrive at one of the four identical conditions
for a possible critical surface:
h2 = h\R (4.3-31)
2
Equation (4.3-27): h2 = h4( 1 + r )
Equation (4.3-26): h2 = h4(l + ^ f ) (4.3-32)

'4
Re-arranging: y\ = h4{h2 - h4) (4.3-33)
Equation (4.3-33) shows that the triangle whose vertices are the points 2, 4 and O2 has
a right-angle at 4 (Figure 4.3-7); regarding 021 as the base of the triangle, j/4 is the
height while Λ,4 and (h2 - h4) are corresponding segments of the base. Thus 02 lies
on a circle passing through points 2 and 4 (and also 6). Since profiles 3, 1, 5 and 4,
2, 6 have the same shape, the ground surface for which relative orientation cannot be
performed is a right circular cylinder containing both camera stations, Ο ι and
If, instead of Equation (4.3-31), we examine the relationships 2h\ = h^R + hsR and
2h 2 = h4R + h^R, we can derive more general conditions for critical surfaces in
relative orientation. For example, a circular cylinder whose axis does not lie in the
vertical plane containing the base (the line joining the two perspective centres), but
which is parallel to that plane, is also a critical surface. Furthermore, the two circular
profiles 3, 1, 5 and 4, 2, 6 can have different radii; that is, a conical surface containing
Ο ι and O2 is a critical surface. Nor is it necessary for the base, 0 \ 0 2 , to be a generator
of the cone; it is necessary only that the points Οχ and 02 should lie on a conical surface
which also contains the relative orientation points on the object surface. Further details
may be found in the technical literature.8
8
Hofmann, W.: DGK, Reihe C, Heft 3, 1953. Rinner, K.: JEK Band IIIa/1. Brandstätter, G.: IAPR
Exercise 4.3-11. Show that the denominator in the άω2 of Equation (4.3-29) is zero
for the case shown below in which the axis of the circular cylinder is not in the vertical
plane containing the base, and thus that the ground surface is a critical surface relative
to O.
Hint: Use the cosine rule to express the length s in the two triangles and hence show
that 2/i2 = (1 + ( S / c ) 2 ) ( h 4 + h6).
In what way do the critical surfaces manifest themselves in relative orientation? Dif-
ficulties arise in the solution of the normal equation system because the equations are
singular or ill-conditioned. Even when the normal equations can be solved, this leads
to very large standard deviations in the ω-ι unknown (see also Section 4.3.6).
Some of the following possibilities may improve the situation:
a) flying with a different type of camera. Figure 4.3-8 illustrates cross-sections of

ground forming critical surfaces for each of three camera types (normal angle,
wide angle and super-wide angle); the axis of the critical cylinder is taken in the
direction of flight. Only when the landscape is very mountainous can it provide a
critical surface for super-wide angle cameras. On the other hand, using a normal
angle camera the relative orientation can relatively frequently be insoluble or,
when the ground surface corresponds approximately to the form of the critical
cylinder, simply very unreliable.
31(B3), Vienna, 1996. Brandstätter concerned himself with critical surfaces in relation to alternative
methods of relative orientation. The relationship to two-dimensional resection in ground surveying
should also be mentioned. As is well known, the position of a new point, A, from which three con-
trol points are observed with a theodolite, cannot be determined if the control points and A lie on a
circle. The two circles in Figure 4.3-7 and the circle in the figure of Exercise 4.3-11 correspond exactly
to this situation.
NA
Figure 4.3-8: Critical cylinders in relation to camera type
b) fly across, rather than along, valleys.
c) in the relative orientation include only points which do not lie on or near the
critical surface.
d) after absolute orientation of a model which is free of parallax, but which is de-
formed as a result of errors in relative orientation, corrections to the elements of
relative orientation may be derived from residual height errors at check points.
(See Section 4.3.6.2 or special literature.)9
e) computation of a combined, single-stage orientation of the two images (Sec-

tion 4.2.2) or computation of a bundle triangulation (Section 5.3) including all
the photographs in the block.
4.3.6 Error theory of relative orientation

Standard deviations of the elements of relative orientation are dealt with in the first sec-
tion while in the following section (Section 4.3.6.2) deformations of photogrammetric
models arising from certain errors of relative orientation are discussed.
4.3.6.1 Standard deviations of the elements of orientation10
At the end of all orientation procedures using least squares an estimate of the standard
deviation of the computed elements is obtained (Appendix 4.1-1). An error compu-
tation of this kind is given in the numerical example of Section 4.3.1. In order for
such accuracy measures to be evaluated, it is necessary to establish accuracy values
9
Regensburger, K.: Photogrammetrie. VEB Verlag für Bauwesen, Berlin, 1990.
l0
Literature: Gotthardt, Ε.: BuL 15, pp. 2-34, 1940.
for the "standard case" of relative orientation, which we define as involving approxi-
mately flat, level ground and six orientation points as in Figure 4.3-6. Application of
the rules of error propagation to Equations (4.3-30) results in the following expressions
for standard deviations of the orientation elements in this standard case:
σ
ψι — σφι
V3 c
σ (4.3-34)
σ
ω2 2 δ2 Ρη
J2 3 c4
++ Tö2 + Τ TJCT.
δΜΐ δ 4 δ4 Pn
It is assumed in the above formulations that the ^-parallaxes are measured with equal
accuracy σΡν at all six orientation points and that no correlation exists among the 77-
parallaxes.
For individual types of cameras and using an assumed value of σΡτ], for example
σρ = ± 5 /im, one can apply Equation (4.3-34) to compute accuracy estimates, as
in the following Table 4.3-1. For example, for a normal angle camera:
σω = 0 5 ρ = ±10m
Ϊ 9 ^ 2 ° · ° g°n(32")
The other values can be calculated in a similar manner:
Camera type σω σφ
Format and p.d. [cm] [mgon (")] [mgon (")] [mgon (")]
Normal angle
±10(32) ± 8 (26) ±34(110)
(23 χ 23, c = 30)
Wide angle
±5(16) ±4(13) ±10(32)
(23 χ 23, c = 15)
Super-wide angle
±3(10) ±2(6) ±5(16)
(23 χ 23, c = 8.5)
Table 4.3-1: Accuracy of elements of relative orientation for different camera types
This table has great significance in practical photogrammetry. In any particular case
the operator, or the software, assesses the extent to which the accuracy from this table
is acceptable or not acceptable. We now carry out such an assessment for the specific
results given in the numerical example of Section 4.3.1, in which case a wide angle
camera was used. The calculated standard deviations (σ Μ = ±15mgon (49"), σΚ2 —
±14 mgon (45"), σψι = σψ2 = σωι = ± 7 mgon (23")) slightly exceed the target values
from Table 4.3-1. This excess stems from the accuracy of the parallax measurements:
a standard deviation of ± 9 μη\ was estimated, while Table 4.3-1 is based on ±5 μτη.
The excess can, however, just be tolerated. The values may clearly exceed the targets
(a factor of more than two should not be accepted), for the following reasons; in each
case a remedy is given:
• insufficient accuracy in measurement of image coordinates or η-parallaxes.

Remedy—repeat the measurements.
• a critical surface for the relative orientation. Remedy—the measures quoted in

Section 4.3.5.
• poor distribution of homologous points (for example, arising from inability to

match points on water surfaces). Remedy—computation of a bundle triangula-
tion (Section 5.3) including all photographs of the block.
The orientation elements constitute an important interim result on the way to the final
photogrammetric result, the reconstituted object points. Monitoring these interim re-
sults is of great importance in quality control. The accuracy of the object points, which
will be considered in Section 4.6, is of still greater interest for the user.
4.3.6.2 Deformation of the photogrammetric model
A relative orientation can be performed with only a specific, limited accuracy and the
resulting elements of orientation will therefore contain specific errors. The question
arises as to how such errors in orientation result in deformations of the stereomodel.
We limit the following discussion to distortion of the ζ coordinate. In general, model
deformations in height are more critical than those in plan (x, y).
Determination of heights in the stereomodel depends on a;-parallaxes. It is neces-
sary, therefore, to derive relationships between x-parallaxes, px = x\ — x2 (x\ and x2
are model coordinates related to the first and second pictures respectively) and small
changes to the orientation elements.
From the first of each of the sets of Equations (4.3-3) and (4.3-4):
Px = x2~ x\ =
h
c
-ςι ~ \h-\ y- άψ\ Η τ—αω\
Η ακ\
, -b , h
ξ
η ι η*- / η x
η2
C \ & ) Cr C
With reference to the relationships (4.3-24) the image-related quantities can be replaced
and instead of the expression (/i£i/c — h^2/c — bx) the small quantity (—dbx) can be
introduced:
„ (x — b) ( x2\ xy
px = -dbx —dbz -hll + j^j άψ\ + ~ ydnι
(4.3-35)
(x — b)\ {x — b)y
+h ( 1 -I 2— J άφ2 άω2 + ydn2
It can be seen from Figure 4.3-9 that:
dz = J p x (4.3-36)
b
O, b o2
Figure 4.3-9: Relation between the x-parallax and the height error dz in the stereo-
model
Substitution in Equation (4.3-35) finally gives us the desired relationship between er-
rors in the orientation elements and the resulting height error in the model, dz.
(x — 6) ί h2 χ2 \ xy yh
dz — -dbx -dbz — I ——I—— J άφι + —άω\ —dn\
b \ b b J b b
(4.3-37)
(x-b)2 (x — b)y yh
+ + άφ2 άω2 + —du 2
b b
The effects of the most important terms in Equation (4.3-37) are illustrated in Fig-
ure 4.3-10 11 ; in that figure instead of dbz (see Figure 4.3-1) we have used dbz2 and we
have added the element dbz 1, equivalent to a negative db 2 . We can see from the various
diagrams of Figure 4.3-10 that absolute orientation will remove the effects of (small)
errors in the relative orientation, except for that arising from error in duj and part of that
caused by άφ.
It should be clearly mentioned that these deformations are superimposed on the surface
of the model; it is as if the height reference plane, the xy plane, of a model free of errors
in relative orientation is distorted into an inclined plane, a cylinder, a paraboloid and
so on.
Exercise 4.3-12. In the mathematical literature one usually finds x2/a2 - y2/b2 = 2ζ
as the equation to a hyperbolic paraboloid. Under what conditions does this reduce to
the equation given above, that is, ζ = cxy in which c is a constant. (Answer: putting
a = b and rotation of the axes by π/4.)
"Albertz and Kreiling: Photogrammetric Guide. 4th ed., 1989.

Tilted plane
Parabolic cylinder
-f-d«.
Hyperbolic paraboloid
db
Tilted plane
Figure 4.3-10: Model deformations in height
Numerical Example.
Given: Principal distance c = 85 mm, format 23 cm χ 23 cm
Photo scale number rrib — 10000
Forward overlap 60%
=> base (actual) = 920 m (Equation (3.7-1))
Side overlap 25%
=> max. y coordinate (actual) = 862 m (Equation (3.7-1))
a) (Ιωι deformation
Let the error du^ be 30mgon (or 1'37"), which is ten times the standard devi-
ation given in Table 4.3-1. We wish to find the maximum height error arising
from this error; the maximum height error will appear along the line χ = 0;
Equation (4.3-37) gives:
= = 8 6 2 ^ = 0.41m
Using four height control points in the corners of the model, this will be reduced
during absolute orientation to about ± 2 0 cm.
b) άψ2 deformation
Let the error άψ2 be 20mgon (or 1'05"), which is again ten times the standard
deviation given in Table 4.3-1. We wish to find the maximum height error arising
from this φι error. From Equation (4.3-37):
(h2 {x-b)2\J
Figure 4.3-11: Bowing of the model as a result of άψ2

As well as the bowing of the model, there is a constant vertical displacement, dz:
h? J 850 2 20
dzk = —άψ2 = -^r.r— = 0.25 m
b 920 ρ
This is in addition to a term varying as the square of the χ coordinate; this will
be a maximum when χ = 0:
(
y r
— h)2 20
dzQ = ' άψ2 = 920— = 0.29 m
b ρ
Under absolute orientation, the uniform displacement in a vertical direction will

be removed completely and the effect of the quadratic term will be reduced to
approximately 1/4 of the maximum value (see Figures 4.3-10 and 4.3-11). The
maximum remaining error from the bowing of the model will amount to dzq/4 =
29/4 cm = 7 cm.
Although we have considered very large errors in orientation, du2 and άψ2, the result-
ing height errors in the model are of the same order of magnitude as the standard de-
viation of height measurement, σζ. According to Equation (4.6-1) σζ is about ± 1 0 cm
for well defined points using this camera, flying height and overlap. In practice, there-
fore, the model deformations are of little significance by comparison. Large orientation
errors and therefore large model deformations can arise only with an unfavourable dis-
tribution of homologous points (choice limited, for example, by water surfaces) or, in
some cases, in the presence of a critical surface (Sections 4.3.6.1 and 4.3.5).
Model deformations are, however, caused not only by orientation errors but also by
sources of systematic error such as lens distortion and film shrinkage. The effects
of these systematic errors on the geometrical correctness of the stereomodel are very
Section 4.4 Absolute orientation 219
complicated. They can manifest themselves as y-parallaxes larger than the measuring
accuracy after relative orientation as well as in large residual errors at the control points
after absolute orientation. A part of the y-parallaxes caused by these systematic errors
is removed during relative orientation, although these parallaxes do not belong to the
relative orientation. As a consequence, the orientation elements are wrong and this,
therefore, results in model deformations. Such systematic errors produce correlations,
especially among the xyz coordinates of neighbouring object points. To a great extent
the inclusion of additional parameters in the mathematical model overcomes this prob-
lem (see Section Β 3.5.6 and Β 5.2.4 in Volume 2 as well as the relevant literature 12 ).
Exercise 4.3-13. Consider the processing of photographs from a stereometric camera

(c = 60 mm, portrait format 80 mm χ 95 mm, Section 3.8.2) in which the right-hand
camera makes an angle of 99.9 gon (89°55') with the base. The object distance is 10 m.
Find what error arises in the direction of the camera axes (depth error); use formulae
for the "normal case". (Answer: The error varies between 13 cm and 18 cm depending
on the χ coordinate; see Figure 3.8-1.)
Exercise 4.3-14. Repeat the above exercise for the case where the zenith distance of
the right-hand camera axis is 100.2 gon (90° 11'); that is, the axis of the right hand
camera is tilted downwards by 0.2 gon (11')· (Answer: The error varies between 0 cm
and 14 cm depending on the χ and ζ coordinates.)
4.4 Absolute orientation

A procedure for the orientation of a stereopair was described in broad outline in Sec-
tion 4.2.3; absolute orientation is the second step. In absolute orientation the stereo-
model, which is defined with respect to an arbitrary coordinate system {xyz), is brought
into a superior, or global, object coordinate system (XYZ) (Figure 4.2-2). The math-
ematical relationship is given by a three-dimensional similarity transformation. The
seven parameters of this transformation are found using control points, points for which
both the model coordinates and the object coordinates are known.
4.4.1 Least squares estimation

For a least squares computation by means of indirect observations (Appendix 4.2-1)
Equation (4.2-6) has to be linearized. For the time being we assume that initial values
are known for the seven parameters of absolute orientation with the model scale number
m « 1. Using these values the original model coordinates χ are transformed into the
object coordinate system. These transformed coordinates are denoted by X°.
This means that:
X « χ = X° => Ω = άΩ, Φ = άΦ, Κ = dK, m = 1 + dm, X u = dX„ (4.4-1)

12
Schwidefsky/Ackermann: Photogrammetrie. B.G.Teubner, Stuttgart, 1976.
As a result the linearized rotation matrix R (2.1-13) is given by:
/ 1 -dK άΦ\
dR= dK 1 -άΩ (4.4-2)
^ -άΦ dü 1)
Expanding the product m R , where m = 1 +dm and ignoring second order terms gives:
/ 1 + dm -dK άΦ \
m R = (1 + dm)dR = dK 1 + dm -άΩ
\ -άΦ dQ 1 + dm J
(4.4-3)
( dm -dK άΦ\
dK dm —άΩ.
\ -άΦ άΩ dm)
in which I represents the unit matrix. Thus the linearized form of Equation (4.2-6) for
a three-dimensional similarity transformation becomes:
X = dXu + (1 + dm)ä R X ° (4.4-4)
Writing this in ordinary algebraic notation:
X = dXu +X°dm +Z°dΦ -Y°dK

Y = dYu +Y°dm -Ζ°άΩ +X°dK +y° (4.4-5)
Ζ = dZu +Z°dm +Υ°άΩ -Χ°άΦ +Z°
For a least squares estimation by the method of indirect observations these simultane-
ous linear equations in the seven unknowns are rearranged as follows:
vx = dXu +X°dm +Z°dΦ -Y°dK -{X - X°)

vy = dYu +Y°dm -Ζ°άΩ +X°dK -(Y - Y°) (4.4-6)
vz = dZu +Z°dm +Υ°άΩ -Χ°άΦ - ( Ζ - Z°)
Equations (4.4-6) may be written in matrix form (Appendix 4.1-1) as:
ν = Ax - 1 (4.4-7)
A full control point results in all three such equations, a plan control point in two and a
height control point only in the last of Equations (4.4-6). The computation proceeds in
the normal manner (Appendix 4.2-1), with the formation of normal equations and their
solution:
A T A x = A T l; x = ( A T A ) " ' A T l (4.4-8)
Since Equations (4.4-6) are only linear approximations, further iterations are normally
required, unless the initial approximations are very good.
Numerical Example. We are given the model coordinates of five points:
Μ 23 24 50 51 45
X 0.303532 0.192638 0.303848 0.204120 0.246931
y
0.595068 0.602834 0.403493 0.434574 0.594227
ζ 0.034298 0.034116 0.026903 0.036672 0.034676
The first three points are full control points; point 51 is a height control point. The
corresponding object points have known coordinates in the control system, frequently
the national coordinate system, as follows13:
[m] 23 24 50 51
X 3321.65 3402.84 1776.75
Y 1167.56 2061.10 1196.79
Ζ 579.48 579.80 493.19 574.62
In addition, the following approximate values are known, for example from the flight
plan (Section 3.7.1) or from one of the methods described in Section 4.4.3:
x i 0 ) = -1400 m, Yu } = 3600 m, z f 1 = 300 m

m (0) = 8000, Ω(°) = φ(°) = 0, Κ® = 300 gon (270°)
Using these initial values the original model coordinates χ are transformed into ap-
proximate ground coordinates X°, by means of the following equation:
/ -1400\ / 0 1 0\ (x\
Y° = 3600 + 8000 - 1 0 0 y
Z ° J \ 300 / \ 001 j \ z J
In what follows we denote these coordinates by X ( 0 ) :
[m] 23 24 50 51 45
3360.54 3422.67 1827.94 2076.59 3353.82
χ ( ο )
y(o) 1171.74 2058.90 1169.22 1967.04 1624.55

574.38 572.93 515.22 593.38 577.41
13
The numerical stability of the computation will be compromised if the control coordinates are very
large by comparison with the model coordinates; that is, if the origin of the ground coordinate system lies
far outside the stereomodel (for relevant literature see, for example, Reinking, J.: ZfV 115, pp. 186-193,
1990). For this reason one should ignore as many of the leftmost digits of the ground coordinate system
as are identical for all control points.
The observation equation system (4.4-6) or (4.4-7) then becomes:
(\ 3.361 0.574 - 1 . 1 7 2 \ I -38.89 \

1 1.172 - 0 . 5 7 4 3.361 -4.18
( dXu \
1 0.574 1.172 -3.361 +5.10
dYu
1 3.423 0.573 -2.059 -19.83
dZu
1 2.059 - 0 . 5 7 3 3.423 +2.20
dmlO3
1 0.573 2.059 - 3 . 4 2 3 +3.87
dti 103 -51.19
1 1.828 0.515 -1.169
1 1.169 - 0 . 5 1 5 1.828 <2Φ103 +27.57
1 0.515 1.169 - 1 . 8 2 8 V irfao 3 / -22.03
1 0.593 1.967 -2.077 \-18.76
The following unknowns are found from Equations (4.4-8):
dXu - 8 2 . 5 0 m dSi -0.00238 —0.152 gon -8'12"

dYu 54.29 m d4> -0.01726 - 1 . 0 9 9 gon -53'24"
dZu - 5 5 . 1 6 m dK -0.02083 — 1.326 gon —1°04'25"
dm 0.008666
From Equation (4.4-4), using (2.1-13) to compute a rigorous form of the rotation ma-
trix, the following ground coordinates are found for the new point 45:
-82.50 \ / 0.999634 0.020825 -0.017256'

54.29 + 1.008666 -0.020787 0.999781 0.002380
-55.16 / \ 0.017302 -0.002020 0.999848
3353.82 /3323.22s
χ I 1624.55 1623.62
-
577.41 \ 582.39
In a manner identical to that for point 45, the model (control) points 23, 24, 50 and 51
can be transformed into the ground system:
[m] 23 24 50 51 45
XW 3320.54 3401.85 1776.20 2042.31 3323.22
y(i) 1166.84 2060.90 1196.29 1995.83 1623.62
zw 580.37 578.18 493.96 575.50 582.39
Using the (approximate) ground coordinates X"^ found in this way for the control
points, an absolute orientation can be once again computed. For the second iteration
the 1-vector of the observation equation systems (4.4-6) or, as the case may be, (4.4-7)
is as follows:
Γ = (1.11,0.72, -0.89,0.99,0.91, -1.38,0.50, -0.77, - 0 . 8 8 )

The unknowns derived in the second iteration are:
dXu = 0.05 m dh = -0.00034 = - 0 . 0 2 1 gon = - 1 Ό 2 "

dYu = 0.23 m άΦ = 0.00021 = 0.013 gon = 42"
dZu = —0.05 m dK = 0.00010 = 0.007 gon = 19"
dm = 0.000308
Now the ground coordinates for the new point 45 and the control points 23, 24, 50 and
51 become:
Μ 23 24 50 51 45
X(2) 3321.61 3402.86 1776.78 2042.90 3324.25
γ( 2) 1167.51 2061.15 1196.78 1996.62 1624.44
z w 579.40 576.90 493.28 574.52 581.27
At this point we break off the iteration process since, as can be seen from the diminution
in the unknowns from the first to the second iteration, a third computation would make
no contribution worth mentioning. From a comparison of the coordinates with the
given ground coordinates X, one obtains the following improvements v.
[cm] 23 24 50 51
vx -4 2 3
Vy -5 5 -1
Vz - 8 10 9 - 1 0
The weight coefficient matrix (Appendix 4.1-1): Q = ( A T A ) 1
/5.098 0.020 0.021 - 1 . 2 8 7 0.046 - 0 . 2 3 6 0.671 \

0.020 5.254 -0.518 - 0 . 6 7 3 0.622 0.040 - 1 . 2 6 5
0.021 - 0 . 5 1 8 6.379 - 0 . 2 5 2 - 1 . 9 7 4 1.071 - 0 . 0 7 0
-1.287 -0.673 -0.252 0.455 - 0 . 0 0 7 - 0 . 0 0 4 - 0 . 0 0 0
0.046 0.622 -1.974 - 0 . 0 0 7 1.415 0.111 0.059
- 0 . 2 3 6 0.040 1.071 - 0 . 0 0 4 0.111 0.473 0.009
V 0.671 - 1 . 2 6 5 -0.070 - 0 . 0 0 0 0.059 0.009 0.458 /
The standard deviation in an observed coordinate:
= ±12 cm expressed in the ground coordinate system
±0.015 mm in the model coordinate system

(1 : 8000)
±12 μπι in the image ( m B ~ 10000)
in which η = number of observation equations ( = 10)

u — number of unknowns (— 7)
The root mean square error in the elements of orientation: σ^ = σο sjqkk
&xu = ±0.28 m σα = ±0.0092 gon = ± 3 0 "

&Yu = ±0.28 m am = ±0.000081 σΦ = ±0.0053 gon = ± 1 6 "
σ χ = ±0.30 m σ κ = ±0.0052gon = ± 1 5 "
An assessment of these accuracies appears in the following section.
Exercise 4.4-1. Calculate the rotation matrix with which the xyz model coordinates
may be transformed directly into the X Y Z ground coordinate system. Hint: As in
Equation (2.1-14), the rotation matrix describing the resultant of successive rotations
is obtained by multiplication of the matrices for the partial rotations. The result is:
-0.020720 0.999640 -0.017045 \

-0.999783 -0.020677 0.002715
0.002361 0.017098 0.999851 /
Exercise 4.4-2. Repeat the numerical example assuming that point 50 is no longer a
control point. Result:
[m] 23 24 50 51 45
X 3321.65 3402.84 1776.95 2042.02 3324.26
Y 1167.56 2061.10 1196.74 1996.53 1624.44
Ζ 579.48 576.80 493.58 574.62 581.25
a X u = ±0.68 m σ Ω = ±0.0113 gon = ± 3 7 "

<ry" = ±0.69m a m = ±0.00018 σ Φ = ±0.0074gon = ± 2 4 "
a Z u = ±0.56 m σ κ = ±0.0115 gon = ± 3 7 "
Since, with this control point arrangement there is no redundancy, one would have been
able to solve the problem (iteratively) on the basis of the linear equation system (4.4-5).
The long way round using a least squares adjustment leads to the same result, including
the error computation (see Equations (4.1-1-11) and (4.1-1-12) in Appendix 4.1-1).
Estimation of the standard errors of the "corrections" is, however, not possible. We
have adopted σο = ± 1 2 cm from the numerical example above.
Exercise 4.4-3. Taking as a starting point the result of Exercise 4.4-1 and the result
of the relative orientation of Section 4.3.1, determine the rotation matrix defining the
orientation of each of the images with respect to the ground coordinate system.
Result:
0.006359 0.999836 -0.016932'
Rι = - 0.999947 0.006495 0.008057
0.008167 0.016880 0.999824
-0.033555 0.998674 -0.039037'
R2 -0.999437 -0.033488 0.002387
0.001077 0.039095 0.999235
Exercise 4.4-4. Repeat the above numerical example assuming that the xy coordinates
are more accurate by a factor two than the ζ coordinates. (Solution: The ground coor-
dinates differ by about 1 cm at most. The explanation for this slight disparity is to be
found in the introduction of Section 5.3.3.)
Further remarks: Every now and then, especially in close range work, the problem
arises in which one has to carry out a photogrammetric evaluation without control
points, either in a single model or in a block of photographs. In such cases an ar-
bitrary object coordinate system may be adopted. For example, one can measure a
distance S and adopt XYZ coordinates (0,0,0) and (S, 0,0) for the end points. If one
adopts the value Ζ — 0 for an (arbitrary) third point the object coordinate system is
defined, although with no redundancy. With an object coordinate system chosen in this
very arbitrary manner, a least squares estimation, which is in reality no such thing (see
Equations (4.1-1-11) and (4.1-1-12) in Appendix 4.1-1), will deliver (exterior) accura-
cies for the new points, dependent on the arbitrary selection of the object coordinate
system. (On the other hand accuracies for distances—interior accuracy—which can be
calculated from coordinate accuracies, taking into account correlations, are indepen-
dent of the object coordinate system chosen.)
In order to reduce the uncertainty associated with an arbitrary selection of object coor-
dinate system, as hinted at above, the free net adjustment, as it is called, was developed.
It is described in detail in Section Β 4.6.3 of Volume 2. To set up the theory of free
net adjustment one requires the normal equations for an over-determined similarity
transformation in three dimensions which can be derived from Equations (4.4-6). Let
the model coordinates and the object coordinates referred to origins at their respec-
tive centroids be denoted by χ and X respectively; when this is done the translations
dXu, dYu and dZu of Equations (4.4-6) vanish. With the substitutions X - X<0> —
dX, Y - Y(0> = dY and Ζ - Z<0) = dZ, the normal equation system in the case of η
full control points then becomes:
f [x{0)2+ Ym+ zm] 0 ο ο \

[F(0)2+^(0)2· (0) <0) (0) (0) /Η
] [-x F ] [-x z ] ™Λ
1 Γ_ν<°>¥°>1
symmetric "χ(0)2 - F (0)2 ]

(4.4-9)
/ [x(0)dx + ¥{0)dY + zl0)dz]\
[z{0)dx-x{0)dz]
The normal equation system (4.4-9) is also of service in the error theory of absolute
orientation which follows.
4.4.2 Error theory of absolute orientation

As a standard case we consider four control points in the corners of the model (Fig-
ure 4.4-la). Based on the normal equations (4.4-9) weight coefficients may be given
for the four relevant elements of absolute orientation, that is the three angles and the
scale. For this purpose we assume approximately flat level ground, that is Z <0i = 0, and
we take the dimensions of the rectangle at the corners of which the control points lie
as 92 mm χ 184 mm; under these conditions the matrix of the normal equations (4.4-9)
becomes a diagonal matrix:
qmm = y/\/(4 χ 46 2 + 4 χ 92 2 ) = 0.00486 m m - 1

qna = y/l/(4 χ 922) = 0.00543 m m " 1
2
ςΦΦ = y / l / ( 4 x46 )" = 0.01086mm" 1
<7kk = v/l/(4 x 4 6 + 4 x 92 ) = 0.00486 m m " 1
2 2
Concerning the accuracy of the scale number m and the rotation K, a crucial matter
is the accuracy in plan of the photogrammetric coordinates of the control points; the
accuracy of the photogrammetric coordinates of the control points in height is decisive
for the angles Ω and Φ. For well-measured control points these two accuracies amount
to about ± 6 μιη in the image and ±0.06%o of the principal distance c (for normal-angle
and wide-angle pictures) or ±0.08%o of the principal distance c (for superwide-angle
pictures) (Section 4.6). Using these accuracies for plan and height coordinates and the
weight coefficients given above, one obtains root mean square errors in this standard
case as compiled in Table 4.4-1.
Camera type Cm
Format, p.d. [cm] [10- 5 ] [mgoin (")] [mgoin (")] [mgoin (")]
Normal angle
±2.9 ±6.2(20) ±12.4(40) ±1.9(6)
(23 χ 23, c = 30)
Wide angle
±2.9 ±3.1(10) ±6.2(20) ±1.9(6)
(23 χ 23, c = 15)
Super-wide angle
±2.9 ±2.4(8) ±4.7(15) ±1.9(6)
(23 χ 23, c = 8.5)
Table 4.4-1: Accuracy of absolute orientation for different camera types
The accuracy of absolute orientation, like that of relative orientation (Table 4.3-1), is
independent of the image scale or of the object distance. As is Table 4.3-1, Table 4.4-1
is of considerable significance in practical photogrammetry. On the basis of this table
the operator, or the computer program, should judge in a particular case whether or not
and to what extent the given accuracy values are acceptable. Such an assessment is car-
ried out below for the numerical example of Section 4.4.1. The root mean square errors
arising in this particular case clearly lie above the values prescribed in Table 4.4-1. The
excesses require some elucidation:
• for the photo scale number m the unfavourable ratio is 2.8

( = 0.000081/0.000029). A factor 2 stems from the fact that in this particular
case an estimated accuracy in plan coordinates of ±12 μιη was used and in the
standard case an accuracy of ± 6 μηι is assumed. A further cause of the large
excess is the different number of plan control points; in the standard case there
are four (Figure 4.4-la) and in this particular case only three.
• for the heading angle Κ the ratio amounts to 2.7 ( = 0.0052/0.0019). The causes
of this large excess are the same as for the photo scale number just reviewed.
• it is conspicuous with respect to the inclination angles Ω and Φ that the rela-
tionship in the standard case between the higher accuracy in Ω and that in Φ
is reversed in this particular example. This apparent contradiction arises from
the Κ rotation of the stereomodel through approximately 300 gon (270°). Taking
account of this transposition, the ratios for the two angles are 1.7
(= 0.0053/0.0031) and 1.5 ( = 0.0092/0.0062). A factor of 1.3 arises because
the particular case under consideration is based on an estimated height accuracy
of ±12 /xm while the standard case an accuracy of ± 9 μιη (— 150000 χ 0.00006)
is assumed. A further cause of the relatively large excess lies in the fact that the
figure containing the control points in the image is somewhat smaller than the
92 mm χ 184 mm rectangle of the standard case.
b) c) e) f)
Full control point Height control point
a) standard situation
b) inclination Φ poor
c) inclination Ω poor
d) inclinations Ω and Φ indeterminate, if the control points lie exactly on a straight
line.
e) heading angle Κ and model scale number m indeterminate
f) heading angle Κ and model scale number m poorly determined
Figure 4.4-1: Assessment of different control point arrangements for absolute

orientation
Absolute orientation provides an important interim result in the photogrammetric pro-

cedure. For quality control of this important step the excess factor mentioned above
should remain below 2. If it exceeds this value there are causes and consequences as
follows:
• determination of the model coordinates of the control points insufficiently accu-

rate. To improve matters: repeat the measurements.
• poor distribution of the control points. In Figure 4.4-1 a number of poor control
point patterns are presented; Figure 4.4-Id) illustrates a seriously bad arrange-
ment of control points. Remedy: Computation of a bundle triangulation (Sec-
tion 5.3) with inclusion of all the images in the block.
From the viewpoint of reliability, a topic to be dealt with in Section Β 7.2.2.2 (Vol-
ume 2), the standard case of absolute orientation is improved by using pairs of control
points in the model corners. In this case the prescribed values of Table 4.4-1 reduce
by a factor 1 /y/2, as one may readily convince oneself with the help of the normal
equations (4.4-9).
The assessment above applies in the case of relatively flat models. In close-range
photogrammetry, and in aerial photogrammetry among very high mountains, the stere-
omodels should be surrounded by control points in all three coordinate directions. For
these cases specific reference values should be derived as outlined above.
Exercise 4.4-5. With the control point arrangement of Figure 4.4-Id) the angles Ω and
Φ cannot be determined. Show that this statement is true. (Solution: Consider the
system of normal equations (4.4-9).
For simplicity take = 0. Choose X '0> = aY1^. If the second equation is mul-
tiplied by the factor (-a) it becomes identical to the third equation. That is, the two
rows of the matrix are linearly dependent, meaning that the matrix is singular.)
Exercise 4.4-6 (Continuation of Exercise 4.4-5). Assess the result of this absolute
orientation with only three control points in the light of Table 4.4-1. (Answer: The
excess factors are 6.2 (m), 2.4 (Ω), 1.8 (Φ) and 6.1 (K). Since three of these factors
exceed 2, the absolute orientation must be rejected.
4.4.3 Determination of approximate values

Least squares estimation assumes approximate initial values for the unknowns. Meth-
ods to find these values, suitable for automation, are given in this section.
The first procedure assumes at least four full control points. For the connection of
the model coordinates (x, y, ζ) and the object coordinates (Χ, Υ, Z) the (linear) three-
dimensional affine transformation (2.1-17), with its twelve unknowns suggests itself.
The unknowns are three translations ao and nine matrix elements a,^. Each control
point provides three linear equations; with four full control points twelve unknowns
may be found. With more than four full control points a least squares estimation should
be carried out.
The question must still be resolved as to how one gets the seven parameters of absolute
orientation (4.2-6) from the twelve transformation parameters of the affine transforma-
tion. Comparison of the relationships (2.1-17) with (4.2-6) gives: αιο = Xu, a20 = Yu
and <130 = Zu. The other elements a,fc of the affine transformation (2.1-17) differ from
the elements r ^ of the three-dimensional similarity transformation (2.1-18) or (4.2-6)
mainly on account of the scale number m. The relationship
yja2u + α\λ Η l· ^33 = V3m (4.4-10)
gives an approximate value for the scale number m. Finally, all the elements α ^ can
be divided by the scale number m so that by means of the relationships (2.1-1-8) of
Appendix 2.1-1 the angles Ω, Φ and Κ can also be found 14 .
Numerical Example. We take up the numerical example of Section 4.4.1, but replac-
ing the height control point with a full control point.
Model coordinates
Μ 23 24 50 51
χ 0.303532 0.192638 0.303848 0.204120
y 0.595068 0.602834 0.403493 0.434574
ζ 0.034298 0.034116 0.026903 0.036672
Object coordinates or national coordinates
[m] 23 24 50 51
Χ 3321.65 3402.84 1776.75 2043.11
Υ 1167.56 2061.10 1196.79 1996.72
ζ 579.48 576.80 493.19 574.62
Setting up and solving the equations results in:
-166.95 8068.09 -107.88 x

/
A = -8069.40 -167.42 39.82
\ 20.49 137.51 8107.24
-0.020685 0.999639 -0.013366 '
R° = I -0.999802 -0.020743 0.004934
0.002539 0.017037 1.004490
Equation (4.4-10) gives ra° = 8070. The elements of the approximate rotation ma-
trix R° are formed as follows: R° = ( l / m ° ) A . Using relationships (2.1-1-8) of
"Further literature: Schmid, H.H., Heggli, S.: Mitt, des Inst, für Geod. und Photogr. an der ΕΤΗ
Zürich, Nr. 23, 1978.
Appendix 2.1-1 one obtains the following approximate values for the angles: Ω 0 =
- 0 . 3 gon = - 1 6 ' , Φ 0 = - 0 . 8 gon = - 4 3 ' and K° = 298.7 gon = 268°50'.
The solution using the three-dimensional affine transformation has two disadvantages:
on the one hand, the method breaks down when the four control points lie in a plane
and, on the other hand, in the presence of only three control points, which is the basic
requirement for an absolute orientation, it cannot be employed. A solution free of
these disadvantages can be found by creating orthogonal vectors in both coordinate
systems from three full control points. Details of this procedure are to be found in
Section Β 4.1.2.2 in Volume 2. Another method, with a detailed example, is included
in Section Β 5.1.1.3, again in Volume 2, which solves for the absolute orientation of
near-vertical photographs without the need for initial approximate values.
4.5 Image coordinate refinement

Image coordinates taken from photographs must be corrected in different respects. This
correction is known as image coordinate refinement. We have already learnt of a num-
ber of corrections:
• correction for lens distortion (Equation (3.1-3)).
• correction for film deformation in metric cameras (Section 3.2.1.2) and in semi-
metric cameras (Section 3.8.4).
• correction for aplanarity (lack of flatness) of the CCD detector array (Section
3.3.4)
• correction for positional errors of detectors in CCD cameras or in film scanners

(Section 3.4.3).
To meet demands for high accuracy one must also eliminate the influence of atmo-
spheric refraction and take into account the curvature of the Earth.
4.5.1 Refraction correction for near-vertical photographs

Because the atmospheric pressure, the temperature and the humidity vary along the
path of a ray of light, so too does the index of refraction, which is dependent on these
three quantities. This means that the rays which give rise to a photograph are bent.
The position of the image point P' must, therefore, be corrected by an amount Ap in
order to get the hypothetical metric image which would have been produced by central
projection (Figure 4.5-1).
In image space, the tangent to the curved light ray will appear to represent the direction
OP. In the case of a near-vertical photograph the angle Ar between the light ray and
Section 4.5 Image coordinate refinement 231
Ο (Χο,Υο,Ζο)
Figure 4.5-1: The effect of atmospheric refraction on a vertical photograph
the straight line OP can be represented with sufficient accuracy by a function of the
angle τ and a coefficient Κ:
At = Κ tan τ = Κ- (4.5-1)
c
It goes without saying that the coefficient Κ will vary according to the meteorologic
data at the time of a particular photograph and, since the index of refraction varies
with the wavelength, also with the wavelength to which the sensor is responsive. For
a normal atmosphere stratified parallel to the surface of the Earth and for light in the
visible wavelengths, Κ can be estimated as follows: 15
Κ = 0.00241 ( ^ 2 _ + 2 5 0 ~ Zo(Z2 - 6Z + 2 5 0 ) ) ^
in which Zq = flying height above sea level (in [km])

and Ζ = height of the ground (in [km])
Normally it suffices to use approximate values of Zq and Ζ representative of the whole
area being surveyed.
l5
Harris, W.D., Tewinkel, S.C. and Whitten, C.A.: Analytic Aerotriangulation. Coast and Geodetic
Survey, Technical Bulletin No. 21, July 1962, corrected July 1963. Schut, G.H.: Ph.Eng. 35, pp. 7 9 - 8 6 ,
1969.
It remains only to give the relationship using which the radial correction Ap can be
found from AT (Figure 4.5-1):
Ap C O S T SS Δ Τ Α / C 2 + p2
c2+p2
Ap « AT
c
Taking Equation (4.5-1) into account:
+ ^ (4.5-3)
Finally, using the relationship (3.1-2), the correction Ap can be resolved in the two
coordinate directions ξ and η.
In Table 4.5-1 the expressions (4.5-2) and (4.5-3) are evaluated for particular cameras,
photo scales and radial distances.
Correction Ap [μπι]
Photo scale c [mm] Zo [km] ρ = 90 mm ρ = 130 mm
1 : 10000 300 3.5 3 5
150 2.0 2 4
85 1.3 2 5
1 : 30000 300 9.5 8 12
150 5.0 6 11
85 3.0 6 13
1 : 100000 85 9.0 15 34
1 : 800000 300 240 1 2
Table 4.5-1: Correction of image coordinates for the effect of atmospheric refraction
(Ground height Ζ = 0.5 km)
One sees from this table that the corrections reach a sufficient number of micrometers
to be taken into account in processing only in small-scale photographs, above all in
the case of super-wide photography. In the case of very precise aerotriangulation for
larger-scale projects, however, refraction also needs to be considered. The last row
shows that satellite images are not adversely affected by refraction.
In conclusion it must be mentioned that the Formulae (4.5-2) and (4.5-3) are valid only
in the case of near-vertical photographs. With larger tilts it is necessary to move to
more general formulae16.
Exercise 4.5-1. An aerial photograph (c = 85 mm) was taken from a height of 9100m
above sea level. Calculate corrections for the effect of atmospheric refraction at 25
points of a 5 cm χ 5 cm rectangular raster. In the left-hand side of the image the ground
16
See, for example, Rinner, K.: JEK, Band IIIa/1, §§ 22-23.
is flat and level with a height of 600 m. It rises from the mid-line of the photographs
with a slope of 30%. (Answer: In two corners of the picture Ap reaches 37 /im; in the
other two corners Ap = 42 pm.)
An important special case concerns horizontal photographs in terrestrial photogramme-

try. In this case the influence of atmospheric refraction is usually dealt with together
with Earth curvature, although treating it separately is didactically more correct. Re-
fraction is actually a physical error in the formation of the image while Earth curvature
is a geometrical "error" in the ground coordinates.
4.5.2 Correction for refraction and Earth curvature in horizontal

photographs
As has already been indicated, the effect of Earth curvature is not a photogrammetric
error but is related simply to the definition of coordinate systems. Photogrammetry
provides coordinates in a three-dimensional Cartesian coordinate system in which the
position of an object point is fixed by orthogonal projection on the respective coordi-
nate planes. In the field survey system, on the other hand, heights of ground points
are referred to the ellipsoid while the X Y coordinates are more or less distorted by
the mathematical projection of the ellipsoid onto a plane. For this reason the ground
survey coordinates of control points should first be transformed into three-dimensional
Cartesian coordinates before the start of the photogrammetric work.
Fortunately, however, the differences between the ground survey coordinate system
and a three-dimensional Cartesian coordinate system are small so that in practice the
problem can be simplified. In Volume 2, Section Β 5.4, however, a full discussion
appears. Here and in the following section the discussion is restricted to approximate
solutions for the coordinate most strongly affected, namely the ground height.
First, the effect of Earth curvature combined with atmospheric refraction for horizontal
photographs is dealt with. The Earth is regarded as a sphere of radius R — 6370 km.
The symbol Ζ is used for heights referred to the sphere as opposed to Ζ for heights
referred to the three-dimensional Cartesian system (Figure 4.5-2).
The effects of Earth curvature AE for the point Ρ are given by the well-known formula:
(4.5-4)
Standard textbooks on surveying 17 give an approximate equation 18 for trigonometric

height measurement which takes account of Earth curvature and the influence of re-
fraction:
17
For example Kähmen, Η., Faig, W.: Surveying. 578 p., Walter de Gruyter, Berlin, 1988.
18
The approximation in this relation is too crude for large distances and large height differences, con-
ditions which, however, seldom arise in photogrammetric practice.
Figure 4.5-2: Atmospheric refraction and Earth curvature for a horizontal photograph
% Ι 2 A2
Ζ = ΖΫ + Αtana + Atana h(l-fc) — (4.5-5)
2R 2R
in which k ... refraction coefficient, usually taken as 0.13
For photogrammetric point determination with horizontal photographs tan a may be

replaced by the ratio η/c:
2
—
z = —z0 + vΑ-+
c
ΑΖο + Ζ ^ , , Α
v-t^—
c In
+ (l-k)—
2H
* ^
(4.5-6)
What is required, however, for use with the three-dimensional object coordinate system
is the significantly simpler relation:
~Z = ~ZQ + —(η + Δη) (4.5-7)

c
To compensate for the use of heights referred to the sphere rather than Ζ coordinates
from the Cartesian coordinate system, the correction Αη is introduced to the image
coordinates η. Subtraction of Equation (4.5-7) from (4.5-6) gives:
A AZP + Z I N A2
Finally, re-arranging gives the desired formula:
+ (4.5-8)
If the image coordinates η are corrected by an amount Αη and if the photogrammetric

measurements are made in the usual way, then the height coordinates, whether of con-
trol points or of new points, will be referred to the sphere; they will also be free of the
influence of refraction. For the computation of the correction Αη, however, the distance
A and the value Ζ are necessary. If Ρ is a new point the following procedure, suitable
for automation, is recommended:
• first measure the point photogrammetrically without correction for Earth curva-
ture and refraction
• using the result so obtained, compute A, Ζ (with adequate accuracy) and Αη
• finally, repeat the photogrammetric computation using the corrected image coor-
dinates.
Table 4.5-2 gives some typical values for the correction Αη.
A [km] 0.1 0.25 0.5 1 5 20

Αη [μπι] 0.7 1.7 3.4 6.8 34 137
Table 4.5-2: Corrections, Αη, to image coordinates for the influence of refraction and
Earth curvature (c = 100 mm, k = 0.13, R = 6370 km, η = 0)
It will be seen that, up to a few hundred metres object distance the correction is smaller
than the measuring accuracy of a few micrometres, but may not be ignored for distances
greater than about 1 km.
Exercise 4.5-2. Compute the correction Αη and the Ζ coordinate for an extreme case
with A = 5000m, Z0 = 1000m, η = 30mm, c = 100mm. (Result: Αη = 42.4/xm,
Ζ = 2502.12 m). Additional exercise: Compute the Ζ coordinate using a refraction
coefficient of 0.10. (Result: Αη = 43.6/zm, Ζ = 2502.18.)
Exercise 4.5-3. Using the data of Table 4.5-2 compute the standard errors of both
the correction Αη and the ground height Ζ under the assumption that the refraction
coefficient k — 0.13 has a standard error of ±0.03. (Solution: for example, if A —
5 km: σ&η = ±1.2 μπι, σ-χ = ±5.9 cm.)
4.5.3 Earth curvature correction for near-vertical photographs

In this section the influence of and correction for Earth curvature in near-vertical pho-
tographs is considered. A stereomodel produced photogrammetrically in the usual
manner from two aerial photographs is referred to the three-dimensional Cartesian co-
ordinate system X Y Z (Figure 4.5-3). Heights from land survey, however, refer to a
sphere with the Earth radius R.
Earth curvature can be taken into account in the photogrammetric process as now de-
scribed. Control point heights Z, referred to the curved datum surface Ζ ο = 0, are
reduced by the correction AE from Equation (4.5-4), in which A is the approximate

distance of the particular control point from the centre of the stereomodel. The result
of this operation is that the control point heights are referred to the tangent plane at the
centre of the model, the X Y plane of a Cartesian coordinate system. Subsequently all
object points are determined in this X Y Z system using the computed image coordi-
nates; finally all Ζ values are increased by the amount AE (Equation 4.5-4).
It is also possible to take account of Earth curvature by "correcting" the image coordi-
nates in aerial photographs in a similar way to the correction in horizontal photographs.
This "Earth curvature correction" is applied in the same way as the correction for re-
fraction. It has the opposite sign, however, and is significantly larger than the Ap
refraction correction (Equation (4.5-3)). This makeshift Earth curvature correction is
dealt with in more detail in Section 5.6.
So that an assessment may be made of the extent to which a correction for Earth cur-
vature is important, the maximum correction AZ?max in parts per thousand of the flying
height is given in Table 4.5-3 for various photo scales. The value Δ£ m a x (Figure 4.5-3)
comes from Equation (4.5-4) in which By/5/2, the semi-diagonal of the model, is
substituted for A:
T^)2 5 B2
^ = ^2ll· = Tr (4·5"9)
In view of photogrammetric height accuracy of 0.06%o for a measured point (Sec-

tion 4.6), Table 4.5-3 shows that Earth curvature must be taken into account:
• for super-wide angle photographs from a photo scale of about 1 : 6000,
• for wide angle photographs from a photo scale of about 1 : 10000 and
• for normal angle photographs from a photo scale of about 1 : 20000.
Maximum Earth curvature correction ΔE max

Photo scale
Η (%0 of the flying height)
c = 85 mm c = 1 5 0 mm c = 3 0 0 mm
1:10000 0.08 0.10 0.06 0.03
1:50000 2.1 0.49 0.28 0.14
1:100000 8.3 0.98 0.56 0.28
Table 4.5-3: Maximum Earth curvature correction AEmax at the corners of a stereo-
model with 60% forward overlap
4.5.4 Virtual (digital) correction image

In digital photogrammetry every completed correction to be implemented in the context
of image coordinate refinement entails a resampling, that is a restructuring of the pixels
in a new orthogonal raster (Section 2.2.3). As a rule every resampling is accompanied
by loss of information. In view of the relatively large number of corrections which
mount up in the course of image coordinate refinement a solution is sought which, if
possible, manages with one single resampling.
With a virtual correction image 19 one achieves this aim. The procedure is explained
with reference to Figure 4.5-4; it is a typical example which can readily be generalized.
Let the digitized photograph have an arbitrary ξ'η' coordinate system (Section 3.4). A
corrected image matrix is required in the ξη system, the coordinate system of a digital
metric camera or a digital metric image (Figures 4.5-4 and 3.1-10). As in indirect
resampling (Figure 2.2-5), an orthogonal raster is first defined in the digital image
19
The idea was adopted from T. Schenk (Digital Photogrammetry. TerraScience, Volume I, 1999).
•η'
Lens distortion
Refraction Δρ
(ξ,,η') = ί·,(ξ,η)
Principal point
V V
ξ ξ'
Figure 4.5-4: Digital metric image, in its position relative to the camera (left), with
corrections for refraction and lens distortion; and transformed into the photograph as
digitized (right)
(Figure 4.5-4, left). Corrections are then computed for the mid-point of every pixel; on
the left of Figure 4.5-4, for example, corrections have been introduced for refraction
(Equation (4.5-3)) and lens distortion (Section 3.1.3). These corrections are recorded
in a virtual correction image with a resolution in the subpixel range. The last step is the
transformation into the digital photograph, an inverse transformation which is based on
the fiducial marks and which eliminates film deformation; depending on the number
and arrangement of the fiducial marks, the parameters of this transformation may be
chosen accordingly (Section 3.2.1.2).
The point whose position has been computed in the above three steps, is marked with
a small circle on the right of Figure 4.5-4; finally, a grey level interpolation is carried
out, as described in Section 2.2.3, and the interpolated grey value is transferred to the
digital metric image. This position is likewise marked with a small circle in the left of
Figure 4.5-4.
By choice, the raster spacing in the digital photograph is somewhat smaller than the
digitizing interval in the digitized photograph. The result is a digital metric image
which can be evaluated in subsequent calculations involving the collinearity equations
(Section 2.1.3) and the coplanarity equations (Section 4.2.3).
4.6 Accuracy of point determination in a stereopair

When the elements of exterior orientation are known the object coordinates of a new
point can be computed directly (Section 4.1). If the elements of exterior orientation
are unknown the object coordinates of new points may be obtained as a by-product of
"two picture resection" (Section 4.2.2). The other orientation procedures permit the
computation, using these orientation elements, of new points in the stereomodel.
Section 4.6 Accuracy of point determination in a stereopair 239
A stereopair from aerial survey approximates to the "normal case". Therefore, for
estimating the achievable accuracy of 3D point measurements in a model the theory
described in Section 2.1.7 can be applied for any photo flight, if the flight parameters
are known and information about the measurement accuracy of image coordinates is
given.
Leaving aside extremely large-scale aerial photographs (perhaps subject to image
movement) the accuracy of object coordinates X and Y (plan accuracy) is directly pro-
portional to the image scale and is constant in relation to the image (Equation (2.1-35)).
The camera type (normal angle, wide angle and so on) has no influence on the accuracy
in plan.
As may be seen form Equation (2.1-33), the height accuracy σζ is directly dependent
on the measurement accuracy of the parallaxes, as well as being indirectly dependent
on the base-height-ratio (B/Z) and on the photo scale ( c / Z ) . σζ may also be regarded
as either linearly or quadratically dependent in the camera-object distance Z. In both
cases represented in Figure 4.6-1 the height accuracy σ ζ is linearly dependent on the
camera-object distance Z. In the diagram on the left of Figure 4.6-1, both the base-
height BIZ and the principal distance c are the same for both camera arrangements;
in the right-hand diagram the photo scale number mB = Z/c and the base B, and
consequently also the forward overlap, are the same for both. The proportionality
between the height accuracy σ ζ and the object distance for different camera types holds
true only for the range from normal-angle to wide-angle cameras; the relationship is
weaker for super-wide-angle photographs20.
V V V
\ 7
/
/
//
// s
>>>>I>/H>l'>>>»ll»i*
Zj Β = const., c = const. Β = const., Zjc — const.
Figure 4.6-1: Height accuracy directly proportional to the camera-object distance
20
Meier, H.K.: BuL 38, pp. 50-62, 1970. Stark, E.: BuL 44, pp. 5 - 1 4 , 1976. Sievers, J., Schürer, Κ.:
Verfeinerter Ansatz vor allem für variables Basisverhältnis und unterschiedliches Auflösungsvermögen.
BuL 50, pp. 101-118, 1982.
From Equation (2.1-33) rules of thumb can easily be derived for rough estimation of
accuracy. In this we limit ourselves to giving information on accuracy which is repre-
sentative of the whole stereomodel. Thus we ignore readily available information (see
Section Β 5.2.3.1, Volume 2) about accuracies which vary within a stereomodel. To a
large extent the information on accuracy is based on empirical investigations.
The conditions under which these rules of thumb apply are:
• forward overlap of 60% (possible cases deviating from this are specifically delat
with; for example in the Numerical Example 2 of Section 6.7.2d)
• at least four full control points in the corners of the model (indirect georefer-
encing); the control coordinates must exhibit markedly better accuracy than the
measured coordinates. Alternatively elements of exterior orientation from GPS
and IMU (direct georeferencing) with corresponding accuracy
• image format 23 cm χ 23 cm (for film-based metric cameras)
• good image quality such that with analytical plotting instruments a coordinate
measurement accuracy of ± 6 μπι at image scale can be achieved
The following values are applicable as guides in the case of targeted points or equally
accurately defined points:
Plan: CTxyjtajg) = ± 6 μπι χ image scale number m s

Height: a X y( t a r g ) = ±0.06%o of height above ground
(normal angle and wide angle) (4.6-1)
±0.08%o of height above ground
(super-wide angle)
If, in determining the object coordinates, one eliminates the systematic part of the im-
age coordinate errors by means of additional parameters in the mathematical model, a
further increase in accuracy of up to 50% is possible (for details see Section 5.3.5). On
the other hand poor quality image material can cause distinct reduction in the numerical
values of the rules of thumb (4.6-1).
These rules of thumb also apply to digital images if they have been generated by scan-
ning aerial films with an appropriate pixel size. In practice, scanning films with 15 μπι
to 20 μπι has turned out as acceptable trade-off between sufficient radiometric quality
and good resolving power. Experience has also shown that for manual measurement if
well defined points in digital images an accuracy of 1/3 of a pixel can be achieved, i.e.
5 μπι to 7 μπι in the case of the above mentioned pixel sizes.
If the above-mentioned conditions do not apply, the rules of thumb will deliver un-
realistic results and Equation (2.1-33) should be used only for rough estimation; for
example, if:
• the accuracy of the image coordinates measurement differs significantly from

6 μπι. Automatic point measurement using, for instance, lest squares matching
(see Section 6.8.1.2) can reach accuracies of 1 /5 to 1 /10 of a pixel and thus much
higher accuracy. On the other hand, measurements in images digitized with large
pixel sizes might be less accurate. In both cases the rules could still be applied,
if the results are multiplied by the ratio of the actual measurement accuracy to
6 μιη.
• other photo flight parameters were used, for instance, forward overlaps signifi-
cantly greater than 60%.
• digital aerial cameras were employed (such as Microsoft-Vexcel Ultracam, Inter-

graph DMC, Leica Geosystems ADS40). For digital cameras a generally valid
rule of thumb is hard to define. There is no standardized principle for digital
cameras and the geometric arrangement for image acquisition may differ signif-
icantly.
Numerical Example. Images taken with a Zeiss RMK 15/23, with a photo scale of 1 :
5000 and 60% forward overlap, were used for photogrammetric point determination.
What is the accuracy of photogrammetrically derived coordinates when targeted points
are involved?
Using the rules of thumb (4.6-1):
σχγ = 5000 χ 0.0006 = ± 3 cm

σζ = 5000 χ 15 χ 6 χ ΙΟ" 5 == ±4.5 cm
Camera arrangements can occur in practice, especially in high mountains, where the
height accuracy varies quadratically with the camera-object distance. In all the cases
sketched in Figure 4.6-2 the base Β and the principal distance c remain the same;
only the camera-object distance varies. The left-hand example represents a flight over
stepped landscape using a constant base and constant height above sea level; in the
right-hand diagram the bases are equal but the flying height varies.
Numerical Example. The second stereomodel on the left of Figure 4.6-2 represents
pictures from a flying height above ground which is double that of the first stereomodel;
the forward overlap in the second model is about 80%. As a result, the height accuracy
given above (that of the first model) worsens as follows:
σχγ = ± 3 x 2 = ± 6 cm
σζ = ±4.5 χ 4 = ±18cm(!)
In comparison with the model with the lower flying height on the right of Figure 4.6-2,
the same deterioration of accuracy also arises in the stereomodel with the greater flying
height. Incidentally, an even more severe reduction in accuracy arises in that area of
the model which lies outside, in the direction of the base, the area between the image
centres. There is certainly a great temptation in practice to use a stereomodel with 80%
forward overlap covering the entire common area—with a base which is too short.
V V V V X7
B,c — const. B,c = const.
Figure 4.6-2: Height accuracy proportional to the square of the camera-object distance
Exercise 4.6-1. A photoflight has been planned over extremely mountainous country
with a forward overlap of 60% and a base of 1000 m. Under these conditions the
Formulae (3.7-1) show that for a 30 cm camera the flying height over the summits
will be 3260 m and for a 21 cm camera it will be 2280 m. What accuracy is to be
expected in the region of the summits and what in the valleys which lie 1000 m beneath
the summits? (Answer: In the summit area, for the 30 cm camera, σζ = ±19.6 cm,
and for the 21 cm camera σ ζ = ±13.7cm. In the valleys, for the 30cm camera,
σζ = ±33.5 cm, and for the 21 cm camera σζ = ±28.4 cm.)
It is worth noting that the superiority, by a factor of 1.43 ( = 19.6/13.7), of the 21 cm
camera over the 30 cm camera in the region of the summits is reduced in the valleys to
1.18 (= 33.5/28.4). 21
In case of natural (non-targeted) points, the guide figures (4.6-1) for the accuracy must
be increased by the uncertainty of definition. This uncertainty of definition can be
related to the image as well as to the object.
In principle, the uncertainties of definition related to the image and their effects on the
standard deviation of the object points may be found by repeated measurements:
21
The suggestion for this example came from Prof. Dr. P. Waldhäusl.
2
'UYi --Ϋ)
σχγ
(°*) = \ — ^ T f τι —1
(4.6-2)
Ι
σ
IT.iZi-ZY
Έ(ζι- - Ζ ) 2
Ζ { nat)
η - 1
where η . . . number of measurements and Χ , Υ, Ζ are arithmetical means
Values related to the image depend on the image quality, the film type (colour, black&
white, etc.), the pixel size, etc. In a practical project one will determine the accu-
racy values σ Χ Υ ( η 3 ή and σ Ζ ( ηΆΙ) for a number of types of points using the relation-
ships (4.6-2). In the most favourable case one obtains σ ΧΥ ( ηίΛ } and σ Ζ ( ηίΛ) , the accura-
cies (4.6-1) for targeted or very precisely defined points.
The uncertainty of definition related to the object may also be determined by repeated
measurements; these can be made either terrestrially or photogrammetrically using a
very large photo scale. Values of the related to the ground for some types of points
have been collected in Table 4.6-1.
Type of point Plan Height

σ
ΧΥ(0,def) [cm] σζ(ο,def) [cm]
House and fence corners 7-12 8-15
Manhole covers 4-6 1-3
Field corners 20-100 10-20
Bushes, trees 20-100 20-100
Table 4.6-1: Uncertainty of definition of natural points 22
Starting with the object related uncertainty of definition, one must summarize by con-
sidering both contributionsCT(targ)and def ):
Planimetry: a X Y ( n a t ) = 7 4 y ( t a r g ) + 4 υ ( ο , < μ )
/ (4.6-3)
Height: <7 z(nat) = ^/4(targ) + 4 ( 0 , d e f )
Numerical Example. With what accuracy can the corners of houses and fields be
measured from wide-angle photographs with about 60% forward overlap and a photo
scale of 1 : 15000? (Use relevant mean values from Table 4.6-1.)
House corners: σχγ - ^ ( 1 5 0 0 0 χ 0.0006) 2 + 9.5 2 = ± 1 3 cm
σζ = VX15000 x 15 x 0.00006) 2 + 11.5 2 = ± 1 8 c m
Field corners: σχγ = ^ ( 1 5 0 0 0 χ 0.0006) 2 + 60 2 = ± 6 1 cm
σζ = y/(15000 χ 15 χ 0.00006) 2 = ±20cm

22
Waldhäusl, P.: Presented Paper, Commission IV, 14. ISP Congress, Hamburg, 1980.
The standard deviation of a distance a s can be found directly from the coordinate error
σχγ·
= σχΥ\ίϊ (4.6-4)
The accuracy of distances derived from photogrammetrically measured coordinates is
essentially independent of the distance.
Exercise 4.6-2. Derive Formula (4.6-4) using the rules of error propagation.
In the field of close range photogrammetry one is often less interested in the absolute
accuracies σχγ and σζ and more interested in relative accuracy with respect to the
largest dimension of the object S. For an object which fills the whole of the format of
a standard aerial camera, therefore, the relative accuracy amounts to:
ψ = ^ = ^ ^ = ^ = 2-6 χ 1 0 - ^ 0.026%o (4.6-5)

S 23cm 23 38000
We have, incidentally, already found this value to be the accuracy of the scale factor in
absolute orientation (Table 4.4-1: ±0.000029 = 0.029%o).
Especially on account of the small format of close range cameras one obtains lower
relative accuracies than with large format aerial survey cameras. For a semi-metric
camera with medium format (6 cm χ 6 cm, Section 3.8.8) the relative accuracy, likewise
applying the value σξ η = ± 6 μηι, comes to:
= ±00006 = = 1 0 - =0 (4.6-6)
6 cm 6 10000
In close range photogrammetry the photographs are frequently arranged in a less regu-
lar pattern than in aerial survey (Section 3.8.8). For this reason it is necessary to extend
the simple accuracy guide (4.6-1) appropriately to suit the diverse camera arrangements
of close range work. This topic is dealt with in detail in Volume 2 (Section Β 4.5.2).
The empirical accuracies given by the relationships (4.6-1), (4.6-5) and (4.6-6) are
chosen fairly conservatively. They can be considerably improved by means of further
mathematical modelling of the photogrammetric process. This is considered in more
detail in Section 5.3.5.
The rules of thumb for accuracies of photogrammetric processing given in this section
and in Section 6.7 are widely accepted in photogrammetric practice. In the calling
of tenders for photogrammetric work, however, tolerances enjoy a greater degree of
acceptance than the standard deviation or root mean square error. Tolerances and their
empirical determination are considered in more detail in Appendix 4.6-1.
Exercise 4.6-3. For purposes of road construction the heights of targeted points are
to be found by photogrammetry with an accuracy of ± 5 cm using a super-wide angle
camera. If the forward overlap is to be 60%, what are the photo scale and flying height?
(Answer: photo scale = 1 : 7350, flying height = 625 m.)
Supplementary Exercise. What reduction in accuracy arises if, over a valley region
200 m lower, the base is not increased? (Answer: The height accuracy in the valley
area is a ζ — ±8.7 cm.)
Exercise 4.6-4. Is the alignment of the image rectangle of the 271 Imaging DMC (Fig-
ure 3.7-10) well chosen for a favourable height accuracy? (Answer: No.) How would
the heighting accuracy be changed if the format were to be rotated by lOOgon (90°)?
(Answer: By a factor 1.7 (= 13500/8000).) What advantage is there in the align-
ment of the rectangular format chosen by Z/I Imaging? (Answer: Fewer strips of
photographs for a given flying height.)
Chapter 5
Photogrammetric triangulation
We discuss first photogrammetric triangulation with aerial photographs, known as aero-

triangulation or aerial triangulation. We assume metric images recorded onto film or
CCD area arrays. In Section 5.5 we will discuss the digital 3-line camera.
5.1 Preliminary remarks on aerotriangulation

Even without GPS and IMU information, aerotriangulation frees photogrammetry from
the constraint of determining at least three control points by ground survey methods in
every stereomodel. It is possible to bridge areas without such ground survey points.
In Chapter 4 we limited the discussion to methods of data capture in two overlapping
photographs, but in this chapter we treat the subject of simultaneous data capture in a
block of such overlapping photographs.
The results of aerotriangulation are the orientation elements of all photographs or stere-
omodels and the X Y Z coordinates of discrete points in a global coordinate system
(usually the ground coordinate system). We speak then of photogrammetric point mea-
surement. The points may be:
• points targeted before the flight

• selected "natural" points in the photographs, usually with accompanying identi-
fication sketches
• points automatically selected in digital metric images, their image coordinates
thus being known (Section 6.8.3.5 and Section Β 2.4, Volume 2)
Figure 5.1-1 illustrates the principle of aerotriangulation, without reference to support-

ing GPS and/or IMU information (this aspect will be discussed in Section 5.4). The
example shows 18 photographs in three strips. The forward overlap of photographs
within a strip is about 60%, the side overlap between strips about 25%. For the ori-
entation of this set of photographs, we have four full control points and four height
control points. We require the orientation elements and the ground coordinates of new
points. These new points tie models together within a strip and tie neighbouring strips
together. For example, new point 7 occurs only in the photographs 12 and 13, new
point 5 in the photographs 11, 12 and 13, new point 6 in the photographs 11, 12, 13,
21, 22 and 23.
Section 5.1 Preliminary remarks o n aerotriangulation 247
A . · · . A
8
1 3
Strip 1 © 4 © © © © ©
O .6 . . . . 0
2 · 10 I
Strip 2 © © © © @ *©
° . · . . · · °
• ·
Strip 3 © © @ ©" © . ©
A · · · . A
A Full control point © Nadir point with photo number

° Height control point
. New or tie point
Figure 5.1-1: Set of photographs for a block adjustment
We have already introduced two techniques which can be extended to deal with blocks
of photographs. The first is the numerical orientation of the two bundles of rays of a
stereopair of photographs (Section 4.2.2). In this technique, the 12 elements of exterior
orientation of the two photographs and the coordinates of new points in the stereomodel
are computed from the known coordinates of control points. The method can be readily
extended to a large number of photographs, e.g. 18 in the example of Figure 5.1-1. The
method is then known as a bundle block adjustment. It is discussed in Section 5.3.
The second technique is numerical absolute orientation (Section 4.4). It is based on the
assumption that a stereomodel has been formed by the numerical relative orientation
of two overlapping photographs (Section 4.3). This stereomodel is then transformed
into the ground coordinate system with the help of control points. The extension of
the technique from a single model to a block of models implies that all models in
the block—15 in the example of Figure 5.1-1—are absolutely oriented simultaneously.
The method is then known as a block adjustment by independent models. It is described
in the following section.
248 Chapter 5 Photogrammetric triangulation
5.2 Block adjustment by independent models

We start with a block of photographs with about 60% forward overlap between pho-
tographs in the individual strips and about 25% side overlap between neighbouring
strips. The adjustment of a single strip by independent models is a special case of the
more general block adjustment.
The adjustment of the block by independent models begins with the model coordinates
derived from the numerical relative orientations and formation of the stereomodels
(Section 4.3.2). In the course of the block adjustment the individual models will be
amalgamated into the single block and simultaneously transformed into the ground
coordinate system. The individual stereomodels are thus the basic units of aerotrian-
gulation by independent models.
As an introduction, we treat first the adjustment of planimetry only. This is a spe-
cial case of the general block adjustment which is a problem in three dimensions. A
purely planimetric adjustment is a plane problem, i.e. we are concerned only with XY
coordinates.
5.2.1 Planimetric adjustment of a block

We require the X Y coordinates of new points in the ground coordinate system. Given
are the model coordinates xy of the relatively oriented and levelled individual mod-
els. The levelling need only be approximate, but must be good enough to ensure that
the influence of the "tilted" model heights on the xy coordinates is less than the pho-
togrammetric accuracy of the planimetric coordinates (Figure 5.2-1).
Figure 5.2-1: Effect of levelling error on planimetric coordinates
The models can be levelled, for example, by means of control points taken from rela-
tively small-scale topographic maps. For this purpose, in addition to the actual aero-
triangulation points, points in the stereomodels are chosen whose heights, or all three
coordinates, can be measured approximately in a topographic map. The image coor-
dinates of these points are also measured and used, after the relative orientation, to
Section 5.2 Block adjustment by independent models 249
compute the corresponding model coordinates. The numerical absolute orientation can
then be performed to level the model with these (approximate) control points.
Exercise 5.2-1. Wide-angle photographs are taken at a scale of 1 : 8000 with a 15/23
camera. A topographic map at a scale of 1 : 5000 is available, from which coordinates
can be derived with an accuracy of ± 3 0 cm. How large may be the height differences
AZ in a model if we use this map for the levelling of the models before a planimetric
block adjustment? (Solution: σχγ = ±4.8 cm (Equation (4.6-1)), AZmm « 80m.)
The principle of a planimetric block adjustment is shown in Figure 5.2-2, while Ta-
ble 5.2-1 contains the initial data for the block adjustment. The model coordinates are
in separate, independent, local coordinate systems for each model. Each of these co-
ordinate systems is displaced and rotated relative to the ground coordinate system and
has an arbitrary scale.
As initial data for the amalgamation of the individual models into one block in the
ground coordinate system we have on the one hand the tie points which exist in more
than one model (points 5, 6, 7, 8 and 9) and, on the other hand, the model and ground
coordinates of the field-surveyed control points (points 1, 2, 3 and 4).
The adjustment can therefore be defined as follows. The models are:
• displaced (two translations Xu, Yu),
• rotated (rotation angle κ) and
• scaled (scale factor τη)
so that:
• the tie points fit together as well as possible and
• the residual discrepancies at the control points are as small as possible.
The mathematical relation between a stereomodel (coordinates x, y) and the ground co-
ordinate system (coordinates X, Y) can be formulated from Equations (2.1-9), (2.1-5)
and (2.1-4) (plane similarity transformation):
(5.2-1)
The non-linearities in the unknowns m and κ can be eliminated by substituting
TO cos κ = a
(5.2-2)
TO sin κ = b
We obtain thus the linear equations
X — Xu + xa — yb
(5.2-3)
Y = Yu + ya - xb
The extension of the system of equations (5.2-3) to a block of models is sometimes

called a chained plane similarity transformation. One possible formulation 1 of such a
chained similarity transformation as a problem of adjustment of indirect observations
is:
Observation equations for a control point:
vx = X_u + χ a — yb — X
(5.2-4)
Vy = Yu + ya + xb — Y
Observation equations for a tie point:
vx = X_u + xa - yb — 2L — 0 (5.2-5)
vy = Yu + ya + xb - Y - 0
The underlined terms are the unknowns. The observation equations (5.2-4) and (5.2-5)
have an unusual form: although the corrections vx and vy derive from the inaccuracies
in the model coordinates x, y, they are to be interpreted as corrections to the (known)
ground coordinates Χ , Y (in (5.2-4)) and to the fictitious observations "0" (in (5.2-5)).
The balance of unknowns and observations for the example of Figure 5.2-2/Table 5.2-1
is:
Unknowns: 4 x 4 = 1 6 transformation elements Xu, Yu, a, b
5 x 2 = 1 0 tie-point coordinates X , Y
==> 26 total
Observations: 4 χ 4 χ 2 = 32 model coordinates χ,y

=>• 6 redundant observations
The 32 observation equations for the 26 unknowns of the schematic example of Fig-
ure 5.2-2 are shown in Table 5.2-2.
After introducing the centroid coordinates χ and y, and assuming all observations have
equal weight, the normal system of equations (4.1-1) shown in Table 5.2-3 is obtained.
If we insert the numerical values of Table 5.2-1, we obtain the normal equations of
Table 5.2-4. The introduction of coordinates referred to the centroid of each model has
had the effect of reducing the submatrix of the transformation elements to a diagonal
matrix.
'Kraus, Κ.: ZfV 91, pp. 123-130, 1966. Van den Hout, C.M.A.: Phia21, pp. 171-178, 1966.
Model 1 Model 2
Pt.No. X y Pt.No. X y
1 148.29 573.28 2 366.93 558.43
5 374.11 561.87 5 154.36 561.30
6 362.77 147.41 6 130.40 143.24
7 138.27 151.39 8 358.30 140.28
Model 3 Model 4
Pt.No. X y Pt.No. X y
3 148.59 139.40 4 359.38 135.30
6 362.10 542.71 6 140.31 578.42
7 159.40 556.85 8 359.34 549.19
9 345.67 128.76 9 141.97 149.87
Ground coordinates of
control points in m
Pt.No. X Y
1 4443.81 8338.54
2 7658.37 7993.67
3 4472.02 1071.18
4 8348.54 1316.60
Table 5.2-1: Initial data for the planimetric block adjustment of the block of Fig-
ure 5.2-2 (Model coordinates in [mm], ground coordinates of control points in [m])
Model 1 Model 2 Model 3 Model 4

- * iff **
7 7
* 7 7
iff 7 7
*00 7 7
£ 7 7
Τ 7
iff 7 7 7 7
*
7 7 7 7
iff 7 7
7 7
^-oo ^^
It. Si^.f Ä-*1» Ss·»?
I Η ι Η ι Η ι Η
a H Ä H Ä H Ä H Ä
η-Ο Ί—r Ti vo ο co η»
i»"»
1 1 1 1
Τι ^t^nt^MvOr^vßiriMnMflOvfliy!
Ö
Ν Nm 1 ri CI csoo tsvo
-Ο Ä^r sun" a«,» Λ".?
[ I l l
Ν
a B Ä H S l H Ä H Ä
-_ —m — ^c — r-
JO
f "H f "h f •-H f "H
~a
~H* " S ""?» ""s» "H"
»3
X
„ „ ~ _
T*i 3
Νa - - - -
N
: 3
X
-><3
H3)H3IHÄHÄ HÄHÄHPiHÄ HÄHÄHPIHPI
S S S S P S S S i
Upper index = model number, lower index = point number
Table 5.2-2: Observation equations for Figure 5.2-2 and Table 5.2-1
r-* Ή
Χ * χ *
|
k ' f
ε λ ε^ϊ
£ 7 ι ι
Ή -e
ε^ r» Je
ΙΗ -C
* 7 I
CO • ' <0 «3 co .' 03 03
-«O •«^•«i ·«->
. . . «Ä Ή . • · ε-„ε·-,
IÄ ΙΗ
7 · 7 1 1 I I
ε·^ •f-i
7 •· 7 I IÄ •· « ?S? -C
(Ν '
ε
•13
6•o
<+s
ε
•S-
ίΤ
ε +
a
ίΓ
CO
3 ο
+ 'S
(Ν ω
>>
εΛ
"a O+l
CO
£
i? e
Es ε
s
cc
"e
»* 3
"e
Table 5.2-3: Normal equations of a planimetric block adjustment

ΟS oo >0 •o ο SO
C
P~ SO
00 >0 en σ\ cs
•tf Ο) co ρ οο
σ\ enoo
en 00 00 ci
cs 00 00 cn Ή «ΛΪ οό es -φ C
CO u-lD as
r~ t- Ul m 00 as < Ν VN
O £P
en V
00 Ρ» o en
Os
oo en oo >
00 en C/->
S VO
1
SO Ρ- V
I
OOs •3
7
CO cs cn
οCS OO
CS asο CS cω
£ 7 7 ö ο ο ο
I I IÖ
cs
as m Ο Οen
* 7 7 ο CS ο Ο cs CS ο
ο ö
I ο ö 3 'S)
SO
SO ο
as cΉ
CS Ο οw
£ 7 7 ÖΟ σ\ ο cs δrt1 ο§
I I ο I ^ 'S
σ·\ 'S
ο ο Os •5 μ
*00 7 7 ο ο CS cs > <41
ο Ο üΟ
I I I •α
rt i-ü
οp- 00 </") >ο ε
ω g1
CS es os CS
£ 7 7 Öο ο ο
ο £ §
1
U") 1 "8
oo ο VI
Os es
CN e Ό
£7 7
ο Ö
CS ο ο
Ö I
S a
ε Λ
(Λ Χ3
ο ο Ο Ο CS Ο
οΟes ο
7 7 7
C S C
7 © Ö Ο es es S ο
ο <Ν
¥cη *υU
Λ
1 1 Iο ο IÖο 3 Jä
or- (C ΝS oo 00 mS J3
sO Ο ο οes ο C CS oo e
3 υ
7 7 7 7 ο CS Ο cs ο ο Ö ο 2 ο
1 ο I I I 2<1 iE
cn 00 Ο 00 ^ uο υ
Ο
CS ο CS os CN Ρ
C
>s 7 7 ο
ο Iο ö Ξ -
00 en c £
ο 00 ο δ ο
(Ν os CS
<N λδ-Ji -c
w
$7 7 ÖÖο ö u £
ο
1 Iο I
to as
σ\
Ό Τ)
ΙΛι cs
en CS CS
Τ) en es öCS ö οο
Ö ν> SO
(Ν ό
Ό so enCS ο
cc «1 r-τ> C Mcs Ο
en •α
c
u-i (Ν C Sο "M 3
cs C Sö ο ο
<N Ö •α
Ö Ο u
y
*3 t>.
/3 <υ
•α
ο
ε
«Ο
4J
Ο
ζ;
><
Table 5.2-4: Normal equations for the example in Figure 5.2-2, derived from the data
of Table 5.2-1
The system of normal equations (Table 5.2-3) of the planimetric adjustment of a block
has a very special structure which can be expressed in matrix form as:
( M (SM".1)
X] ...unknown transformation parameters
X2 ...unknown tie-point coordinates
Di,Ü2 ...diagonal matrices
Ντ, Ν ...sparse submatrices which give rise to correlation between the un-
knowns xi and X2
ni . . . vector of the absolute terms of the unknowns xi (the corresponding
vector for the unknowns X2 is the null vector)
As a result of the diagonal matrices in the normal equations (5.2-6), it is particularly

easy to eliminate the unknowns xi or x 2 before solving the equations by the Gaussian
or the Cholesky algorithm.
a) normal equations after eliminating the transformation elements xi in (5.2-6):
DLXL + NX2 = NI Ί XL = D " 1 (N! - Nx2)

T
N x i + D2X2 = 0 J N Dj~1ni - N T D f ! N x 2 + D2x2 = 0
T
(5.2-7)
( n t D - ! N - D 2 ) x2 = N T D f 1 n i (5.2-8)
This method of reducing the normal equations is particularly suitable for block
adjustments with a large number of models and a small number of tie points. This
situation is typical in analytical photogrammetry where the operator chooses a
small number of tie points.
b) normal equations after eliminating the tie-point coordinates X2 in (5.2-6):
D l X l + NX 2 = η , 1 x2 = - D " 1 (NTxi)
N t X I + D2X2 = 0 J D]X] — N D ^ ' 1 N T x i = ni
K }
( D l - N D j ' Ν τ ) X] = m (5.2-10)
This method of reducing the normal equations is particularly suitable for block
adjustments with a small number of models and a large number of tie points.
This situation is typical in digital photogrammetry where a large number of tie
points are selected.
If new points occur only in one model (single points), we can write two observation
equations (5.2-5) for each such point and compute the coordinates X , Y of the new
points from Equation (5.2-8) or (5.2-9). Since these single points contribute nothing
to the block adjustment by independent models however, it is better to eliminate them
from the initial data and then transform them in a separate process by the transforma-
tion parameters xi (Equation (5.2-7) or (5.2-10)) into the ground coordinate system.
In order to keep the computing effort and storage requirements for larger blocks within
reasonable limits, computer programs2 take advantage of the large number of zeros in
the normal matrix when solving the normal equation (5.2-8) or (5.2-10). (This proce-
dure is treated in more detail in Volume 2, Section Β 5.2.2.)
Exercise 5.2-2.
Given: The model coordinates of three models
Δ 13> 11 Δ.
(in [mm]) and the ground coordinates (in [m]) of
five control points. Required: Set up the obser-
vation equations according to Table 5.2-2 and
perform the adjustment by indirect observations.
© © ©
A 11 1I λ. Model 2
Pt. X y
3 686.90 542.66
Model 1 4 687.32 447.69
Pt. X y 5 759.46 494.88
1 697.46 572.52 6 758.68 543.16
2 698.64 463.63 7 760.64 440.05
3 808.18 571.81 Ground coordinates
4 808.50 475.26 of control points
Model 3 Pt. X Y
Pt. X y 1 1131.50 2331.50
6 609.38 578.29 2 1138.62 1142.22
7 610.80 461.60 5 3143.78 1782.21
8 692.30 579.75 8 3951.05 2332.99
9 692.89 464.34 9 3957.72 1201.05
Solution: For example ax = 10.9228 ± 0.0018, 6, = -0.0542 ± 0.0018, =

2341.04 ± 0.25 m,y 3 = 2317.74 ± 0.25 m.
5.2.2 Spatial block adjustment

In a spatial block adjustment we compute the X Y Z coordinates of points in the ground
coordinate system. As initial data we have the xyz model coordinates of points in the
stereomodels formed in relative orientation. In addition to the model coordinates of
the tie points and control points, we introduce the model coordinates of the perspec-
tive centres (Figure 5.2-3), derived from the numerical relative orientation and model
formation (see solution to Exercise 4.3-3).
2
For example: Ackermann, F., Ebner, Η. and Klein, Η.: BuL 38, pp. 218-224, 1970, and Ph.Eng. 39,
p. 967, 1973.
Figure 5.2-3: Model connection with perspective centres
The perspective centres stabilize the heights along the strip. A similar stabilization
perpendicular to the strips is not possible, however, so that, as shown in Figure 5.2-4,
chains of height control points perpendicular to the strips are required. A very good
perpendicular stabilization of heights in the block could also be achieved by side over-
laps between strips of about 60%.
Figure 5.2-4: Spatial block adjustment by independent models
The principle of spatial block adjustment can be seen in Figures 5.2-3 and 5.2-4: the
points in each model are defined in an independent, spatial coordinate system which
can be transformed into the ground coordinate system by the seven elements of absolute
orientation (Section 4.4). For the simultaneous absolute orientation of all models in the
block, we have on the one hand the model coordinates of the tie points (including the
perspective centres) and on the other the model and ground coordinates of the control
points.
The principle of spatial block adjustment can therefore be defined as follows. The
models are:
• displaced (three translations Xu, Yu, Zu),
• rotated (three rotations Ω , Φ, K ) and
• scaled (scale factor m )
so that:
• the tie points (including the perspective centres) fit together as well as possible
and
• the residual discrepancies at the control points are as small as possible.
The mathematical relation between the model and the ground coordinate system was
defined in Equations (4.2-6) and is known as a spatial similarity transformation. The
mathematical formalism of the simultaneous computation of the elements of absolute
orientation of all models in the block is, by analogy with the 2D case, sometimes called
a chained spatial similarity transformation.
The extension of spatial transformation (Equations (4.2-6)) to a chained spatial trans-

formation is in principle exactly the same as the extension of a plane transformation
(Equation (5.2-1)) to a chained plane transformation (Equations (5.2-4) and (5.2-5), as
well as Table 5.2-2). In contrast to the plane transformation, however, a spatial transfor-
mation is non-linear. W e have already seen a method of linearizing the similarity trans-
formation of a single model in the system of Equations (4.4-5). The observation equa-
tions (4.4-6) for an adjustment by indirect observations have already been formulated.
The observation equations (4.4-6) correspond to the observation equations (5.2-4) for
a chained plane similarity transformation: the XYZ coordinates arising in the equa-
tions (4.4-6) correspond on the one hand to the (known) control point coordinates. On
the other hand, for a new point the XYZ coordinates in equations (4.4-6) should be
interpreted as (unknown) new point coordinates. Since such new points appear in sev-
eral models they have a connecting function within the block of stereomodels. Just as
the system of observation equations in Table 5.2-2 can be found using the observation
equations (5.2-4) and (5.2-5), so the system of observation equations for the spatial
block adjustment with independent models can be found from the correspondingly
adapted observation equation (4.4-6). Due to space restrictions a detailed presentation
will not be made.
In the system of observation equations for the spatial block adjustment with indepen-
dent models, the unknowns are small additions to the approximate starting values. (In
contrast, no linearization and no starting values were necessary for the system of equa-
tions for planimetric block adjustment.) Automatic generation of approximate starting
values to initialize the linearization is presented in Volume 2, Section Β 5.1.
5.2.3 Planimetrie and height accuracy in block adjustment

by independent models
The reduction in the number of field-surveyed control points—the objective of aero-
triangulation—leads in general to a reduction in accuracy compared with the absolute
orientation of every model on four field-surveyed points in the model corners. For
the application of aerotriangulation it is therefore of prime importance to be able to
estimate this reduction in accuracy.
In a spatial block adjustment, the planimetric accuracy is not affected by the accuracy
of the model heights or by the layout of the height control points if the ground is
relatively flat. Similarly, the height accuracy is independent of the accuracy of the
model coordinates x, y and the layout of the planimetric control points. The planimetric
and height accuracies are therefore treated separately. The results for planimetry are
valid for planimetric as well as spatial block adjustment.
5.2.3.1 Planimetric accuracy
Since the X Y coordinates of the tie points are computed from a least squares adjust-
ment by indirect observations (Appendix 4.1-1), we can derive their accuracy—more
correctly, the weight coefficients (cofactors) qxx and ςγγ—by inverting the normal
equation matrix (Table 5.2-3). As a result of the symmetrical structure in X and Y,
the weight coefficients qxx and qyy are identical; they are named qpp below. We
then have the accuracy σβ,ρ of the coordinates X and Y of a tie point after the block
adjustment:
— \fqppoο = y/qppOM,p (5.2-11)
σο is the standard error of unit weight of the adjustment, i.e. the accuracy σχ or συ of
a model coordinate χ or y, expressed in the ground coordinate system (see the com-
ment following Equation (5.2-5)). The quantity can therefore be regarded as a
factor which, multiplied by the accuracy σΜ,ρ of the X Y coordinates in the individual
models, gives the planimetric accuracy σβ,ρ of the block.
Figure 5.2-5 3 shows these factors sßTpp for various sizes of blocks, each containing a
control point in the four corners of the block. The quantity a m a x denotes the maxi-
mum planimetric error in the block, while CTmean denotes the root mean square of the
planimetric error of all tie points in the block, i.e.:
Numerical Example. Given: Aerial photographs at a scale of 1 : 6000. Each model

has 4 tie points which are as accurately defined as targeted points. The block is com-
posed of 32 models (case dl in Figure 5.2-5) and is adjusted by independent models.
Required: a m a x of the adjusted coordinates and for the entire block.
3
These and the further diagrams of planimetric accuracy were published by Ackermann (BuL 35,
pp. 114-122, 1967).
al (1 χ 2) = 1.31 σ„
= 1.44 σ ο
ο » , = 3.02 σ
σ„„ = 4.08 σ „
Figure 5.2-5: Planimetrie accuracy with four points in the block corners, for square
blocks ns χ nm (ns... number of strips, nm... number of models in a strip)
Accuracy in a single model (Equation (4.6-1)): gm,ρ = 0.0006 χ 6000

- ±3.6cm
Block accuracy (maximum error): ^B,p,max - ±2.28 χ 3.6
±8.2cm
Block accuracy (r.m.s.): &B,p,mean = ± 1 . 8 5 x 3 . 6
= ±6.7 cm
We see from Figure 5.2-5 that:
• the accuracy falls significantly as the size of the block increases and
• the maximum error occurs in the middle of the block edges.
It is therefore obvious that if we wish to increase the overall accuracy we should pro-
Figure 5.2-6: Planimetrie accuracy with a dense pattern of control points along the
block edges, for square blocks ns χ nm
vide a dense pattern of control points along the edges of the block. The success of such
a strategy is shown in Figure 5.2-6.
The accuracy
• is almost independent of the size of the block and
• is close to the accuracy in a single model.
The question of further improvement in accuracy by introducing control points in the

centre of the block is answered in Figure 5.2-7:
Comparison with Figure 5.2-6 shows that:
• control points inside the block bring no significant improvement in accuracy.
Some generalization of these accuracy results for square blocks has been published
Figure 5.2-7: Planimetrie accuracy with a dense pattern of control points along the
edges of the block and a control point in the centre of the block, for square blocks
n s Χ Τίγγι
by Ackermann 4 , Ebner 5 and Meissl6 for the patterns of control points shown in Fig-
ure 5.2-8 ( P I corresponds to Figure 5.2-6, i.e. the number i of bridged models corre-
sponds to two baselengths along the block edges; PA corresponds to Figure 5.2-5) and
summarized in Equations (5.2-12), for which six tie points per model, as opposed to
four in the model corners, were assumed.
-PI : σΒ,Ρ,mean « (0.70 + 0.29 log n s ) a M , p

P2 : <TΒ,P,mean ~ (0-83 + 0.02η β )σ Μ ,ρ
: <J Β,Ρ,mean « (0-83 + 0.05η8)σΜ,ρ
ΡΑ : & Β, Ρ,mean « (0.47 + 0.25η 3)σΜ,ρ
ns ... number of strips in the block
Numerical Example (to verify Equations (5.2-12) for a square block with eight strips).
Case 1 (block edges with dense control = g3 in Figure 5.2-6):
Equation 1 of (5.2-12): as,p,mean ~ (0.70 + 0.291og8)aM,p ~ 0.96σ Μ ,ρ
Figure 5.2-6: σΒ,ρ^η = 1-03σ Μ ,ρ
Case 2 (four control points, in the corners of the block = gl in Figure 5.2-5):
Equation 4 of (5.2-12): σΒ,ρ<ΠΚ3η ~ (0.47 + 0.25 χ 8)σ Μ ,ρ « 2.47σ Μ ,ρ
Figure 5.2-5: σΒ,ρân = 3.02σ Μ ,ρ
4
B U L 36, pp. 3 - 1 5 , 1968.
5
BuL 40, pp. 214-221, 1972.
6
ÖZfV, pp. 61-65, 1972.
•••••••Dm Β—Β—Ξ—Β—g
Π [)
11
[1 P1
K l
11
[] Ρ2
11 ,h 11
16 PCP
[] l = 2b π
Π []
Θ—Β—Β—Β—Β
• Β α
Ρ3
Ο [1
8 PCP
η Β α
Figure 5.2-8: Pattern of planimetric control points for the Equations (5.2-12). Note
that the block size is variable; for PI the number of control points increases with the
size of the block, for P2-P4 the number is constant ( P C P = planimetric control
point).
The difference between 0.96 and 1.03 and between 2.47 and 3.02 can be tolerated
for rough estimates of block accuracy such as concern us here. The difference arises
partly from the fact that Equations (5.2-12) are based upon the use of six tie points in
each model while Figures 5.2-5 to 5.2-7 are based upon the use of four tie points per
stereomodel.
Numerical Example (for project planning). Model control points are to be established
by photogrammetry in an area of 10 km χ 10 km, with a coordinate accuracy of ± 10 cm.
These control points will be natural points for which we assume an uncertainty of
definition of ± 5 cm. Eight field-surveyed control points exist in the block, four in
the corners and four in the middle of the edges (P3 in Figure 5.2-8). What scale of
photography must be adopted?
Trial 1, photo scale 1 : 15000:
Distance between strips for 25%

A = 0.23 χ 15000(1 - 25/100) = 2590 m
side overlap (Equation (3.7-1)):
Number of strips (Equation (3.7-1)): ns = [10000/2590 + 1J = 4
Accuracy in single model
(Equation (4.6-3)): cm, ρ = V92 + 52 = 10.3 cm
Block accuracy (Equation (5.2-12)): σΒ,ρ = (0.83 + 0.05 χ 4)10.3 = ± 1 0 . 6 c m
Trial 2, photo scale 1 : 12000:

A = 0.23 χ 12000(1 - 25/100) = 2070 m

ns = [10000/2070 + 1J = 5
σΜ,ρ = V7.2 2 + 5 2 = ± 8 . 8 c m
σΒ,ρ = (0.83 + 0.05 χ 5)8.8 = ±9.5 cm
A photo scale of 1 : 12000 is therefore suitable for the task.
Exercise 5.2-3. Repeat this example of project planning with the assumption that 16
field-surveyed points are available (P2 in Figure 5.2-8). (Solution: Four strips at a
photo scale of 1 : 15000 are sufficient.)
The rules for accuracy shown above are also valid, to a rough approximation, for rec-
tangular blocks, but not for the extreme case of a single strip. Accuracy in a strip is
therefore discussed separately in Section 5.2.3.4. The near-equality between block ac-
curacy and single-model accuracy for blocks with dense control around the edges—the
most important case in practice—also holds good if the edges of the block are irregular.
Up to this point, the accuracies quoted have all referred to the tie points in the corners
of the models. Single points, i.e. points lying in one model only, will be less accurate.
In the case d3 of Figure 5.2-6, the accuracy of single points is 1.33σΜ,ρ 1 , a decrease
of about 33%, a value which can also be applied as a guide in blocks of other sizes and
with other patterns of control. In the same publication, Ebner also studied the effect on
accuracy of single points of using significantly more than four or six tie points in each
model. For the same case d3 of Figure 5.2-6, 60 (!) tie points per model brought an
improvement in accuracy of only about 25%. The number of tie points thus has little
effect on the accuracy of the coordinates of single points after a block adjustment.
In the interest of blunder detection and blunder location, however, the statements made
above need to be revised: one should use at least eight tie points in each model, two in
each corner, and groups of at least two control points rather than single control points.
(This aspect is treated in depth in Volume 2, Section Β 7.2.2.4.)
Exercise 5.2-4. A cadastral survey is to be performed by photogrammetry in an area

of 5 km χ 5 km. The maximum difference between lengths derived from photogram-
metrically determined coordinates and lengths measured on the ground (check lengths)
is to be As = 10 cm. Assume signalized points and compute the smallest photo scale
together with the pattern of field-surveyed control points. Hint: A maximum error
in length of As = 10 cm corresponds to a root mean-square planimetric accuracy of
σρ = 10/(3\/2) = ±2.4 cm (division by 3: relation between maximum and root mean-
square error; division by y/2: Equation (4.6-4)). (Solution: Arrangement of control
points as PI in Figure 5.2-8, 1 : m e = 1 : 4150, 8 strips. Alternatively, arrangement of
control points as P2 in Figure 5.2-8, 1 : m j = 1 : 3960, 9 strips. Another alternative is
the arrangement of control points as P3 in Figure 5.2-8, 1 : rriB — 1 : 2790, 12 strips.)
7
Ebner, Η.: Na. Ka. Verm. I, Heft 53, pp. 51-71, 1971.
5.2.3.2 Height accuracy
The accuracy of heights after a block adjustment can be derived from an inversion of the
normal equation matrix, in the same way as for planimetric accuracy. The analogous
relation to Equation (5.2-11) is then:
&B,Z — y/θζζσο = \/qzz(TM,z (5.2-13)
The height accuracy is primarily dependent on the number i of models between two
chains of height control points (perpendicular to the strips). It is also good practice to
improve the height accuracy along the upper and lower edges of the block by introduc-
ing height control points on these edges at intervals of i/2 models. The ideal pattern of
height control is then as shown in Figure 5.2-9.
-— i/2 models —
ο-
ι models
Figure 5.2-9: Ideal pattern of height control points
Figure 5.2-10 8 shows the relation between the accuracy of heights of corner points in
the models and the number of models i bridged between control chains. It shows the
root mean-square value for the entire block (amc-dn) and the maximum value a m a x in the
most unfavourable position in the block. The corresponding equations assuming six tie
points are:
<?B,Z,mean (0.34 + 0.22ί)σ Μ ,ζ
σ
(5.2-14)
Β,Ζ, max (0.27 + 0 . 3 Η ) σ Μ , ζ
8
Ebner, Η.: BuL 40, pp. 214-221, 1972 (Forster, B.C.: UNISURV G28, pp. 84-93, 1978, demon-
strates a further generalization).
— (mean)
i (Number of models bridged)

5 10 15 20
Figure 5.2-10: Height accuracy of a block adjustment (σ Β ) compared with the accu-
racy in a single model (ctm)
Numerical Example. Given: Aerial photographs at a scale of 1 : 6 0 0 0 ; 15/23 camera;

tie points to the accuracy of signalized points; adjustment of a 72-model block by
independent models, with the control pattern shown below:
Λ-
Δ-
A- Δ Full control point
ο Height control point
Required: a max of the adjusted heights and a mean for the height accuracy of the com-
plete block.
Accuracy in a single model
&M,Z = ± 5 . 4 cm
(Equation (4.6-1)):
Block accuracy (maximum
error): σΒ,ζ,max = ± ( 0 . 2 7 + 0.31 χ 6 ) 5 . 4 = ± 1 1 . 5 c m
Block accuracy
σΒ,ζ,mean = ± ( 0 . 3 4 + 0 . 2 2 χ 6 ) 5 . 4 = ± 9 . 0 c m
(representative value):
Exercise 5.2-5. What is the change in accuracy if four chains of height control points
are used instead of three? (Solution: σβ,ζ,πηχ = ±8.2 cm, σΒ,ζ,mean = ±6.6 cm.)
We see that, when compared with planimetric accuracy, particularly with dense control
along the edges of a block, height accuracy is significantly less favourable. An interval
of about three models between chains of height control is the maximum if we wish to
introduce no significant loss of accuracy compared with that of a single model—but

we then affect adversely the economy of the entire process. As a compromise between
accuracy and economy, an interval of at least four models between chains of control
is usual. The introduction of GPS data 9 (Section 3.7.3.2) into the block adjustment
frees the adjustment from this very restrictive pattern of height control (Section 5.4
and Section Β 5.3.5.2.2, Volume 2).
In the interest of blunder detection and blunder location one should—as for planimet-
ric adjustment (see the end of Section 5.2.3.1)—introduce at least eight tie points per
model, two in each corner, and provide control points in groups rather than as single
points (see Volume 2, Section Β 7.2.2.4). In high mountain regions care should be
taken that height control points do not only lie in the valleys or on mountain peaks;
they should cover the full height range of the area.
5.2.3.3 Empirical planimetric and height accuracy
The theoretical accuracies of block triangulation by independent models shown above

are based on the assumption of random errors in the individual models. The effects of
varying accuracy within a model and the undoubted existence of correlations resulting
from systematic errors are ignored.
Nevertheless, extensive empirical tests of accuracy made on field-surveyed check

points 10 have confirmed the essential soundness of the theoretical results, so that we
can use Equations (5.2-12) and (5.2-14) as rough approximations for project planning.
5.2.3.4 Planimetric and height accuracy of strip triangulation
The accuracy of points in a strip adjusted by independent models depends primarily on

the number i of models bridged between control points. As before, we are interested in
two values—a mean accuracy as imea n which is representative for the strip as a whole
and a maximum value as,max of the error occurring between control points.
Since we must here also differentiate between signalized, natural and artificial points,
it is convenient to relate the accuracy of strip triangulation, a s , to the error, gm, of
single models introduced in Section 4.6.
The results of theoretical11 and empirical 12 studies of accuracy of strip triangulation,

which, in part, produce greatly different values, are sketched in Figure 5.2-11. The
'Hein, G.W., van der Vegt, H.J.W., Andersen, O., Colomina, I.: Schriftenreihe des Inst, für Pho-
togrammetrie, Uni. Stuttgart, Heft 13, pp. 261-325, 1989. Lucas, J.R., Mader, G.L.: Journal of Surveying
Engineering 115/1, pp. 78-92, 1989. Friess, P.: BuL 58, pp. 136-143, 1990.
10
OEEPE Publ. No. 8, 1973.
"Ackermann, F.: DGK, C 87, 1965.
12
Stark, Ε.: OEEPE, No. 8, pp. 49-82, 1973; BuL 45, pp. 183-190, 1977.
single models (ctm)
extreme simplification represented in this diagram has the practical advantage of simple
interpretation; note, however, that differences of up to 50% from the values shown in
Figure 5.2-11 can occur in individual cases.
Numerical Example (of project planning). Signalized points along a strip must be
measured with a planimetric accuracy, σρ, of ± 3 cm and a height accuracy, σζ, of
± 5 cm. Because of the high cost of control, as few field-surveyed control points as
possible are to be used. A 15/23 camera is to be used. The field-surveyed points are at
the ends of the strip and in the middle.
Trial 1. Photo scale 1 : 4500, with the following pattern of control points.
Single-model accuracy (Equation (4.6-1)): σΜ,ρ = 0.0006 χ 4500

= ±2.7 cm
σ Μ Ζ = 0.00006 χ 15 χ 4500
= ±4.0 cm
Strip accuracy (Figure 5.2-11, i = 4): as,ρ — ±2.7 χ 1.2
= ±3.2 cm
= ±4.0 χ 1.5
= ±6.0 cm
Section 5.3 Bundle block adjustment 269
Trial 2. Photo scale 1 : 3000, with the following pattern of control points.
Single-model accuracy: &M,P = 1-8 cm

c m , ζ = ±2.7 cm
Strip accuracy (Figure 5.2-11, i = 6): as,ρ = ±1.8 χ 1.4
= ±2.5 cm
= ±2.7 χ 1.9
= ±5.1 cm
Strip accuracy in the models central between the control points:
a s , Ρ = ±1-8 χ 1.7
= ±3. l c m
a s , ζ = ±2.7 χ 2.2
= ±5.9 cm
The requirements for the strip as a whole are met, while the errors in the models be-
tween the control points are only just higher than allowed.
This numerical example shows that the reduction in accuracy caused by increasing
the number of bridged models is significantly less than the improvement in accuracy
resulting from the increase in photo scale.
5.3 Bundle block adjustment

In a bundle adjustment of a strip or block of photographs, with at least 60% forward
overlap and 20% side overlap, we compute directly the relations between image coordi-
nates and object coordinates, without introducing model coordinates as an intermediate
step. Thus, the photograph is the elementary unit in a bundle adjustment.
5.3.1 Basic principle

Figure 5.3-1 shows the basic principle of a bundle block adjustment.
The image coordinates and the associated perspective centre of a photograph define a
spatial bundle of rays. The elements of exterior orientation of all bundles in a block
are computed simultaneously for all photographs. The initial data consist of the image
coordinates of the tie points (points existing in more than one photograph) together
with the image coordinates and object coordinates of the control points.
The adjustment principle can therefore be defined as follows: the bundles of rays are:
• displaced (three translations XQ, YQ, ZQ) and
• rotated (three rotations ω, ψ, κ)

Figure 5.3-1: Principle of a bundle block adjustment
so that the rays
• intersect corresponding rays as nearly as possible at the tie points and
• pass through the control points as nearly as possible.
The mathematical relationship between the image coordinates and the global ground
or object coordinate system can be found in Appendix 2.1-2. The differential quotients
for the corresponding collinearity equations can be found in Appendix 2.1-3.
5.3.2 Observation and normal equations for a block of photographs

We can now use the differential quotients in Appendix 2.1-3 to write down the lin-
earized observation equations of a least squares adjustment by indirect observations
(Appendix 4.1-1), for (new) points P, whose image coordinates have been measured
in the photograph with the index j. Each measured image point yields two observation
equations.
0 ο
Vîj = dY0j +
(5.3-1)
- Φ
while the expression for v v i j has a similar form. The unknowns are the six elements of
exterior orientation of the photograph with the index j and the three ground coordinates
of point Pi. If a ray passes through a fixed point, the terms dXiy dYi and dZt disappear.
The differential quotients ()° are calculated from approximate values of the unknowns
by the relations in Appendix 2.1-3. and 77? are image coordinates computed from
Equations (2.1-19) with the help of the approximations to the unknowns, ξ^ and ή13 are
the measured image coordinates. The necessary approximations to the unknowns can
be derived in various ways. For example, for near-vertical photographs ω ο = φ ° = 0;
κ° is known from the flight plan. A block adjustment by independent models (Sec-
tion 5.2) yields the coordinates X°Y°Z° of the perspective centres and the coordinates
X Y Z of the new points (this problem is taken up again in Volume 2, Section Β 5.2.1).
A schematic example with four photographs will help to clarify the procedure. The
balance of observations and unknowns is:
1A 2 A A1 A2
@
% 4. .3
*->
Observed image coordinates (2 χ 6 χ 4) = 48
Unknowns:
is. 6. .5 .6£
3 x 4 = 12 perspective-centre coordinates
3 x 4 = 12 rotations
3. 4. 3 x 4 = 12 new-point coordinates
(N) => 36
0 5. 6. Redundancy: 48 - 36 = 12
.c
CL·7 A 8
A
Observation equations
In place of the differential quotients ()° in observation equations (5.3-1), the following
table uses variables a and b which are introduced in Appendix 2.1-3.
Xoi Yoi Zoi ωι ψ\ Xq2 Y02 Ζ02 ω2 ¥>2 «2

«ξΐι 4 a° <
«7)116° bl
Pt. 2 similarly
« ξ 3 1 a! <
«7)31 6° b°4 b°5 69
Pts. 4,5,6 similarly
«ξ12 α» a° a° a° a° «7°
«7)12 6° 6° 6° 6°
Pt. 2 similarly
« ξ 3 2 a« «3° «4 ^
«7)32 68 6§ 6§ 60
Pts. 4,5,6 similarly
XQ4 Y04 Z04 ΟΛ( ψ4 -X3

«4 Y3 ^3 X(, Υβ Ζβ
(61
(mi-'/?,)
n° n° n°
a
8 9 " l O (61-&)
60 6° 6j0 (»Öl - » & )
(^12 - r/?2)
(7° 8 u
9 u
1 0 (62-^2)
6°
°8
6° h°
9 ° 1 0
(V32 ~ V32)
In matrix notation:
ν = Ax — 1 (5.3-2)
Normal equations:
A Ax = A 1 (5.3-3)
If we replace A T A by Ν and A T 1 by η the normal equations become:
Nx = η (5.3-4)
5.3.3 Solution of the normal equations

The normal equations (5.3-4) of the schematic example given in Section 5.3.2 have the
structure:
Photo! Photo2 Photo3 Photo4 Pt3 Pt4 Pt5 Pt6
Nil
N22
In matrix notation:
N„ N12\ / X , \ = /m
(5.3-5)
NJ2 N22 Y V X 2/ VN2
The upper left matrix N n is a hyperdiagonal matrix with submatrices each of 6 χ 6
elements. The lower right matrix N 2 2 is similarly a hyperdiagonal matrix with subma-
trices of 3 χ 3 elements. The inversion of the two hyperdiagonal matrices is particularly
simple: each submatrix can be inverted independently.
The system of normal equations reduced by the new-point coordinates x 2 can thus be
formed with no great computing effort. By analogy with Equations (5.2-9) and (5.2-10)
we have from Equation (5.3-5):
( N n - N, 2 N 2 - 2 1 Ni 2 )xi = ni - N i 2 N ^ n 2 (5.3-6)
An adjustment yields corrections to the approximate values of the elements of exterior

orientation of each photograph and to the approximate coordinates of the new points.
If the approximations are very poor, the corrected values must be treated as new ap-
proximations. The adjustment should be repeated until there is no further significant
change in the unknowns. (Further details on the solution of the normal equations will
be found in Volume 2, Section Β 4.4.)
Further treatment of bundle block adjustment will be found in: Wong, K.W., Elph-
ingstone, G.M.: Ph.Eng. 39, pp. 267-274, 1973. ASPRS-Manual of Photogramme-
try, 5 th ed., Falls Church, 2004. Triggs, B„ McLauchlan, P., Hartley, R„ Fitzgibbon,
Α.: Bundle Adjustment—A Modern Synthesis. Vision Algorithms'99, LNCS 1883,
pp.298-372, 2000.
5.3.4 Unknowns of interior orientation and additional parameters

If amateur cameras are used instead of photogrammetric cameras, the elements of in-
terior orientation can also be determined within the bundle block adjustment. For this
purpose, the elements of interior orientation are introduced as unknowns in the obser-
vation equations (5.3-1). The solution of the normal equations (5.3-3) then yields the
elements of interior orientation of the particular camera used. The larger number of
unknowns in such a bundle block adjustment obviously requires a larger number of
control points and tie points.
It is typical of photographs from amateur cameras that the theoretical central projec-
tion is significantly deformed by lens and film distortion. These influences can be taken
into account in a bundle block adjustment by introducing correction polynomials in the
observation equations, whose coefficients are determined in the adjustment. Such an
adjustment is called a bundle block adjustment with additional parameters or with self-
calibration. The technique is not only used for amateur-camera photographs, but more
and more for photographs taken with proper photogrammetric cameras. Additional
parameters provide an extremely powerful method of compensating for systematic er-
rors. It is of no importance whether these systematic errors occur as a result of lens
or film distortions or as a result of anomalies of refraction etc. This concept is further
discussed in Volume 2, Sections Β 3.5.6 and Β 5.2.4.
Further reading: Jacobsen, K.: Phia, pp. 219-235, 1982. Kilpelä, Ε, Helkilä, J, Inkelä,
Κ.: Phia, pp. 1-12, 1982.
5.3.5 Accuracy, advantages and disadvantages of bundle block

adjustment
Although the mathematical model of a bundle block adjustment differs significantly
from that of block adjustment by independent models—central projection with the im-
age coordinates as observations on the one hand and spatial similarity transformation
with the model coordinates as observations on the other hand—the rules of accuracy for
independent-model adjustment (Section 5.2.3) can be applied, more or less, to bundle
block adjustments.
The "single-model accuracy", which is required by the rules in Section 5.2.3, however,
can be based on the following accuracy guides. These apply to targeted and other corre-
spondingly accurate points in a regular block with 60% forward overlap and 20% side
overlap, processed by a bundle block adjustment with additional parameters which re-
duce systematic errors. The values are 50% better than the guide values in Section 4.6.
In the case of analogue images, they are:
Planimetry: ^ y ^ g ) = ± 3 μιη in the photograph χ image scale m e

Height: crZUarg) — ±0.03%c of camera distance (NA - WA) (5.3-7)
= ±0.04%o of camera distance (SWA)
Exercise 5.3-1. Repeat the numerical examples of Section 5.2.3 under the assump-
tion that the block triangulation has been performed by the bundle method with addi-
tional parameters rather than by independent models. (Solution: Numerical example
according to Figure 5.2-5: ομ,ρ = ± 1 . 8 cm, σΒ,ρ,πχ&η = ± 3 . 3 cm; numerical example
according to Figure 5.2-6: gm,z = ± 2 . 7 cm, crß,z,mean = ±4-5 cm.)
Extensive details concerning the accuracy of the highly developed bundle block ad-
justment can be found, amongst other, in the following literature: Schwidefsky/Acker-
mann: Photogrammetrie. Teubner-Verlag, Stuttgart, 1976. Brown, D.C.: Ph.Eng. 43,
p. 447, 1977.
At the end of this section we list the advantages and disadvantages of bundle block
adjustment relative to independent-model adjustment:
Disadvantages:
• non-linear problem, for which approximations can only be established after

lengthy procedures
• computer-intensive methods
• always a spatial problem, so that separate adjustments in planimetry and height

are not possible
Advantages:
• most accurate method of aerotriangulation (direct relation between image and

ground coordinates without the intermediate step of model formation)
• simple possibility of extending the technique to compensate for systematic errors

(see Volume 2, Sections Β 3.5.6 and Β 5.2.4)
• simple possibility of incorporating observed elements of exterior orientation into

the adjustment (see Section 5.4)
• possibility of using unconventional camera dispositions and amateur-camera

photographs such as are often necessary in close range photogrammetry (Sec-
tion 5.3.4)
• possibility of deriving the elements of exterior orientation to be set in particular

analogue or analytical plotters. The photogrammetrie measurement of control
points for subsequent stereoprocessing is then unnecessary.
5.4 GPS- and IMU-assisted aerotriangulation

Aerotriangulation, as discussed in the previous paragraphs, is intended to determine
the exterior orientation elements of the metric images. With GPS and IMU, which are
increasingly used (Section 3.7.3.2), the exterior orientation elements can be directly
determined and the images then processed using the methods of direct georeferencing
(Section 4.1.1). In principle, GPS and IMU make aerotriangulation redundant. In prac-
tice, however, the following factors argue against direct georeferencing which employs
only GPS and IMU information and makes no use of control points:
• GPS positioning "on the fly" is relatively demanding and cannot, at present,
reach the centimetre accuracy level which is the objective of precision aerial
photogrammetry.
• unavoidable cycle slips and/or multipath effects can disturb continuous GPS po-
sitioning.
• the combined calibration of the three units, GPS, IMU and imaging sensor, is a
very complex exercise in which there are often considerable shortcomings (Sec-
tion 3.7.3.2).
• GPS values, which are determined in a global coordinate system, must be cor-
rected for various influences, such as the geoid, in order to correspond to the
current ground (national) coordinate system in which the photogrammetric re-
sults are expected. In other words, GPS/IMU data often have datum problems
with respect to the ground coordinate system.
Instead of direct georeferencing by GPS and IMU, which is known as direct sensor
orientation, GPS and IMU information should be used only to support aerotriangula-
tion. The combination of GPS and IMU data with control points for orienting image
recording systems is known as integrated sensor orientation.
The following discussion concentrates on GPS-assisted aerotriangulation. The per-
spective centres of the metric images can, as suggested in Section 3.7.3.2 (note in par-
ticular Figure 3.7-12), be determined in the global GPS coordinate system with the aid
of GPS and IMU recording. Due to the previously indicated datum problems, amongst
others, the GPS coordinates Xoj, Yoj and Zoj of the perspective centres cannot be intro-
duced as known parameters (i.e. constants) in the adjustment equations (5.3-1). (They
would be better treated as "observed unknowns", see Section Β 3.5.7, Volume 2.) In-
stead, all perspective centres fixed by GPS should be grouped into a single model in a
three-dimensional coordinate system, the GPS model, and processed in a hybrid block
adjustment. Here the bundles of rays from each metric image (bundle block adjust-
ment, Section 5.3) and the GPS model just mentioned (spatial block adjustment with
independent models, Section 5.2.2) are processed together. The ground coordinates of
the perspective centres are regarded as unknowns in this hybrid block adjustment; the
unknown perspective centres connect the ray bundles of the individual metric images
and the GPS model. Each perspective centre with its three GPS coordinates provides,
Section 5.5 Georeferencing of measurements made with a 3-line camera 277
in addition to the adjustment equations of the bundle block adjustment, three further
adjustment equations. Within the hybrid block adjustment the GPS model is subject
to three translations, three rotations and a scale change. These seven unknowns prin-
cipally eliminate the above mentioned datum problem between GPS data and ground
coordinates.
The other errors in the GPS data, mentioned above, are largely eliminated by a division
into several GPS models. In particular, a GPS model for every image strip is very
suitable.
The GPS models support the perspective centres inside the photogrammetric blocks.
GPS-assisted aerotriangulation therefore avoids the need for height control point chains
(Section 5.2.3.2). Further details about GPS-assisted aerotriangulation and its accuracy
can be found in Volume 2, Section Β 5.3.5.
Where IMU data are available the process is similar. Thus, the IMU orientation angles
Wj, ψ] and Kj should not be introduced as known parameters (constants) in the adjust-
ment equations (5.3-1). Instead, these orientation angles should be left as unknowns
in every metric image and an "orientation angle model", the IMU model, introduced
into a hybrid block adjustment. This IMU model should have some unknown parame-
ters, by analogy with the GPS model. A division into independent IMU models for
individual strips is, in many cases, to be recommended.
5.5 Georeferencing of measurements made with a

3-line camera
Aerial survey flights using 3-line cameras are only made in combination with GPS and
IMU (Section 3.7.2.3). GPS determines the flight path, and IMU the angular orienta-
tion, of the 3-line camera. With this GPS/IMU information, the object coordinates of
new points identified in the three strips can be simply determined (Section 4.1.2). This
direct solution, or direct sensor orientation, will be expanded on in the following sec-
tions. For instructive reasons, the GPS/IMU information will initially be disregarded.
In a 3-line camera (Figure 3.3-3, left), each line corresponds to a central projection
with its own individual elements of exterior orientation. The six elements of exterior
orientation change dynamically and are a function of time or line index. These six
orientation functions are illustrated in Figure 5.5-1. Normally they are described by
cubic polynomials between nodal points JVj (cubic spline interpolation). The separation
of the nodes should be chosen such that the dynamically varying orientation elements
are approximated to a sufficient accuracy by the successive cubic polynomials.
The abscissae ti of the nodes Ni,which are often regularly spaced, are known; at each
node the ordinates of the six orientation functions Xo(ti), Yo(i?;), Zo(tt), ω(ίί), φ(ίι),
κ ( ί ί ) η are required. Interpolation is implemented between the nodes TV, by means of
cubic polynomials whose coefficients can be determined from the nodes. (Details on
13
Figure 5.5-1 implies that the ω φ κ rotations of the sensor take place about the three coordinate axes of
the aircraft. In Figure 3.7-13 this coordinate system has been designated as the body coordinate system.
Figure 5.5-1: Flight path of a line sensor and 6 orientation functions
cubic spline interpolation can be found in Lancaster, P., Salkauskas, K.: Curve and
Surface Fitting. Academic Press Ltd., 1986.)
The obvious question now arises as to how a point Ρ with coordinates xt, = (xb,yb, zt,)T in the XbVbZb
body coordinate system can be transformed into the X Y Z object coordinate system. The following
rotational axes are available for this exercise:
• primary rotation ω (roll) about the Xb axis until the yb axis is parallel to the X Y plane.
• secondary rotation ψ (pitch) about the yb axis (already rotated by ω) until the χι, axis also lies
parallel to the X Y plane.
• tertiary rotation κ (yaw/heading) about the Zb axis (already rotated by ω and ψ) until the Xb and
yb axes are parallel to the X and Y coordinate axes.
Using the rotational and axial directions introduced in Figure 2.1-5, the following result is obtained by
analogy with the presentation of Appendix 2.1-1:
X = RL^kXÖ (5.5-1)
The rotation matrix Έ Ι ω φ κ introduced in Section 2.1.2 needs therefore only to be transposed. (This result
is not surprising given the knowledge from Section 2.1.1 about reversal of transformations involving
rotation matrices.) From the numerical form of a given rotation matrix, the roll, pitch and yaw angles
of an aircraft (Section 3.7.1) can be determined with the help of the transposed algebraic form of the
rotation matrix (2.1-13) in the usual way.
Exercise 5.5-1. The rotation angles of a body coordinate system are to be determined from the matrix
R.2 given in the numerical example of Section 2.1.2: Solution, making use of the formulations (2.1-1-8)
in Appendix 2.1-1: ω (roll angle) = - 1 . 0 8 8 6 gon ( - 5 8 ' 4 7 " ) , ψ (pitch angle) = 0.1499 gon (8'6"), κ
(yaw angle/heading) = 101.3224gon (9Γ11'25").
Further reading: Baumker, Μ., Heimes, F.: OEEPE-Publication No. 43, pp. 197-212, 2002.
Section 5.5 Georeferencing of measurements made with a 3-line camera 279
Before discussing the determination of the ordinates of the six orientation functions
at all nodes, a particular feature of the 3-line camera (Figure 3.3-3, right) should be
emphasized. If the three strips in the triplet are aligned along the same time index, then
the lines arranged in vertical rows in Figure 5.5-2 have the same orientation elements;
in other words: for all strips in a triplet a single set of six orientation functions Xo(U),
Yo(U), Z0(ti), co(ti), ip(U), k(U) is required.
Forward
_JJt - r 9
Nadir ? wY [Hy
1
Figure 5.5-2: Image triplet from a 3-line camera showing the common line index
The basic equations for georeferencing images from a 3-line camera are the observation
equations (5.3-1). However, the six exterior orientation elements Xqj, Yoj, Zq3, Uj, ψj,
Kj, which would be inserted here for each line triplet, are replaced by the six ordinates
Xo(ti), Yo(U), Z0(ti), oj(ti), <p(ti), n(ti) at all nodes Nt. Since the nodes Ni are
considerably more widely spaced than the original lines, the number of unknowns in
the adjustment is significantly reduced. (Details about this substitution can be found in
Forkert, G.: Geow. Mitt, der TU Wien, vol. 41, 1994.)
Every control point identified in a triplet of strips provides 2 x 3 = 6 adjustment equa-

tions for determining the nodes of the six orientation functions. The ξ coordinate in
the forward-looking strip corresponds to line spacing a while the negative ξ coordinate
in the backward-looking strip corresponds to line separation b (Figure 4.1-1). (The η
coordinates of a point identified in the three strips are only slightly different.) Every
tie point determined in a triplet of strips also provides 2 x 3 = 6 adjustment equations
for the determination of the six orientation functions. Also with the tie points the ξ
coordinates are known from the a and b values. In adjustment equations based on tie
points, the (unknown) ground coordinates XiYiZi of a tie point also appear (see Equa-
tion (5.3-1)). A tie point in the overlapping region between two neighbouring triplets
of strips provides 2 x 2 x 3 = 12 adjustment equations to determine the nodes of the 6
orientation functions in each strip; the three (unknown) ground coordinates Xi,Yi, and
Zi for such a tie point again appear in the system of equations.
Aerotriangulation, i.e. indirect georeferencing, can also be implemented with these ad-
justment equations. This would require many control and tie elements. Control and
tie elements include control lines and tie lines in addition to control and tie points, as
indicated in Figure 5.5-1, left (the relevant mathematical analysis can be found in Sec-
tion Β 4.3, Volume 2). The high number of control and tie elements can, however, be
considerably reduced if the GPS/IMU information is included in the sensor orientation.
Such a solution is very similar to GPS/IMU-assisted aerotriangulation (Section 5.4).
For this purpose, the six orientation functions are determined for all strips from the
acquired GPS/IMU data. The abscissae of the nodes Ni are taken from the indirect
georeferencing explained above. The ordinates of the nodes Ni can be determined
by interpolation between the GPS/IMU observations at the nodes Ni. Restricting fur-
ther discussion to the three orientation functions Xo(t), Yo(t) and Zo(t), the result is
a single GPS model for the complete area of interest. This GPS model contains three-
dimensional Cartesian coordinates of all nodes Ν in a global GPS coordinate system.
This GPS model can be jointly processed with the adjustment equations of the indirect
georeferencing in a hybrid block adjustment, as explained in Section 5.4. In this hybrid
block adjustment the ground coordinates of the nodes are treated as unknowns. The un-
known nodes Ν connect the GPS model and the orientation functions Xq(1), Yo(t) and
Zo(t). Within the hybrid adjustment the GPS model is translated, rotated and re-scaled
in order to compensate for possible datum problems and other error effects arising be-
tween the GPS data and the ground coordinate system. A division into separate GPS
models for every strip can also be advantageous. In the same way as GPS models can
be formed for the orientation functions Xo (t), Yo(t) and Zo(t), so IMU models can be
setup for the orientation functions ω(ί), φ(ί) and κ(ί).
The method presented here is a combination of direct and indirect sensor orientation,
i.e. an integrated sensor orientation, which has proven itself in practice.
Further reading for Section 5.5: Ries, C., Kager, Η., Stadler, P.: Publikation der
Deutschen Gesellschaft für Photogrammetrie und Fernerkundung, Band 11, pp. 59-
66, 2002. Gervaix, F.: PFG 2002(2), pp. 85-91. Hinsken, L. et al.: ASPRS, St. Louis,
USA, 2001.
5.6 Accounting for Earth curvature and distortions due to

cartographic projections
Image coordinates are introduced as observations in the bundle block adjustment (Sec-
tion 5.3), in the GPS- and IMU-assisted aerotriangulation (Section 5.4) and in the geo-
referencing of images from 3-line cameras (Section 5.5). In a prior process the image
coordinates are corrected for the effects of objective lens distortion, film deformation,
unflatness of the CCD detector surface, positioning errors in the detectors of CCD
cameras or film scanners and refraction, as explained in Section 4.5. This section on
image coordinate refinement also treats the correction for Earth curvature. The method
given (Section 4.5.3) cannot be directly applied to the wide areas normally covered in
aerotriangulation. Alternatives to the Earth curvature correction are therefore required.
Section 5.6 Accounting for Earth curvature and distortions 281
(a) the first method, which can be applied where there are relatively minor height
differences and small image tilts, makes radial corrections to the image coordi-
nates in a similar way to refraction (see Equation (4.5-3)). This radial correction
is given by the following equation (for the derivation of the formula see Sec-
tion Β 5.4.3, Volume 2):
(5.6-1)
ρ ... radial distance from principal point

h ... flying height above ground
R ... radius of the Earth = 6370 km
c ... principal distance
The ground coordinates of the control points, and the GPS locations of the per-
spective centres (where available) in ground coordinates, remain unchanged.
Any available IMU data relating to roll and pitch angles can also be adopted
without change. However, the heading angle (yaw angle), which is also provided
by the IMU, is referenced to the meridian and so a convergence of the merid-
ian must be applied. The convergence of the meridian 7 is the angle between
the meridian and the North/South gridline in the ground coordinate system, fre-
quently a Transverse Mercator coordinate system. It is given by:
λ (5.6-2)
7» = - ο) cos ψί
ψί, Xi . . . geographic coordinates of the point for which the

convergence of the meridian 7, must be calculated
Λο . . . datum meridian in the ground coordinate system
(b) a second method fundamentally solves the problem that the object coordinate
system in photogrammetry is a three-dimensional Cartesian coordinate system
and the ground coordinate system is not. The control points and GPS positions
of the perspective centres are given in the ground coordinate system; the pho-
togrammetric results are expected in this coordinate system. The second method
has the following steps:
• the control points and the GPS positions of the perspective centres (if avail-
able) are transformed from the ground coordinate system into a tangential
coordinate system. This is tangential to the Earth ellipsoid at the centre of
the area of interest.
• photogrammetric processing is done in this tangential system.
• the photogrammetric results are subsequently transformed from the tangen-
tial system into the ground coordinate system.
The IMU information requires special treatment. The discussion is here restricted to
the roll and pitch angles which are referenced to the physical plumb line. The Ζ coordi-
nates in the tangential Cartesian coordinate system are perpendicular to the X Y plane.
Depending on the rigour of the analysis, the plumb lines are perpendicular to the spher-
ical or ellipsoidal surface of the Earth; being specific, the tangents to the curved plum
lines at the IMU's point of measurement are the datum for the IMU information. The
IMU measurements must therefore be rotated by the angle between the physical plumb
line and the Ζ axis of the tangential coordinate system. This correction, which depends
on the IMU's measurement position as taken from the GPS data, can be calculated as a
rotation matrix using either a spherical or ellipsoidal model of the Earth's surface and
applied to the IMU data in the first calculation stage mentioned above.
Details of this rigorous method, which include accounting for the geoid, can be found
in Sections Β 5.3 and Β 5.4, Volume 2.
Exercise 5.6-1. Table 4.5-3 shows the maximum Earth curvature in stereomodels for
various image scales and camera types. Calculate the corresponding radial correction in
the image using Equation (5.6-1). (Solution: for the 1st value in the 1st row: c = 85 mm,
h = 850 m, ρ = 140 mm, Δρ = 25 μτη.)
Further reading: Wang, S.: DGK, Reihe C, Nr. 263, 1980. Ressl, C.: VGI 89, Heft 2,
pp. 72-82, 2001. Cramer, M.: DGK, Reihe C, Nr. 537, 2001. Heipke, C„ Jacobsen,
K., Wegmann, H. (Eds.): OEEPE Publication No. 43, 2002.
5.7 Triangulation in close range photogrammetry14

The sets of photographs in close range photogrammetry are often widely different from
the examples we have discussed so far, with 60% forward overlap and 20% side over-
lap. Instead of the photogrammetric stereomodel we have the photogrammetric bundle
of rays or directions (see Figure 3.8-11). A photograph records simultaneously a bun-
dle of spatial directions to all object points visible in the photograph. The conversion
of points initially recorded in the image plane into spatial bundles of rays occurs either
by manual measurement of the image coordinates (analytical photogrammetry) or au-
tomatic target point recognition (digital photogrammetry). If measurement or recogni-
tion is monocular, object points must either be targeted or must be natural points which
can easily be identified and accurately measurable in the individual photographs. For
small objects with an extent of a few metres at maximum, there is also the option of
projecting a pattern onto the object (see Section C 2.3, Volume 2. In Section 8.3 of this
volume an interesting variant is described).
The bundles of rays of the individual photographs are tied together by common object
points. Figure 5.7-1 shows such a network of photogrammetric bundles. Unmarked
camera stations are normally chosen (equivalent to free stations in surveying). The
camera may also be used without a tripod, so that the exterior orientation of the bundle
of rays is entirely unknown. The essential criteria for the photographs are: the individ-
ual object points must appear in at least two photographs, better in at least three, and
the rays must not intersect at narrow angles.
14
A comprehensive textbook covering this topic may be found in Luhmann, T., Robson, S., Kyle, S.,
Harley, I.: Close Range Photogrammetry. Whittles Publishing, 2006.
Section 5.7 Triangulation in close range photogrammetry 283
A least squares adjustment of such a network of photogrammetric bundles is performed

as a bundle block adjustment, described in detail in Section 5.3. The unknowns are the
six elements of exterior orientation of each photograph and the three coordinates of
each new point. Some control points are required on the edges of the object.
The X Y Z coordinates of particular imaging stations (perspective centres), which have
been determined by GPS or some other technique, can be introduced as known para-
meters (constants) in the adjustment equations (5.3-1). Due to the unavoidable inac-
curacies in perspective centre coordinates determined in this way, it is preferable to
introduce the terms as "observed unknowns" in the sense described in Section Β 3.5.7
(Volume 2). In this way, perspective centres determined by GPS, or some other method,
take on the role of control points.
In many cases, for example if the object is to be reconstructed in an arbitrary coordinate
system, it is possible to dispense with control points and execute a free net adjustment,
already mentioned in Section 4.4.1 and discussed in detail in Section Β 4.6.3, Volume 2.
In order to ensure that the object reconstructed in a local coordinate system has the
correct scale, at least one (reasonably long) known object point separation must be
included in the adjustment. The reconstructed object can be levelled by including new
points along vertical lines and on horizontal planes 15 .
At close ranges, a significant part is played by information on object form, such as the
distribution of points on straight lines, curves, planes and surfaces in three-dimensional
15
Some relevant publications include: Wester-Ebbinghaus, W.: IAPR 25(A3b), pp. 1109-1119, 1984,
and ZfV 110, pp. 101-111, 1985. Fräser, C.: IAPR 27(B5), pp. 166-181, 1988.
space. The corresponding extension to photogrammetric triangulation is examined in

Section Β 3.5.5, Volume 2. Section Β 4.1, Volume 2 also handles the very important
and practical issue of calculating approximate values for close range bundle adjust-
ment.
In close range work, semi-metric cameras (Section 3.8.4), amateur cameras (Section
3.8.5) and video cameras (Section 3.8.7) are often employed. Camera calibration, or
at least an improvement in the calibration, is incorporated within the photogrammetric
triangulation (this was discussed in Section 3.1.2 under the theme of on-the-job calibra-
tion or self-calibration). For this purpose the adjustment equations (5.3-1) are extended
beyond the unknown elements of interior orientation to include additional parameters
which compensate for optical distortion and other error sources in such cameras, as
already discussed in Section 5.3.5.
Also at close range GPS and IMU are increasingly employed. One of the many com-
binations of GPS/IMU with digital cameras is the mobile mapping system presented at
the end of Section 3.8.8, Figure 3.8-12. The bundle adjustment method plays a signifi-
cant role in evaluating the multi-sensor data from such systems. The bundle adjustment
equations (5.3-1) should be re-arranged so that the perspective centre coordinates Xoj,
Yoj, Z0j do not appear as unknowns in the equations but as directly observed parame-
ters (constants) or, better, as "observed unknowns" (Section Β 3.5.7, Volume 2). If
IMU data are available then the angles Wj, ψ 3 , Kj should be handled in a corresponding
way. The grouping of GPS and IMU data into GPS and IMU models, as introduced
in Section 5.4 for GPS and IMU-assisted aerotriangulation, is not normally required in
close range work. Potential datum and calibration problems, and other error sources,
can be more easily handled in terrestrial applications than in survey flights16.
For the arbitrarily configured bundles found in close range photogrammetry, the accu-
racy of close range bundle adjustment can only be very approximately estimated with
the previous rules of thumb (4.6-1) and (5.3-7). Here it is a condition that the camera
rays intersect the object points at good angles and the entire block is surrounded by
control points. The critical parameter in deciding accuracy is, as always in photogram-
metry, the camera-object distance.
Numerical Example. Signalized points on a truck are to be measured using a P31

normal-angle metric camera (c = 20 cm, see Section 3.8.3), at distances of about 5 m
from the truck. Equations (5.3-7) yield the following estimates of accuracy:
perpendicular to the camera axis: σ = ±0.003 χ 5/0.2 = ±0.075 mm
along the camera axis: σ — ±0.00003 χ 5000 = ±0.15 mm
Where images fill the format with a useable extent of 100 mm, the relative accuracy
(Section 4.6) is:
Perpendicular to the camera axis: 3/100000 « 1/33000 = 0.03%o
Exercise 5.7-1. How would the accuracy change if a P31 wide-angle camera (c =
10 cm) were used instead of the normal-angle camera? The camera-object distance
16
Extensive further reading on mobile mapping can be found in Section 3.8.8.
Section 5.7 Triangulation in close range photogrammetry 285
could then be halved without increasing the number of photographs. On the other hand,
the shorter distance (wider angle) may increase the number of dead (invisible) areas in
the object. (Solution: σ (perpendicular to camera axis) = ± 0 . 0 7 5 mm, σ (along camera
axis) = ± 0 . 0 7 5 mm.)
More comprehensive accuracy rules than the rule-of-thumb methods of Equations

(4.6-1) and (5.3-7) have been given by, amongst others, Fräser 17 , Grün 18 and Schlögel-
hofer 19 . These take into account control point layout, object shape and the number and
quality of the intersections of rays through new points (Schlögelhofer's results are pre-
sented in Section Β 4.5.2, Volume 2). Since networks of photogrammetric bundles vary
widely from project to project, the computer programs for the adjustment must estimate
the variance of each new point from the inversion of the matrix of normal equations.
In this way, it is possible to check whether the intersections in the object points are too
narrow, whether a sufficient number of rays intersect in a point and whether the num-
ber and layout of the control points are satisfactory (if no free network adjustment has
been made). Luhmann et al. 20 also promote computer-supported measurement plan-
ning using the inverse of the normal matrix. Various orders of network optimization
are defined for this concept 21 .
"Ph.Eng. 53, pp. 4 8 7 - 4 9 3 , 1987.

18 IAP, Commission V, Stockholm, 1978.
19 Geow. Mitt, der TU Wien, vol. 32, 1989
20 Luhmann, Τ., Robson, S., Kyle, S., Harley, I.: Close Range Photogrammetry. Whittles Publishing,
2006.
21 Further reading: Fräser, C.: in Atkinson (ed.): Close Range Photogrammetry and Machine Vision,
pp. 2 5 6 - 2 8 1 , Whittles Publishing, 1996. Mason, S.: Mitt. Nr. 53, Institut für Geodäsie und Photogram-
metrie, ΕΤΗ Zürich, 1994.
Chapter 6
Plotting instruments and

stereoprocessing procedures
Stereoprocessing is central to this section, in the narrow sense of reconstruction of

three-dimensional object models from corresponding points and lines in two-dimen-
sional metric images. The determination of the elements of exterior orientation, either
for the whole block (Chapter 5) or separately for individual stereomodels (Chapter 4),
however, must precede the actual stereoprocessing. This section is to a large extent lim-
ited to the processing of two overlapping metric images. Automatic digital processing
will certainly be considerably more robust and reliable when more than two images are
dealt with simultaneously. A human operator can observe only two images in stereo
simultaneously. For a human operator the process of measurement, or of matching cor-
responding points and lines in two overlapping photographs, can most conveniently,
quickly and accurately be carried out with the help of stereoscopy. The different sys-
tems for stereoscopic observation as well as the principle of stereoscopic measurement
are now considered, not only for analogue but also for digital images.
6.1 Stereoscopic observation systems

A short preamble on natural spatial vision is given before the introduction of the dif-
ferent stereoscopic systems used by photogrammetric operators.
6.1.1 Natural spatial vision

When we observe a point Ρ and its surroundings, the axes of the two eyes are both
directed towards P ; this is known as convergence. In addition the focal length, and also
the aperture, of the lenses in the eyes are changed so that a sharp image is generated on
the retina; this is known as accommodation. In the plane of observation, defined by the
eyebase BA (which varies around 65 mm) and the fixed object point P, there will be
two different images of the one object, one on each retina. In Figure 6.1-1 can be seen
the difference in parallax pA which relates to the difference in distance between the
object point Ρ and its neighbouring point Q. For simplification in the representation,
in Figure 6.1-1 the direction of Q from the left eye is the same as that of P; their images
in the left-hand retina coincide. With the head held upright, the plane of observation
Section 6.1 Stereoscopic observation systems 287
P,= Q
Horizontal plane Vertical plane
of observation
Figure 6.1-1: Natural stereoscopic vision
is approximately horizontal; for this reason we speak, in the context of natural spatial
vision, of horizontal parallax and differences in horizontal parallax.
The fusion of the two retinal images, differing in horizontal parallaxes, resulting in a
perception of three-dimensional space, is known as stereoscopic vision or stereoscopy.
The smallest perceptible difference in angle d*y — 7 P — 7q determines the resolution
of stereoscopic perception or the stereoscopic acuity. This angle d-y lies in the range
of about 5" to 10". Monocular acuity, on the other hand, is of the order of about 30".
In relation to the standard "minimum distance of distinct vision" of about 25 cm, the
stereoscopic acuity lies in the interval 6 μιη to 12 /im and the monocular acuity is about
40 μπι 1 . Photogrammetric measuring accuracy is distinctly better (see Sections 2.1.7
and 4.6).
This improvement arises from optical or digital enlargement of the images, so that
the acuity of the human eye is not the limiting factor of the man-machine processing
system.
On the other hand images exhibiting such differences in horizontal parallaxes ρ a can
be fused into a single three-dimensional image only when the angle d-y is smaller than
about 1.3 gon (about 1°12')· If this angle is exceeded either the foreground or the
'Albertz, J.: ZfV 102, pp. 490-498, 1977. In comparison with the accuracy of measuring individual
image coordinates, the different monocular and stereoscopic acuities result in a higher accuracy in the
measurement of ξ-parallaxcs, although not in the anticipated ratio of 1 : 6 to 1 : 3. Further reading, which
also contains advice on additional literature, may be found in: Schenk, Τ.: Digital Photogrammetry. Terra
Science, 1999.
288 Chapter 6 Plotting instruments and stereoprocessing procedures
background appears to dissolve into two separate images; one can easily verify this by
experimentation (but see Section 6.1.2(a)).
In contrast to the horizontal parallaxes ρ a (Figure 6.1-1), which lie in a plane con-
taining the eyebase B A , vertical parallaxes are defined as being perpendicular to the
eyebase. Homologous rays emanating from object points produce no vertical paral-
laxes since the rays are of necessity coplanar. In natural stereoscopic vision there are,
therefore, no vertical parallaxes.
Exercise 6.1-1. For an observer lying on his side, his eyebase is vertical. Think criti-
cally about the implications for the definitions of horizontal and vertical parallaxes.
6.1.2 The observation of analogue and digital stereoscopic images

Suppose that two images have been taken with parallel axes normal to the base (the
"normal case"); they will exhibit horizontal parallaxes in the direction of the base but
no vertical parallaxes perpendicular to the base (Section 2.1.5). They can subsequently
be presented to the two eyes. The observer interprets the horizontal parallaxes as sig-
nifying differences in depth in a virtual spatial image. One speaks also of a three-
dimensional image (3D picture).
Before dealing more closely with the observation of stereoscopic images, principles
should be set down as to how one arrives at stereoimages with horizontal parallaxes
but without vertical parallaxes:
• "normal case" photographs taken, for example, with a stereometric camera (Sec-
tion 3.8.2)
• conversion of arbitrarily oriented photographs into the "normal case" using the
elements of interior and exterior orientation (the favoured method in the process-
ing of digital metric images Section 6.8.5.2)
• projection, with oriented projectors, of arbitrarily oriented photographs onto the

xy plane of a relatively oriented stereomodel (the standard method in analogue
photogrammetry with optical projection Section 6.3(a))
• observation of arbitrarily oriented photographs not in a common image plane but

in two separated image planes; the photographs have previously been oriented
according to the elements of relative orientation (standard method in analogue
photogrammetry with mechanical projection instruments, Section 6.3(b))
• observation of arbitrarily oriented photographs in a common image plane while

a mechanism is running in the background using the elements of relative orienta-
tion to remove vertical parallaxes (standard method of analytical photogramme-
try, Section 6.4)
(a) observation of analogue stereoscopic images

In the observation of two stereoscopic photographs lying side by side at approximately
25 cm, the minimum distance of distinct vision, an uncommon demand is made of
the two eyes. With their axes almost parallel, that is with near-zero convergence, the
eyes are being asked to accommodate at a short focussing distance, that is, they are
being asked to overcome the unconscious reflex which normally imposes coordination
of focus and convergence. After long practice and with the exercise of considerable
will power, perhaps aided by the positioning of a sheet of card or paper between the
two images so as to separate them visually, fusion can be achieved with the naked eye,
resulting in a more or less sharp stereoscopic image.
Stereoscopic observation can, however, be made significantly easier if one places the
photographs under a lens stereoscope (Figure 6.1-2). The rays entering the eyes are
almost parallel and, if the photographs are a distance / below the lenses, their virtual
images are at infinity and may be viewed with ease and comfort.
Large format photographs require the use of a mirror stereoscope (Figure 6.1-3). In
most such instruments optical systems are placed above the lenses L\ and L2 so that
sections of the images can be seen magnified.
Figure 6.1-3: Mirror stereoscope {e\ + ej + e-i — / )

In most analogue and analytical stereoplotters (for example Figure 6.3-2b) stereoscopic
observation occurs by way of lenses or mirror stereoscopes in conjunction with com-
plicated optical systems.
Some analogue stereoplotters employ superimposed stereoscopic images (see, for ex-
ample, Figure 6.3-1). Figure 6.1-4 shows superimposed stereoscopic images of a pyra-
mid. The eyes converge and accommodate in the accustomed way at a comfortable
short distance but each eye sees both of the two-dimensional images. Suppose, how-
ever, that the left-hand image has been coloured cyan and the right-hand one red and
that a filter in the complementary colour has been placed in front of each eye (Sec-
tion 3.2.2.1). Then each eye will see only one image in black, while the other image
will disappear; the two black images will be fused in the usual manner into a three-
dimensional, stereoscopic image. This anaglyph system, as it is known, is very suitable
for the printing of stereoscopic images. Coloured images, however, cannot be observed
in this manner.
Figure 6.1-4: Superimposed stereoscopic images of a pyramid (from cameras as illus-

trated in Figure 6.1-8). Solid lines for the left image, broken lines for the right
(b) observation of digital stereoscopic images

Digital stereoscopic images can also be observed with either a lens stereoscope or a
mirror stereoscope if the two images are presented to the eyes, either on two screens
side by side or on a single, split screen. The anaglyph system mentioned above can
also be applied.
Digital stereoscopic images offer other, very interesting, variations, one of which is
sketched in Figure 6.1-5. The two images alternate on the same screen and are ob-
served through spectacles incorporating LCD shutters. An infra-red system controls
the synchronization of the spectacles and the screen (Figure 6.1-5); if the frequency of
alternation of the images exceeds 50 Hz a continuous stereoscopic image is observed.
A second variant is shown diagrammatically in Figure 6.1-6. Under the control of a

synchronizer the two images alternate in time in a bitmap memory and on a polar-
Figure 6.1-5: Shutter glasses for stereoscopy
Image display
monitor
Polarized screen
Polarized
glasses
Figure 6.1-6: Polarization for stereoscopy
ized screen. Stereoscopy results with the use of correspondingly polarized spectacles
(Figure 6.1-6). Polarization today also uses LCD systems.
Both these variants of stereo-observation can be used simultaneously by more than one
person. As a result, difficult problems of interpretation can be discussed by a group
while observing the stereomodel.
(c) additional information concerning stereoscopy

The second set of methods for the stereoscopic observation of digital images, which
may be denoted passive stereoviewing, requires simple spectacles, as opposed to the
first methods which may be denoted active stereoviewing. From the point of view of
the operator, both variants have advantages and disadvantages, the first usually provid-
ing better image quality since there is no filter while the second variant uses light and
cheap spectacles. Stereoviewing over many hours is relatively demanding whichever
variant is considered, especially when the frequency of the images is low. For this rea-
son, for detailed stereoscopic study many operators still prefer analytical stereoplotters
with analogue images on film. At present great efforts are being made in research and
development to produce stereoscopic images which could be viewed without specta-
cles, filters and so on. Such systems would offer many advantages. Interesting illusions
can be produced also when observing stereoscopic images. If, for example, one inter-
changes the red and cyan filters in the anaglyph method, the tip of the pyramid in
Figure 6.1-4 no longer appears to lie above the square base but beneath it. Instead of
the orthoscopic effect the pseudoscopic effect is observed. When viewing aerial pho-
tographs pseudoscopically one gets an inversion of the relief in which mountain ridges
become valleys, streams flow towards their source and so on.
Stereoscopic perception can also be disturbed. Thus the limit of 1.3 gon (1°10') (Sec-
tion 6.1.1) should not be exceeded in the section of the image under observation.
Schwidefsky/Ackermann gives another measure for the limiting value for stereoscopic
vision: the difference in scale between the sections of the left and right images being
viewed should not be greater than 14%. Since in high mountains and in terrestrial im-
ages this limit is frequently exceeded, provision of individual zoom controls for each
image is of great assistance in stereoplotters.
As well, the density or the colour of the two images should not be greatly different.
The procedures for enhancement of digital images reviewed in Section 3.5 are also of
great significance for stereoscopy. All in all, however, the human eye is very tolerant:
even when one eye is presented with an RGB image (Red/Green/Blue) and the other
eye with a B&W (Black and White) image a spatial impression (in colour) is obtained.
The sole condition is that the intensities of the RGB image should, to a large extent,
match those of the B&W image; one makes use of this when observing images from a
three-line camera (Section 3.7.2.3).
Provided that the geometrical parameters of the photographed scene are of no impor-
tance but that it is the significance of individual objects that is under investigation,
one frequently dispenses with a complete orientation of the images and with three-
dimensional geometrical processing. Instead, for such photointerpretation one makes
only a makeshift orientation as sketched in Figure 6.1-7 in which it is assumed that the
photographs have been taken with the camera axis near-vertical, typically the case in
air photo-interpretation. This consists simply of laying the two photographs on a plane
surface for observation and aligning the principal point bases (Figure 6.1-7) whereby
the vertical parallaxes are removed, admittedly only along this common straight line;
with increasing distance of the image region from this favoured line, however, the
vertical parallaxes increase, depending on such things as the tilts. Nonetheless, with
stereometers or parallax bars, as they are called, horizontal parallaxes, and thus the
^-parallaxes between two particular points can be measured and from these the height
difference between both 3D points computed. Detailed discussion of these methods
and their limitations were to be found in older textbooks.
Photo 1 Photo 2
pp
Ip2 pp
, 1 2 ,
PP^
Figure 6.1-7: Orientation of near-vertical images for stereoscopy
(d) spatial impression in stereoscopic observation

The question remains to be asked concerning the extent to which the virtual spatial
image arising from two stereoscopic images is distorted with respect to reality. The
relationships with respect to the photograph, both wide angle and normal angle, are
shown on the left of Figure 6.1-8 and on the right are sketched those relating to stereo-
scopic viewing. BA is the eyebase. The virtual spatial image has come nearer to the
observer by the factor H/DA ; Η is the flying height and DA the apparent distance at
which the virtual model is formed.
The quantity DA depends on the design of the stereosystem, on the observer and also
on the image quality. Unexpectedly, DA is independent of the magnification of the ob-
serving equipment and also of the separation of the two images as they lie beneath the
stereoscope. One can easily satisfy oneself with respect to this physiological phenom-
enon by experimental magnifications and shifting of the images. For lens and mirror
stereoscopes with greatly differing magnifications and different image material the ap-
parent distance DA is found to lie in a range 2 from 35 cm to 60 cm. In the case of dig-
itally superimposed images, as in Figures 6.1-5 and 6.1-6, DA is the distance between
the eyebase and the screen; in this case DA can, in principle, be arbitrarily increased
from the initial value of the minimum distance of distinct vision, approximately 25 cm.
As the observed virtual 3D-image comes closer by the factor Η / D A its scale in relation
to reality also decreases, to be precise, in object planes parallel to the image planes by
a factor H/(cv), where Η is the flying height, c the principal distance of the taking
camera and υ the enlargement factor of the stereo-observing instrument. For digital
images on the screen υ is the zoom factor.
Of greatest importance is the question of the degree to which, in the observation of
stereoimages, the dimensions of objects in the direction of the camera axes are exag-
gerated in relation to vertical distances. Using the notation of Figure 6.1-8 this vertical
2
Miller, C.: Ph.Eng. 24, pp. 810-815, 1958, and 26, pp. 815-818, 1960. Collins, S.: Ph.Eng. 47,
pp. 45-52, 1981. (For good stereo-observers DA is small.)
exaggeration factor is defined as:
62 : (6 1
= Δ ^ ΔΧ · "1)
The values ΑΧ and ΑΧ a can be expressed as the 71th part (for Figure 6.1-8 η = 4) of
the camera base Β and of the eyebase Β A 65 mm) respectively:
AX = Β/η AX A = ΒA/η (6.1-2)
The values AZ and AZA follow from the first relationship of the equation system
(2.1-32), in which —Z is substituted for the flying height H :
ΑΖ = Ηυ-ΗΡ = ^ - - ^ - = (6.1-3)
Ρξ,υ Ρζ,Ρ Ρξ,υΡξ,ρ
The parallax ρξ// is also there in the observation of the virtual spatial image, so that
the "apparent principal distance" CA may be derived from the first relationship of the
equation system (2.1-32) as follows
CABA
Da = => cA = —
Ρξ,υ Β A
Section 6.2 The principles of stereoscopic matching and measurement 295
Then, in accordance with equation (6.1-3):

Ρξ ρ ρξ υ D
Δ Ζ λ = 0ΑΒΑ '- ' = m,uBAPt,P-P(,u = DAPi,P-Pt,u
Ρξ,υΡξ,ρ Ba Ρζ,υΡζ,ρ Ρς,ρ

Substituting from Equation (6.1-2) to (6.1-4) in Equation (6.1-1) gives:
_ D A P m l _ D A B _ Β BA Λ
62
~ BaC ~ BaH ~ Η ' DA ^ j
Result: The height exaggeration factor ez is the ratio of the base/height factor B/H
of the stereopair to the base/distance ratio ΒA/DA in the observation of the virtual
three-dimensional image.
Table 6.1-1 gives the height exaggeration factor ez in the case of normal angle, wide
angle and super-wide angle photographs with an image format of 23 cm χ 23 cm and a
60% forward overlap (base/height ratios are given in Table 3.7-1).
Normal angle Wide angle Super-wide angle

c [mm] 300 150 85
B/H 1 : 3.3 1 : 1.6 1 : 0.9
ez 1.9 3.8 6.7
Table 6.1-1: Height exaggeration factor ez in the observation of stereoimages with

60% forward overlap when BA/DA = 65 : 400 = 1 : 6.2
A disproportionate spatial impression, an increased spatial quality, is observable when

examining normal aerial photographs stereoscopically. This exaggeration of relief in
the virtual spatial image plays a significant role in photo-interpretation. In photogram-
metry, on the other hand, there is usually no need to worry about the phenomenon.
The stereoscopic effect serves simply to hasten the matching of corresponding points
in both images and to aid in accurate setting of the measuring mark on a required point
in the spatial image.
Exercise 6.1-2. Make a superimposed drawing in two complementary colours of Fig-

ure 6.1-4 and observe it, both orthoscopically and pseudoscopically, through filters of
complementary colours.
Exercise 6.1-3. In order to see an undistorted stereoscopic impression of images from
Wild or Zeiss fixed-base stereocameras (Section 3.8.2), what camera-to-object distance
should be used? (Answer: 2.5 m and 7.4 m, respectively, for camera bases of 40 cm and
120 cm.)
6.2 The principles of stereoscopic matching and measurement

A (virtual) three-dimensional measuring mark is used for the geometric analysis of
the stereomodel, the three-dimensional image, presented to the operator. The operator
sets this three-dimensional measuring mark on points and traces it along lines, in this
manner abstracting information from the images, converting it to discrete values. The
stereoplotter becomes a 3D digitizer. In this process of analysis the photogrammetric
operator brings his knowledge and judgment to the choice of points and lines, interpret-
ing their significance and capturing their meaning in encoded form. In this way human
vision, recognition and recall are put to service in the processing of the stereomodel.
Machine vision, which is not considered in more detail until Section 6.8, attempts to
transfer this whole process to the computer.
This section concentrates on the design and operation of the three-dimensional mea-
suring mark mentioned above, beginning with those measuring marks used with lens
and mirror stereoscopes.
(a) beneath a lens or mirror stereoscope

One uses two real measuring marks which can be moved over the surface of the
two-dimensional images (Figure 6.2-1). The two images fuse into a virtual three-
dimensional image and, so long as no vertical parallax exists, the two measuring
marks M\ and M2 also fuse into a single virtual three-dimensional measuring
mark, known as the floating mark. If the right-hand mark moves from the posi-
tion (M 2 ) to the correct M2, then the virtual three-dimensional measuring mark
will appear to move from the position (M) to the position Μ in the virtual spa-
tial image. Shifting of the measuring marks in the two-dimensional images is
perceived as a vertical displacement of the floating mark in the virtual spatial
image.
After the floating mark has been set on a definite object point, the real measuring
marks M\ and Mi each coincide with the image of that point in their respective
two-dimensional photograph. With appropriate measuring systems the image
coordinates can be determined in each image (Section 6.4).
Digital images may be viewed in analogue form as images on two screens or on
a single split screen; in this case the cursors on the screen(s) can serve as the real
measuring marks.
(b) with digitally superimposed images

The discussion is limited to digitally superimposed images that are displayed ei-
ther as alternating real image sequences (Figure 6.1-5) or as alternating polarized
images (Figure 6.1-6). There is only a single cursor which one sees in both im-
ages. If the cursor is not located on homologous images then the floating mark
is perceived as hovering either above or below the virtual surface. In order to
set the floating mark on the virtual surface one or both of the images must be
shifted in the direction of the base. The cursor then becomes a 3D cursor; its ap-
parent vertical movement relative to the virtual surface can be controlled by the
operator, using a wheel on a mouse-like device. With the wheel one of the two
images can be displaced in the direction of the base until the 3D cursor lies on
the surface of the stereomodel (as in Figure 6.2-1). Following this stereoscopic
setting on a point one knows the positions, in other words the image coordinates
Section 6.3 Analogue stereoplotters 297
marks
ξ and η, in both images; so both corresponding points are located and stereo-
scopically matched. Using the recorded image coordinates and the elements of
interior and relative orientation, or in other cases the exterior orientation, the xyz
coordinates are computed in real time, or in the latter case the X Y Z coordinates
(Sections 4.3.2 or 4.1).
6.3 Analogue stereoplotters

Between about 1910 and about 1970 probably more than a hundred different types of
analogue plotters were designed and constructed. They are mentioned here for histori-
cal reasons and are of minor importance nowadays.
In an analogue stereoplotter the position and orientation of the taking cameras is re-
produced at a reduced scale. The relationship is established, by means of optical and
mechanical components, between the model of the object at a reduced scale and the
two metric images at their original scale. Two principles of design which have been
developed may be described either as optical projection or as mechanical projection.
(a) principle of optical projection

A widely used stereoplotter with optical projection was developed by the Amer-
ican firm Kelsh (Figure 6.3-1). Both projectors are set up for an image format
of 23 cm χ 23 cm. They have the same interior orientation as the camera used
to take the photograph. After setting of the exterior orientation the intersection
points of corresponding rays define a true model without geometric distortion.
Figure 6.3-1: Kelsh-Plotter 3
Each projector may be rotated through angles ω, ψ and κ; this allows the rela-
tive orientation (independent model method) to be carried out. The rotations Ω
and Φ required for absolute orientation (levelling the model) may be set using
the footscrews (5). Since the model scale is determined by the ratio of the base
length in the model to the base from the photoflight, the desired model scale may
be set by shifts of both projectors. Note that the complete process of relative and
absolute orientation may be carried out without computational aids.
Stereoscopic observation is by the anaglyph method (Section 6.1.2(a)). A light
spot on the table (6) is used as the floating mark for stereoscopic setting on the
model surface. The links (10) ensure that the lamps (9) project corresponding
portions of the images.
(b) principle of mechanical projection
In mechanical projection each straight line joining image point, perspective cen-
tre and object point is represented by a link, known as a space rod, usually in
the form of a slender rod of cylindrical cross-section which rotates about one
of the perspective centres. The principle of mechanical projection is briefly ex-
plained with reference to the Wild Aviograph (Figure 6.3-2a). The object point
is driven in the x, y and 2 directions on a three-dimensional cross slide system.
The space-rod is mounted
• at the object point in a ball-joint

• at the perspective centre in stationary gymbal axes (also known as a Car-
dan(ic) mount)
3
Kelsh, H.T.: Ph.Eng. 14, p. 11, 1948.
Section 6.3 Analogue stereoplotters 299
• at the image point in a ball-joint connected to the photo-carrier which may

be translated in a plane
The portions of the space rod connected to the ball-joints may be lengthened or
shortened along their axes, often referred to as being telescopic.
Each photo-carrier can be rotated through angles ω and φ, the axes of rotation
passing through the respective perspective centre. In addition each image can be
rotated in a plane about its principal point (κ rotation). The longitudinal tilt of
the model Φ can be set on an adjustment screw (Φ in Figure 6.3-2a). The other
inclination of the model for absolute orientation Ω is set directly with ω\ and ωι
on the two photo-carriers.
Corresponding to the principal distance is the distance of each perspective centre

from that plane which is parallel to the tilted plane of the image and which passes
through the centre of the image point ball-joint. This ball-joint, at the upper end
of the space rod, moves the photo-carrier on a two-dimensional cross slide under
the control of the space rod.
A measuring mark M, a light-spot illuminated by the light source L (see the

left-hand part of Figure 6.3-2b), is fixed in the lower, stationary part of the tilted
photo-carrier. The beam-splitter F causes the measuring mark Μ to be reflected
into the optical system and brought into coincidence with the image point Pi.
After absolute orientation which, as has already been mentioned, can be carried
out without numerical computation, the operator sees a virtual three-dimensional
image not only of a section of the images but also of the measuring mark. Now
he moves the x, y, ζ cross slides, and with them both images, until the two mea-
suring marks coincide with the image points Pi and P2; that is, the floating mark
has been set stereoscopically on point Ρ and its object coordinates have been
found.
An important special feature of almost all analogue stereoplotters is the Zeiss

parallelogram. On the one hand it allows the perspective centres to remain in
fixed positions in the instrument and, on the other hand, it avoids the difficult
problem of mechanical intersection of the space rods in order to find the object
point. The base is introduced by relative movement not of the perspective cen-
tres, as in the Kelsh plotter, for example (Figure 6.3-1), but of the ends of the
space rods; this is achieved in the Aviograph AG1 by adjustment of b and
(Figure 6.3-2a). More information on the Zeiss parallelogram and on analogue
stereoplotters is to be found in older textbooks.
L Illumination for the F Beam splitter

measuring mark Κ Eyepiece
Μ Measuring mark
(b) Optical system
Figure 6.3-2: Aviograph AG1
6.4 Analytical stereoplotters

It is an interesting fact that until the end of the first decade of the 20 th century virtually
all photogrammetry applied analytical methods. From that time for more than half a
century analytical methods were displaced by analogue procedures in plotting. This
situation is rapidly changing with the advent of high-speed computers.
6.4.1 Stereocomparators
It is very simple to explain the mode of operation of a stereocomparator in terms of the
cross-slide system shown in Figure 6.4-1.
Section 6.4 Analytical stereoplotters 301
Figure 6.4-1: Diagram of a stereocomparator
The two images may be moved in coordinate directions corresponding to the perpen-
dicular cross-slides. The two measuring marks M\ and Mi are fixed in the focal planes
of the eyepieces. Images of parts of the photographs are also focused in these planes,
so that the measuring marks may be set stereoscopically on object points (Section 6.2).
Disregarding the zero settings for the time being, such a stereoscopic setting yields the
image coordinates ξι and η\ on the lower cross-slides and the parallaxes ρς = ξ\ — ξι
and ρη = ηι — r/2 on the upper cross-slides from which the image coordinates £2 and 772
may be derived. So that these uncorrected image coordinates from the stereocompara-
tor scales may be transformed into the camera coordinate system, it is also necessary
to measure the coordinates of the fiducial marks (Section 3.2.1.2).
When the image coordinates are recorded by a human operator, the X Y Z object coor-
dinates of the measured points are found by computation using analytical photogram-
metry. Since in its early days photogrammetry was concerned exclusively with ter-
restrial photographs, usually conforming to the "normal case", the labour involved in
calculation of object coordinates was kept within limits 4 .
Advances in analytical photogrammetry received a very decisive stimulus with the de-
velopment of electronic data-recording and electronic data processing. At this point
electronic data registration should be introduced.
4
In 1909 Eduard von Orel constructed the Stereoautograph, an instrument for "automatic" plot-
ting from stereoscopic photographs, circumventing this computational labour (Mitteilungen des k.u.k.
Militärgeographischen Institutes XXX, pp. 63-93, Wien, 1910). In England F. V. Thompson was slightly
before von Orel in the design and use of the Vivian Thompson Stereoplotter (Atkinson, K.B.: Ph.Rec.
10(55), pp. 5-38, 1980. Atkinson, K.B.: Ph.Rec. 17(99), pp. 555-556, 2002). Both Thompson and Orel
solved the simple equations of the "normal case" (3.8-1) by means of lineals; at one end these were
fixed to the slides of the stereocomparator; at the other end they controlled a drawing device. The in-
struments permitted the continuous drawing of plan detail and contours. Recent publications touching
on the historical development of the stereocomparator and the stereoplotter include Meier, H.-K.: ZfV
128(1), pp. 6 - 1 0 , 2 0 0 3 , and Luhmann, T., Robson, S., Kyle, S., Harley, I.: Close Range Photogrammetry.
Whittles Publishing, 2006.
6.4.2 Electronic registration of image coordinates in the

monocomparator
An efficient system for the registration of image coordinates is explained on the ba-
sis of the precise comparator PK1 from the firm of Zeiss 5 (Figures 6.4-2 and 6.4-3).
Two glass scales at right angles to each other are fixed with respect to the baseplate.
The image to be measured lies on a photo-carriage which can be translated in two di-
mensions without rotation; the left-hand and the upper borders of this photo-carriage
each carries a series of equidistant lines. Two linear pulse generators measure the dis-
placements of the image. The measuring mark Μ is stationary. As in the instrument
shown in Figure 6.4-1, it is situated in the focal plane of the optical system. The rel-
ative movements of the individual parts of the monocomparator can be studied with
the aid of Figure 6.4-2 which shows, in broken and continuous lines respectively, their
positions when measuring points 1 and 2.
Base plate
position for measurement of point 1

position for measurement of point 2
Figure 6.4-2: Principle of the Zeiss precision comparator PK1
5
Schwebel, R.: ZfV 104, pp. 157-165, 1979.
Photograph
40 μΓΠ Photo-carriage
Scale
tr
Liym
Intensity of illumination/
magnitude of current
10 20 30 40 μιτι
Figure 6.4-3: Linear impulse generator (Since it has no influence on the measuring
principle, the parallel guidance system is not illustrated). The photograph, the photo-
carrier, the light source and the photodiode move together; the scale and the measuring
mark are fixed.
The linear pulse generator (Figur 6.4-3) warrants a closer examination. On the mea-
suring scale as well as on the photo-carriage, not only the width but also the separation
of the lines is 20 μηι. Rays coming from the light source and falling on a line are fully
absorbed; they are totally reflected from the regions between the lines on the photo-
carriage. When the lines on the photo-carriage are moved 10 μηι to the left from a
position in which they are aligned with the graduations on the scale (Figure 6.4-3),
only 25% of the emitted light falls on the photodiode D, which translates the incident
light into electrical current in proportion to the intensity of illumination. If the photo-
carriage is then moved to the left the current will increase. The maximum of 50% is
reached when the graduations on the photo-carriage are moved 20 μηι out of alignment
with the graduations on the scale. A further movement to the left reduces the current,
and so on.
The number of minima and maxima transmitted from the diode D to a counting system
defines the image coordinates as a multiple of 20 μτη intervals. Subdivision of the
20 μηι intervals occurs through digitization of the electrical signal; in the case of the
PK1 this is done in 20 steps which corresponds to an increment of 1 μηι. The image
coordinates ξ and η which are passed to the computer are obtained by measuring the
extreme value and digitizing the electrical signal.
The accuracy of a comparator is very much dependent on its design, especially on the
extent to which the Abbe comparator principle is fulfilled. This fundamental princi-
pie means that the distance to be measured and the measuring scale should, so far as
possible, lie in one straight line.
The greater the distance d between the scale and the segment I to be measured (Fig-
ure 6.4-4), the larger the measurement error Al attributable to out-of-squareness, such
as might be caused by a defective parallel guidance system of the photo-carriage. The
Abbe comparator principle is strictly adhered to in the monocomparator PK1 (Fig-
ure 6.4-2) as opposed to the stereocomparator sketched in Figure 6.4-1 in which it is
not.
For this reason no great accuracy requirement is placed on the parallel guidance system
for the photo-carrier in the PK1. The accuracy of the PK1 is more or less determined
by the accuracy of the glass scales and the accuracy with which the measuring mark
Μ can be held fixed in relation to the two scales. All these causes of error together
result in a mean square coordinate error of only ±1 μτη, an error which is less than
that arising for other unavoidable reasons in photogrammetry, such as irregular film
distortion which can amount to about ± 3 μτη (Section 3.2.1.1).
Figure 6.4-4: Measurement of segment I without regarding the Abbe comparator prin-
ciple
6.4.3 Universal analytical stereoplotter

The union of a stereocomparator (Figure 6.4-1) and electronic image coordinate regis-
tration (Section 6.4.2) results in an analytical stereoprocessing instrument, with which
the operator is able to register corresponding points in both images, upon which the
3D coordinates of the object point are automatically determined in the computer linked
to the instrument. Each photo-carrier requires its own electronic coordinate registra-
tion system, by virtue of which such an analytical stereoplotter may be classified as
an image space plotter. Image coordinates are the starting point for command of the
instrument and there is no feedback from the computer to the photo-carriages or the
measuring mark.
EiWWSl
1
)M 2
Computer
Image 11 Image 2
ryi φ ΓχΊ
Terminal
Figure 6.4-5: Universal analytical stereoplotter
The efficiency and generality of such an instrument can be increased in a fundamen-

tal way by means of a dominant incorporation of the computer in the processing, as
illustrated by the schematic diagram in Figure 6.4-5. This kind of instrument is fre-
quently classified under the rather clumsy heading of an "analytical plotter with object
coordinates primary" since the starting point for command of the instrument consists
of the XYZ object coordinates 6 . Using two handwheels and a footwheel the oper-
ator sets an object point in the XYZ coordinate system; alternatively, a 3D mouse
may be used as described in Section 6.2. From this xyz coordinate vector and using
the elements of interior and exterior orientation, the computer determines the corre-
sponding image coordinates ξι,ηι,ξ2,η2, also taking account of all image refinements
(Section 4.5). A prerequisite is that four servomotors should be capable of driving the
two photo-carriages to the appropriate positions corresponding to image coordinates
ζ\,ν\,ξ2,ν2', to accomplish this positioning operation four linear impulse generators
(Figure 6.4-3) register with the computer the image coordinates corresponding to the
current position of the photo-carriage. These coordinates are then compared with the
required image coordinates and, corresponding to the existing differences, commands
are given to the four servomotors for translation of the photo-carriers. This loop is re-
peated approximately 50 times per second. For all practical purposes the procedure is
instantaneous; such processes, running rapidly under digital control through electronic
data-processing, are known as real-time processes and programs necessary for them as
real-time programs.
6
Plans for a universal analytical stereoplotter were patented by U. Helava as early as 1957 (Phia
XIV, pp. 89-96, 1957/58 and Ph.Eng. XXIV, pp. 794-797, 1958). It was another 20 years before the
well-known firms extended their product range to include analytical plotters. Many references for further
reading are to be found in: Petrie, G.: ISPRS Journal 45, pp. 61-89, 1990, and ITC-J, Volume 4, pp. 3 6 4 -
383, 1992.
The control process by which the two photographs are moved under the stationary
measuring marks M\ and Mj (see Figure 6.4-5) is set in motion by the operator of
a universal analytical plotter using rotation of the two handwheels and the footwheel,
that is by the setting of the object coordinates Χ, Υ, Z. If the object coordinates Χ, Υ, Ζ
describe a continuous sequence of points on the surface of the object, the operator sees
the floating mark moving smoothly along this point sequence on the stereoscopically
observed model. The operator must aim to turn the handwheels and the footwheel
in such a way that the floating mark remains in coincidence with the surface of the
stereoscopically perceived object model.
Exercise 6.4-1. Using aerial photographs (1 : mB = 1 : 30000) a stereomodel is

set up in an analytical plotter. After introduction of the Earth curvature correction
the terrain height differs by 75 cm between the middle of the model and the corners.
Check whether this discrepancy agrees with the fact of Earth curvature. (Answer: This
difference is exactly in accordance with Earth curvature for this stereomodel)
6.5 Digital stereoplotting equipment

Because analogue photographs are used in analytical stereoplotters, relatively expen-
sive components are necessary, for example drive mechanisms employing servomotors
for the photo-carriages. If, on the other hand, digital photographs exist, translations of
the two images can take place on the computer screen beneath stationary measuring
marks. With such a digital stereo-processing instrument, therefore, the operator con-
trols the processing by moving a "3D mouse" (Section 6.2) on the table and rotating
the handwheel. The "3D mouse" defines the X Y Z coordinates of a point sequence in
the object model. If this digitized line runs along the object surface the operator sees
the floating mark glide down this line in the stereoscopic model. Just as with an an-
alytical stereoplotter, the task of the operator consists of moving the "3D mouse" and
rotating the handwheel so that the floating mark remains in contact with the perceived
stereoscopic model.
The requirement for digital displacement of both images in real time according to this
principle of a digital stereoplotter is very demanding. For the observation of digital
stereoimages (Section 6.1.2(b)) liquid crystals provide an image repetition frequency
of at least 50 Hz for each of the two images. If two images in sequence are shifted by
just one pixel, for a typical pixel size of 15 μιη the speed of displacement of the image
is:
50 Hz χ 15/im = 0.75 mm/s = 45mm/min
That is a relatively low speed for operator digitizing of 3D lines in the stereomodel.
Therefore, for higher digitizing speeds or in the case of images with a smaller pixel
size, the image repetition frequency must obviously be raised above 50 Hz.
A still greater problem than that of repetition frequency is the large image matrix of
digitized photographs which, as an example, for an aerial photograph with a pixel
dimension of 10 μπι amounts in total to 23000 χ 23000 pixels requiring approximately
0.5 GByte of space. The mentioned image data management can be handled as follows
Section 6.6 Computer-supported manual methods of analysis 307
(Figure 6.5-1). The screen has for example 1280 χ 1024 pixels. In what is called the
display memory of the graphic processor, which permits real-time access, a section
of the total image is held available, perhaps four times larger than the screen. If the
image excerpt on the screen approaches the edge of the segment in the display memory,
the display memory is appropriately updated from the main storage of the computer,
usually from a disk. For this purpose a precalculated direction of movement in the
stereomodel, on the basis of the line already digitized, is very helpful. In the case when
only coarse digitizing of a particular line is required (because, for example, the snake
algorithm will carry out automatic refinement of the measurements, Section 6.8.7.1) a
coarser image from the image pyramid (Section 3.6.1) can be held ready. As a result
the region of the image from which the screen display can be updated in real-time is
expanded. In many digital stereoplotters the zoom adjustment, controlled from the "3D
mouse", occurs in conjunction with different levels of the image pyramid.
Expected direction
of roaming
Computer
screen
Digital Display
image memory
Figure 6.5-1: Different portions of an image for roaming in real-time
Further details of the roaming, so important for digital stereoplotters, may be found in
Schenk, Τ.: Digital Photogrammetry, TerraScience, 1999, and elsewhere.
As a result, digital stereoplotters are distinguished by the fact that they do not only
permit manual operation but also reach a high level of automation. This is not con-
sidered in detail until Section 6.8. First various manual procedures, with considerable
computer support, are dealt with. Subsequently (Section 6.7) the accuracy of computer
supported processing is discussed.
6.6 Computer-supported manual methods of analysis

Dependent on the desired results there are various ways of working. Individual object
points can be selected in analytical and digital plotters by manually changing the X Y Z
coordinates, thus determining object coordinates (single point measurement). If one
follows along a line in the model in three dimensions, such as the edge of a street or a
wood, the coordinates of a dense sequence of points on this line are determined (line
measurement). If the Ζ coordinate is left unchanged and if the floating mark is kept in
contact with the surface of the stereoscopic model while changing only the X and Y
coordinates, a contour line is digitized (contour measurement). When measuring lines
a choice may be made between recording at constant intervals in time or at constant
intervals in distance. In areas of higher curvature a greater density of points results
when recording at constant time intervals, for here the operator moves more slowly.
The interplay of various methods of working will be examined more closely in the
following sections.
6.6.1 Recording in plan

Photogrammetric processes are always three-dimensional. The nature of many tasks is
such that only two-dimensional measurement is required (measurement in plan). The
third dimension, the height, is always available, but is either ignored or registered as an
attribute of the plan point.
Figures 6.6-1 and 6.6-2 show typical results in plan. Figure 6.6-1 portrays photogram-
metric output from an analogue instrument (Section 6.3). It is a manuscript which
must be revised and completed, a process which requires about as much time as for its
production on the analogue plotter. Computer support, therefore, is of great economic
importance in photogrammetric plotting.
It is seen that, to a great extent in the large-scale field, computer-supported plotting
delivers the final product. (Generalized representation at smaller scales, on the other
hand, is carried out by cartographers, frequently by hand.) Prerequisites for creating
the final product with a computer-supported system are convenient means of erasure,
insertion and changing (editing) of recorded information.
Figure 6.6-2: Graphical output from

Figure 6.6-1: Photogrammetric manu- computer assisted photogrammetric
script from analogue plotting plotting
Figure 6.6-2 shows a typical computer assisted photogrammetric product. Without

going further into the algorithmic and technical computing side, the elements of such
a plot are created as follows:
Buildings: The operator drives around the corner points of a building in turn. After
storing these points one by one the operator gives the command to "square up" the
connecting lines. The area defined in this way can be automatically hatched. Many
computer assisted plotting systems offer the further possibility that, when the operator
has driven around all but one of corner points of a building, the computer completes
the final one.
Roads and railway lines: Instead of laboriously driving down long lines in plan as is
necessary with an analogue plotter, the operator of a computer assisted plotter has only
to set accurately on a few points; the computer then connects these by interpolation
either with a polygon or with a higher order curve, as desired. A further possibility
which is of special interest when plotting railway lines is that only one line is recorded
in this way; the computer is instructed to add one, or more, parallel lines.
Fences and hedges: With computer assisted plotting there is also a suite of commands
for various kinds of lines in the object. Thus, in the graphical output, fences, hedges
and so on are directly drawn in their final cartographic form.
Symbols: As is the case for different kinds of lines, there is also a suite of commands
for different kinds of object points. After the setting of the floating mark at a certain
point and input of the command for the particular kind of object point, the desired
symbol will be drawn at this point on the plan being output. The placement of symbols
throughout the interior of a polygon can be particularly advantageous. The hatching of
buildings has already been mentioned in this connection but drawing the symbols for
trees, marshes and so on can also be very efficiently carried out in this way.
Spot heights: Spot heights deserve special mention. A single point is set at the desired
position with the attribute that this point should be appropriately represented on the
plan and should have the terrain height attached.
Text: Most computer-assisted systems also have a text generator. The operator man-
ually chooses the position for the lettering and then enters the letters or sequence of
numbers using the keyboard; these will ultimately be drawn on the plan in their final
form.
Map sheet preparation: One great gain from this technology is that the map sheet can
be prepared almost automatically. By that is meant the computer-controlled plotting of
such items as the control points, the rectangular border, a map coordinate grid, refer-
ence to latitude and longitude, a legend or key and text. In so far as not only a plan but
also elevations of the object have to be produced, a three-dimensional coordinate sys-
tem is defined and in individual drawings the appropriate two-dimensional coordinate
system is portrayed.
To conclude this section on plotting, field completion requires a brief mention. In
field completion, details which are not recognizable in the stereomodel, or cannot be
interpreted unambiguously, are gathered there and then on the spot. Field completion
also includes the registration of administrative information; this means such matters as
differentiation between official and private buildings, classification of paths and roads,
names of streets and places, limits of political and administrative units and so on.
6.6.2 Determination of heights

When it became possible, now many decades ago, to trace contours directly in the
stereomodel, it represented a great advance in topographic mapping and became the
highlight of analogue stereoplotting. While the procedure can also be used with ana-
lytical and with digital stereoplotters, the direct following of contours demands under-
standing of topography and great skill on the part of the operator. The operator should
not stare at the floating mark but should rather look ahead continuously along the con-
tour. Directly drawn contours can be regarded as a cartographic end-product only at
large scales and only for work of modest cartographic standard. Frequently it is neces-
sary in a re-working to adjust uncertainties and to make sharp bends at the breaklines
more precise 7 . Figure 6.6-3 conveys an impression of the appearance of directly traced
contours before any revision.
Within forested areas in large-scale stereomodels the contour lines cannot be directly
followed. One can frequently make do by measuring individual points in clearings
with subsequent interpolation of the contours. Laserscanning, which is dealt with more
closely in Section 8.1, offers an exceedingly interesting alternative in forested regions
and when the demand for accuracy is high. In small-scale stereomodels contours can
be followed directly even in forested parts. To do this one measures the tree height
photogrammetrically at the edge of the forest and then, after appropriate change in
7
Finsterwalder/Hofmann: Photogrammetrie. de Gruyter, 1968.
8
Taken from E. Aßmus: Geow. Mitt, der TU Wien, vol. 8, 1976.
the height reading, moves the floating mark over the crowns of trees representing the
forest canopy. Some operators on the other hand leave the height reading unchanged
and dive, as it were, into the forest with the floating mark. In either case, contours in
forest are considerably less accurate than on open land, a topic dealt with more closely
in Section 6.7. Not only in forested areas but also in flat land, direct following of the
contours is very risky; indeed on horizontal land contours are indeterminate.
Hence, alternative methods of photogrammetric height evaluation are to be considered
which have become so competitive that, with the exception of high mountains, they
have more or less displaced direct tracing of contours.
Figure 6.6-4 shows the result of a modern photogrammetric procedure whereby con-
tours are extracted indirectly from a digital terrain model (DTM).
Figure 6.6-4: Contours determined indirectly from a digital terrain model (DTM), for
comparison with those of Figure 6.6-3
The recording of the ground begins with the three-dimensional digitizing of the break-
lines (shown as broken lines in Figure 6.6-4) marking streams and often ridges. After-
wards a static raster measurement, as it is called, takes place (Figure 6.6-5) in which
the task of driving along a meandering path over the model is entrusted to the com-
puter, the operator meanwhile having only to control the Ζ movement. The recording
of the XYZ coordinates during this operation is usually arranged in such a way that
the points lie on a square grid in the XY plane. For increased accuracy the travel along
the profiles is not continuous; before the X Y Z coordinates are taken the movement in
XY comes to a complete stop. If satisfactory stereoscopic heighting is not possible at
the pre-set X Y position, because of vegetation or the presence of a house for example,
that raster point is skipped or a substitute point near the defined X Y point is used.
By interpolation from this data and the digitized breaklines, the next step is to produce
a Digital Terrain Model (DTM), sometimes known as a DEM or a DGM (elevation
and ground respectively), from which by-products, for example contour lines, can be
derived. Such indirectly derived contour lines give a very good 3D impression (Fig-
ure 6.6-4); the integrated breaklines emphasize important geomorphologic elements.
Figure 6.6-5: Static raster measurement
6.6.3 Recording of buildings

The high potential of stereophotogrammetry can be demonstrated on the basis of the
recording of three-dimensional buildings. In this task the topological information, that
is the connectivity between points, lines and surfaces is especially important. The
topology in the case of simple buildings is captured in the computer-supported system
in the form of standard models as knowledge bases. Figure 6.6-6 shows a graphical
user interface with a ridge roof as a standard model. After having chosen the rele-
vant standard model the operator has only to measure manually the corner points of the
building. In this way the topological model receives its scale. Before taking a measure-
ment on the building point in the stereoscopic model the operator clicks on the relevant
building point in the topological model (Figure 6.6-6). If corner points cannot be set
in the photogrammetric model they are added by the computer, using the corner points
already measured and the geometric conditions implicit in the current standard model.
In this work and in the case of two blended standard models (as, for example, a ridge-
roofed extension to a ridged roof such as those in Figure 6.6-7) this measurement of
building points and building edges is of great benefit. To this end the operator clicks in
the topological standard model (Figure 6.6-6) around the middle of the ridge on which
the point to be measured will lie; then he sets a point in the stereomodel on the chosen
building ridge. Such a working technique is designated topology assisted recording.
To assist such a computer supported plotting system, more accurately a topology as-
sisted system, a superposition system is required. This offers the possibility for those
I Building editoi
Identifier Γ"Τ;§
I - Buidng prirribve
VRML.
- - - ~•
J
I I Saddleback Rod 3 Op»*« | Union 3 Newpm*iyc |
Remove pumftive
Accept pnmfove
Reconstruct
J
Undo
J
NewbuMng Accept! lim I Removetxjlcing |
Ctee j
Figure 6.6-6: Graphical user interface for topology assisted recording (the standard
model is known as a "primitive")
Figure 6.6-7: Superposed building measurements in one image (Taken from: Rotten-
steiner. F.: Geow. Mitt., vol. 56, 2001.)
points and lines which have already been measured to be overlaid on one of the two dig-
ital photographs (Figure 6.6-7). For this purpose the X Y Z object coordinates are to be
transformed into the relevant photograph, using the parameters of interior and exterior
orientation and taking account of the image coordinate refinement (Section 4.5). Still
better is a three-dimensional superposition in the optical stereomodel; this means a su-
perposition of the already measured features in both of the digital images. Apart from
the check on accuracy which this provides, the operator is constantly kept informed of
what he has already measured and what he still must do.
Figure 6.6-8: Oblique view of photogrammetrically processed buildings (Taken from:

Rottensteiner, F.: Geow. Mitt., vol. 56, 2001.)
Figure 6.6-8 shows a rendered 9 oblique view of an assembly of buildings which co-
incides to a large extent with Figure 6.6-7. As a result of the topological information
concerning the building surfaces it is no problem to declare, for example, that roof sur-
faces are dark. In Section C 1.1.3, Volume 2 there is a description of how it is possible
for the surface topology to be structured in a relational database.
Incidentally, the Level of Detail (LoD) can also be efficiently regulated using the stan-
dard building model. In the above example a standard model without chimneys was
used; larger extensions were, however, modelled. The final roof form in the case of a
ridge roof with a ridge-roofed extension is formed by the blending of two ridge roofs
as one can see for oneself in some of the blended buildings. Suites of CAD (Computer
aided design) programs provide such blending functions. (The subject of more highly
automated photogrammetric measurement of buildings is entered into in Section 6.8.8
in this volume.)
'Rendering: generation of a two-dimensional image from a three-dimensional model with the appli-
cation of various lighting effects (shadows, surface reflections, etc.).
The fa?ades of buildings cannot be acquired from aerial photographs. Facades must
be captured with terrestrial photogrammetry, and evaluated either in two or in three
dimensions. Figure 6.6-9 shows a typical example. The fa$ade was plotted with an
analytical stereoplotter. Computer support is especially efficient in such work. Three-
dimensional geometric models of windows, ornamental features and such-like can be
created, taking account not only of topology but also of proportions, and can be placed
in multiple positions in the CAD model. In this way, for example, one of the windows
can be accurately measured and then this window can be duplicated in positions chosen
by the operator. In this kind of work the superpositioning system mentioned above is
especially important for it allows the operator continually to check for any possible
departures from a "standard window".
I O S E P H O II IANSA
ET
MARIA T H E R E S I A AVGVSTA
IMPERANTIB
ERECT CIDIDCCLXXV
Figure 6.6-9: Fagade of the Gloriette of Schloss Schönbrunn (Plotted by: J. Tschan-
nerl, I.P.F.)
6.6.4 Transition to spatially related information systems

Photogrammetric plotters can build up relatively universal systems with which a very
wide variety of products can be generated. Thus, for example, when digitizing an ar-
bitrary three-dimensional line in real-time a running count of the plan distance can
be computed and on a graphical output device a profile, showing height against dis-
tance, can be drawn, folded over and unwound. Or the computer can be commanded
to drive the floating mark along geometrical figures in plan such as circles, polygons,
and clothoids while the operator inputs the Ζ, height, coordinate. The result is a length
profile over a pre-determined mathematical route traced in plan.
The developments towards universal systems outlined above at present assume less
and less importance: photogrammetric stereoprocessing equipment is taking on the
task only of three-dimensional digitizing. Digital object models are constructed from
the resulting data. As required, various derivatives are acquired from the object mod-
els such as contours, straight and curved profiles, slopes, curvatures, oblique views,
shading, and others.
In this connection several definitions are presented: digital object models are a simpli-
fication of the real world which originate through idealization and discretization and
which are accessible for systematic use in electronic data processing. To the digital
object models belong therefore not only the data, the three-dimensional coordinates of
discrete points, but also
• encoding of the significance of points, lines and surfaces

• elements concerned with data structure
• the algorithms for conversion from the discrete points to curves and surfaces
If the digital object models are designed as the central database then the next step
emerges a spatially related information system.
These definitions can be still further refined in the case of topography: digital topo-
graphic models are a simplification of the natural and cultural landscape which arise
through idealization and discretization and which are accessible for systematic use in
electronic data processing. The digital topographic models are the central database of a
topographic information system which is a subset of superior geo-information systems.
Digital topographic models can be still further divided:
• the digital terrain model (DTM: see Figure 6.6-10) refers to the surface of the
terrain, without buildings or vegetation
• the digital surface model (DSM: see Figure 6.6-10) refers to the surface with the
inclusion of buildings and vegetation
• the digital building model refers to the buildings, including their roofs
• the digital street model refers to roads and street
• etc.
Section 6.7 Operator accuracy with a computer assisted system 317
DTM
Figure 6.6-10: Terrain model (DTM) and surface model (DSM)
6.7 Operator accuracy with a computer assisted system

Measurement accuracies attained by the operator of a computer assisted photogram-
metric instrument are given in this section. With technology as it stands today, there is
no need to differentiate between analytical and digital equipment. Rules of thumb are
given for rough estimation of accuracy, which also serve well in project planning. The
structure of the section is directed towards the products discussed in Section 6.6.
6.7.1 Measurement in plan

6.7.1.1 Point measurement
With analytical and digital stereosystems, as explained in Section 6.6.1, objects are
recorded, above all, through individually measuring single points which are connected
by means of lines. The accuracy of point measurement was dealt with in Section 4.6
and the following numerical example should recall the statements made there.
Numerical Example. The photogrammetric plot in Figure 6.6-2 originated with a

photoflight at a flying height of 1600 m using a 21 cm camera. What is the accuracy
of the local path which was recorded point by point and that of the fence, likewise
recorded point by point? What is the accuracy of the spot height? The photo scale
comes to 1 : m B = 1 : 7600 (= 1600/0.21).
According to the rule of thumb (4.6-1) the accuracies for very well defined points are
as follows:
Plan: σχγ = ±4.6 cm (= 7600 χ 0.0006)
Height: = ±9.6 cm (= 160000 χ 0.00006)
For natural points the uncertainty of definition in plan and in height, σ χ γ ^ 0 def) and
σζ(o,def)> according to the relationships (4.6-3), must be taken into account. We make
the following assumptions: for the local path a X Y ( 0 i i e q = aZ(o,def) = ±1 cm and for
the fence CTxy(0 def) = a z ( f i M ) = ± 8 cm, while for the spot height az{0,def) = ± 2 cm-
Using these values the relationships (4.6-3) yield the following accuracies for the plan
drawing of Figure 6.6-2:
Local path: σχγ = ±4.7 cm σζ = ±9.7 cm

Fence: σχγ — ± 9 . 2 c m σ ζ = ±12.5cm
Spot height: = ±9.8 cm
Note: the point density along the lines must be chosen such that no significant er-
ror of interpolation will be introduced. The object-specific uncertainties of definition
σ
χγ(o,def) and σΖ(o,def) should be evaluated in the photogrammetric instrument, prefer-
ably with large-scale images, by means of repeated measurements.
Exercise 6.7-1. The accuracy needs to be improved, above all in height. For a practical
project, for example with respect to the local path, one expects accuracy of σχγ —
± 2 cm and σ ζ = ±2.5 cm. How can this accuracy be achieved? (Answer: Use of a
15 cm camera (wide-angle) instead of a 21 cm camera. Making full use of the accuracy
potential of bundle block adjustment with additional parameters (Section 5.3.4) and
reducing the flying height.)
For a first trial:
Flying height 800 m, implying 1 : m s = 1 : 5300.
For very accurately defined points: σχγ — ± 1 . 6 c m σζ = ± 2 . 4 c m

For the local path: σχγ — ±1.9cm σζ — ± 2 . 6 c m
The requirement for plan accuracy is met. In height the value is slightly excessive; this
can be remedied by a slight reduction in flying height.
6.7.1.2 Processing of lines
Continuous guiding of the floating mark along a line is afflicted with considerably
greater inaccuracies than in single point setting. By empirical investigation as early as
1954 Heißler 10 found an accuracy of about
σι — ±45 μχη at photo scale (6.7-1)
Considering a drawing accuracy specification in final cartography of ±0.2 mm, an en-

largement ratio from image to graphical output no greater than 1 : 4.5 may be chosen.
Such an enlargement ratio is perfectly normal when mapping at scales from 1 : 1000 to
1 : 2000. It is not possible in small scale mapping, however, to use such high enlarge-
ments from image to plot, which would imply the use of very small-scale photographs;
this is determined not by limits on graphical accuracy but by limits on interpretability
of the content of images at very small scales. Thus, for example, the operator who must
portray lines in plan in a topographical map at 1 : 50000 can only just make them out
in photographs at about 1 : 60000. The optimal choice of photo scale in view of
10
BuL 22, pp. 67-79, 1954.
• accuracy in large scale mapping and
• interpretability in small-scale mapping
is expressed in the following rule of thumb (6.7-2) or in Table 6.7-1 (the parameter k
varies between 200 and 300 depending on the quality of the aerial photographs and on
the demands on the end-product):
TUB = ky/rriM (6.7-2)
1 : mM mB
1 1000 6300 -9500
1 5000 14000 -21000
1 10000 20000 -30000
1 25000 32000 -47000
1 50000 45000 -67000
Table 6.7-1: Relationship between map scale TUM and photo scaleTOΒ
The draughting accuracy of ±0.2 mm plays a central role in the considerations of

Heißler, mentioned above. With the digital graphical output of the results produced
using analytical or digital instruments, the relationship between draughting accuracy
and map scale is largely lost. The operator can choose, almost, any arbitrary scale for
the graphical representation. Just as tricky are the zoom possibilities of the graphical
systems which permit arbitrary enlargement of the content and give the user the im-
pression, because of the fine lines even when the enlargement is very high, of superior
accuracy of the data set. These risks can be removed if, at the same time as the graph-
ical representation is magnified, the lines are intensified. The map scale numbersTOΜ
given in Table 6.7-1 and in the Formula (6.7-2) refer to drawings in which one expects
an accuracy of ±0.2 mm. Graphical representations with considerably worse accuracy
than this do not warrant description as photogrammetric products.
6.7.2 Height determination

6.7.2.1 Directly drawn contours
The conventional method of contouring in analytical and digital instruments is by direct

tracing in the model (Figure 6.6-3 shows an example). The accuracy of such contours
is dependent on the slope a of the land as first expressed by Koppe:
= er ζ + ffc tan a (6.7-3)

σζ — 0.2%o of the distance from camera to ground, which is the accuracy of con-
tinuously traced lines from photographs with 60% forward overlap. Com-
pared with the accuracy of Equation (4.6-1) this is considerably worse
since this is a dynamic event unlike the measurement of an individual
point. In areas of forest this value is too optimistic.
σο = 100 μτα (at photo scale) which is the plan accuracy of directly drawn con-
tours. This is distinctly worse than the plan accuracy of Formula (6.7-1)
since the contours in the stereomodel are not marked in the photographs
but have to be sought.
For a statement of the accuracy—with reference to the object—of directly drawn con-
tours, in the case of aerial photographs in open terrain, Equation (6.7-3) is modified as
follows:
(6.7-4)
h ... flying height above ground;

c ... principal distance (in [mm]);
tan α . . . slope of the ground
The dimensions for are the same as those chosen for h.
Numerical Example. The contours of Figure 6.6-3 originated with a photoflight using
a 15 cm camera and a photo scale of 1 : 30000. Equation (6.6-3) gives the following
accuracy for the contours (flying height = 4500 m):
σΗ[τα] = ±(4500 χ 0.0002+(0.10/150)4500tana)

σΗ[m] = ± ( 0 . 9 + 3 tan a )
In flat land, therefore, the height accuracy comes to ±0.9 m; on a 10% slope to ± 1.2 m;
and on a 25% slope to ±1.7 m; and so on. In forested areas these figures are increased
by a further ±2.0 m.
6.7.2.2 Relationship between contour interval and heighting accuracy
In topographic mapping it is usual to choose a contour interval which, in metres, is one

thousandth of the map scale number; thus for a scale of 1 : 5000 a contour interval
of 5 m.11 In flat country the interval should be reduced. In order that the character of
the contours should not be negatively influenced, the mean square error of the contours
should not exceed 1/4 to 1/8 of the contour interval, according to the quality demanded
for the end product.
"imhof favoured distinctly smaller contour intervals (Hake, G., Grünreich, P., Meng, L.: Kartogra-
phie. 8th ed., Walter de Gruyter, 2002).
Numerical Example (Continuation of the numerical example of Section 6.7.2.1). The

contour lines of Figure 6.6-3 were produced for an orthophoto at 1 : 10000 with a con-
tour interval of 10 m, which is in accordance with the above rule of thumb. Depending
on the quality stipulated for the end product, the accuracy of the contours should there-
fore lie between ±1.25 m and ±2.5 m. Since the maximum slope of the terrain is 25%,
application of Formula (6.7-4) results in a heighting accuracy of ± 1.7 m—as is already
reported above. The height accuracy of the contours is therefore too poor to meet very
high quality map requirements; for medium quality it is satisfactory. This statement
is valid for open country. In forested areas, with 25% slope, the height accuracy de-
teriorates to about ±2.6 m (— \/1.7 2 + 2.02), so that, even for very modest quality
requirements, the accuracy of the contours is too poor. The contour lines which touch
each other in Figure 6.6-3, bear eloquent witness to the poor height accuracy. Such
contancting lines may appear, or even cross one another, especially in forested areas.
Exercise 6.7-2. Re-consider the numerical example of Sections 6.7.2.1 and 6.7.2.2
under the assumption that, instead of a 15 cm camera, a 21 cm camera is used. For
reasons considered more closely in Section 7.3.2, the 15 cm camera is not popular for
orthophoto work. (Answer: The vertical accuracy of contours for a 25% slope is, in
open country, σπ = ±2.1 m and, in forested country, = ±2.9 m. This means that,
perhaps, instead of a 10 m vertical interval a 12.5 m interval should be selected.)
6.7.2.3 Contours obtained indirectly from a DTM
Contours obtained indirectly via a digital terrain model (DTM) have the accuracy of
the DTM. This amounts to:
(6.7-5)
h ... flying height above ground;

c ... principal distance ([mm]);
tan α . . . slope of the ground
The dimensions for are the same as those chosen for h.

The constant term in the expression (6.7-5), 0.15%o of the flying height, corresponds
approximately to the vertical accuracy of static raster measurement (Figure 6.6-5). In
forested country another value has to be added, as in Formula (6.7-3).
Numerical Example. The contours in Figure 6.6-4, derived indirectly from a DTM,
have the following accuracy (the parameters for the photoflight are those of Section
6.7.2.1:
a H [ m ] = ± ( 4 5 0 0 χ 0 . 0 0 0 1 5 + 0 . 1 5 χ 4 5 0 0 / 1 5 0 tan α )
a/i[m] = ± ( 0 . 7 + 4 . 5 tan α)
In flat, open country, therefore, the vertical accuracy of the contours comes to ±0.7 m,
in land with a 10% slope to ±1.1 m, and with a 25% slope it is ± 1.8 m.
The choice of spacing for the static raster measurements, the choice of vertical interval
in relation to the grid spacing of the DTM and also the geomorphologic quality of the
interpolated contours should be closely considered.
6.7.2.4 Measurement of buildings
The recording of buildings, introduced in Section 6.6.3, proceeds through single point
measurement in stereoplotters, mainly digital stereoplotters. Hence, the accuracy cor-
responds to that of single point measurement as reviewed in Section 4.6.
Numerical Example. For the buildings of Figure 6.6-8 image material was available
at a scale of 1 : 4500, taken with a 15 cm camera. On the basis of experience in this and
similar projects, the following may be taken from Table 4.6-1 as the appropriate values
for uncertainty of definition: — ± 7 cm, = ± 8 cm. Assuming these
values, the following accuracies for stereoscopic recording of buildings are derived:
σχγ = ± 7.5 cm = ±9.0 cm
The accuracy of points on buildings measured from large-scale images will therefore
be dominated by uncertainties of definition.
Numerical Example. For the recording of the fagade of Figure 6.6-9 image mater-
ial was available at a scale of 1 : 290, taken with the 45 mm camera Wild P31 (Ta-
ble 3.8-2). The distance from camera to object was 13 m ( = 290 χ 0.045). Because of
visibility problems increasing with enlargement of the base, the full format of the P31
in the base direction was not exploited. Two stereopairs were chosen each with a base
of 7 m, giving a photo-base of 24 mm ( = 7000/290). As is general at close range, the
accuracy should be estimated not from Equation (4.6-1) for very well defined points,
but from the relationship (2.1-35) in which one normally takes ± 6 μτη as the measuring
accuracy at image scale.
Accuracy in the facade,

at right angles to the camera axis, = ±1.7 m m ( = 290 χ 0.006)
Accuracy perpendicular to the facade,
in the direction of the camera axis, = ± 3 . 2 m m ( = 1.7 χ 13/7)
Additionally, using Equation (4.6-3), an object-related uncertainty of definition should

be taken into account; for a point on a fagade this is about ± 5 mm.
Exercise 6.7-3. With the same positions of the camera, but using a 100 mm P31 camera
instead of the 45 mm camera, how would the accuracies of this numerical example
change? (Answer: Accuracy in the fa$ade = ±0.8 mm; accuracy perpendicular to the
facade = ±1.4 mm. This increased accuracy comes at the cost of losing the upper
portions of the facade—when using the "landscape" orientation of the camera they do
not fall within the image.)
Section 6.8 Automatic and semi-automatic processing methods 323
6.7.3 Checking of the results

Photogrammetric processing should be validated at certain intervals by means of check
measurements which, for example, are carried out with a photo scale distinctly larger
than the original or with GPS measurements on the ground. Such comparative mea-
surements expose all error sources, beginning with potential errors in calibration of the
camera, through possible inadequacies in the operator, up to errors in the photogram-
metric instrument and shortcomings in the software. This expensive procedure can, to
a large extent, be replaced by checking of the model joins. In general errors are at their
largest at the margins of the models, so that the discrepancies in measurements between
two different stereomodels allow a good quality control. In this procedure errors in rel-
ative and absolute orientation are also revealed. Errors arising in the overlap zones of
the stereomodels also give a very reliable basis for consideration of the oft-mentioned
uncertainties of definition for object points and lines.
For this section and the preceding Section 6.7.2, Appendix 4.6-1, dealing with empiri-
cal determination of standard deviations and tolerances, can be of interest as well.
6.8 Automatic and semi-automatic processing methods

Digital photographs open up the possibility of automatic processing. In place of au-
tomatic processing methods, names such as machine processing or procedures based
on computer vision are also common. Some photogrammetric problems can be solved
fully automatically, others only partly automatically. Partly automated procedures,
which are treated in Sections 6.8.7 and 6.8.8, are known as semi-automatic procedures.
The topic of automatic processing commences here with the correlation algorithms.
They are central to many automatic procedures. Correlation processes take the place of
humans at photogrammetric instruments. Instead of human sight and action, machine
sight and action take over.
6.8.1 Correlation, or image matching, algorithms

Correlation algorithms solve the task of finding corresponding patches in two images,
usually images taken from different camera positions. In this connection one speaks of
image matching and also of the maximum similarity or best agreement of the two im-
age patches. One of the two images can even be a geometric figure, in which case the
term pattern recognition is very appropriate. In the following passages a well-known
correlation algorithm is explained on the basis of such pattern recognition. The geo-
metric figure might be a cross which may exist in the form of an artificially generated
digital image (Figure 6.8-1, left). This image is known as the reference image or the
reference matrix. One speaks also of a pattern matrix or a template. The second image
may be known as the search image.
The position of the cross, that is of the reference image, is sought in the second image.
The terms search window or search matrix are used. In Figure 6.8-1 a 5 χ 5 template
is compared with a 12 χ 12 search window; for the sake of simplicity the grey levels
are assumed to lie only between 1 and 9.
The cross obviously lies in the search image in position i = 8 and j = 7. The automatic
search for this position is made more difficult because
• on the one hand the search image has faded out, that is the grey levels contain
accidental defects, and
• on the other hand, on account of the finite size of the detectors with which the
search window was created, the grey levels are smeared, that is mixed pixels
appear at the edges.
111111
12 1 1 1 1
11111
11111
2 1111
119 11 9 1111
119 11 9 2 11 1
9 9 9 99 9 9 9 921
1 19 11 1111
119 11 1111
111111
111111
Template (cross) Search window
Figure 6.8-1: Reference matrix (template) and search matrix (search window)
For the sake of simplicity the solution of this correlation task is demonstrated, not in
two dimensions but with reference to a one-dimensional example (Figure 6.8-2).
6.8.1.1 Correlation coefficient as a measure of similarity
The desired position of the template in the search window is found by calculating
correlations. A measure for the correlation or similarity is the correlation coefficient r.
This is computed as follows from the standard deviations σΓ and as of the grey levels
gr and gs in both images and also the covariance ars between the grey levels of the two
images:
r ^ < ? r s = Σ(9ν - lh){gs - gl) (6 8-1)
2
>/E(Sr - E ( S s " 9s)
in which gγ and g J are the arithmetic means of the grey levels of the template and those
of the corresponding section of the search window respectively.
9rf
9
Template
1 2 3 4 5 6 7 8 9 10 11 12
Search window
1 Τ - k
1 2 3 4 5 6 7 8 9 10 11 12
-ξ[μηη]
10 50 100
Figure 6.8-2: Template and search window of a one-dimensional pattern
The correlation coefficient r is evaluated for all possible positions of the template in the
search window. The position with the largest correlation coefficient r is the position
sought.
Numerical Example (with reference to Figure 6.8-2). The correlation coefficient for
the 1 st position, computed with i = 1 , . . . , 7 (or k = 1 , . . . , 7), is:
E(gr-4.43)(g6-2.43) -26.3
η = = -0.40
y/U9r- 4.43)2 U9s- 2.43)2 10.5 χ 6.3
The correlation coefficients for all possible positions are collected in the following
Table 6.8-1 in which the matching ξ coordinates (the pixel size Αξ = ΙΟμπι) are also
given:
Position 1 st 2 nd 3 rd 4th 5 th 6 th
k 1-7 2-8 3-9 4-10 5-11 6-12
r - 0 . 4 0 - 0 . 5 1 0.00 0.73 0.92 0.24
40 50 60 70 80 90
Table 6.8-1: Correlation coefficients for the example outlined in Figure 6.8-2
The 5 th position is that sought; thus ξ = 80 μττι. The neighbouring correlation coeffi-
cients (r4 = 0.73 and = 0.24) mean that the optimal position must lie somewhat
before ξ = 80 μιη. Thus a fit in the subpixel region is sought in a second step.
6.8.1.2 Correlation in the subpixel region
The following will provide a fit in the subpixel region
• formation of a continuous correlation function in the neighbourhood of the lar-

gest correlation coefficients and
• determination of the maximum of this correlation function by setting the first

derivative to zero
Numerical Example (continued). The following second order polynomial is found

from the last three ξ coordinates:
r ( 0 = 24.96 + 0.67150ξ - 0.00435£ 2
The first derivative is:

r ' ( 0 =0.67150-0.00870$
Setting the first derivative to zero results in:
ξ = 77.2 ^m
Further details of this method, in particular in the subpixel region with two-dimensional
images, are to be found in Section Β 6.1.2.2, Volume 2.
A very widely used method of correlation in the subpixel region employs a least squares
estimation method (Appendix 4.1-1) and is known as least squares matching (LSM). In
LSM one starts from an approximate fit which can be first found using the correlation
procedure described in Section 6.8.1.1. If, after the approximate fit, the template and
the search window are referred to the same coordinate system, in our case that of ξ
coordinates, then the required translation of the template with respect to the search
window is only small, usually less than a pixel (Figure 6.8-3). Looking at the template
and the search image together in Figure 6.8-3 one can conclude that the template should
be moved somewhat to the left for an optimal fit, that is ξ < 80 μτη.
Let this small translation be b. The positions of the two grey level ensembles gr and gs
are therefore related as follows:
<?*(£) = <?r (·£ + &) (6.8-2)
Let the grey level gs contain a random component v. The right-hand side of Equa-
tion (6.8-2) must be linearized; since b is small, this results in the following equation:
(6.8-3)
9rl.
9
8 - Template
7-
6 -
5-
4-
3-
2 -
1-
ς [μη]
10 50 100
Search window
ξ Dm]
10 50 100
Figure 6.8-3: Template and search window after an approximate fit
=
5r(£)>Ss(£) corresponding grey levels in the two images, the number of such
pairs of grey levels being determined by the size of the template
(d ( Jr(C)/9ξ) = the slope of the grey level profile of the template at the particular
position; that is, (θςΓ(ξ)/δξ) = Ag/Αξ. For the first pixel Ag
is taken as Ag = g2 - g\ and for the last pixel, the n t h pixel,
Ag = 9n — 9n—ι i for those pixels lying in between (1 < i < n)
Ag is expediently chosen as Ag — gi+\ — gt ι; for these interior
pixels Αξ is twice the pixel dimension.
The procedure follows the formal system of adjustment by the method of indirect ob-
servations (Appendix 4.1-1)
ν = A T v — 1, ν = (ATA)_1AT1 (6.8-4)
The accuracy of the position unknown b comes out of the computation (Appendix 4.1-
1) after the standard deviation of unit weight σο has first been estimated from the el-
ements of the vector v. The standard deviation at, of the unknown b comes from the
following:
N=(ATA), Q = Ν"1, σ6 = ν ^ σ 0 (6.8-5)
The possibility of such error computation is a great advantage of the LSM over other
correlation methods.
For our example the least squares equations are as follows (6.8-3):
f 0 \ /
0.4 2
0.4 -1
0 b - -1
-0.4 -4
-0.4 0
y o V 0)
From the normal equation system 0.646 = 2, the required translation is found to be
6 = 3.1; hence the final position of the template in the search window has the following
ξ coordinate: ξ = 80 - 3.1 = 76.9 μπι.
The standard deviation of unit weight σο may be obtained from the values
ν = (0, - 0 . 7 6 , 2 . 2 4 , 1 . 0 0 , 2 . 7 6 , - 1 . 2 4 , 0 ) T
=>· σο = ^ 1 5 . 7 5 / ( 7 - 1) = ± 1.62 greylevels. The standard deviation of the ξ coor-

dinate, that is of the result of the correlation in the subpixel region, amounts to (Equa-
tion (6.8-5)):
μ
σ ( = σ6 = λ / ΐ / 0 . 6 4 σ ο = ±1.25 ™ , σ0 = ±2.0 um
greylevels
In this simulated example the accuracy of positioning in the subpixel region is almost
an order of magnitude better than the pixel size, which in our example is 10 μπι. This
high accuracy has been confirmed many times in comprehensive correlations with real
image data.
The reference image and the search image frequently differ from each other not only
to a small extent in positioning but also in grey levels. Following the example of
Section 3.5.1, a parameter c (contrast adjustment) and a parameter d (brightness ad-
justment) may be introduced into the grey level gr. Equation (6.8-2) is extended by the
inclusion of these two parameters:
9s(0 = cgrti + b) + d (6.8-6)
These equations can be linearized and re-arranged into correction equations for an ad-
justment by the method of indirect observations. The unknowns are the three parame-
ters b, c and d. The simultaneous determination of position in the subpixel region and of
the adjustments for contrast and brightness is not always to be recommended as it can
seriously impair the convergence, especially with a small correlation window and very
faded grey levels. In these cases the adjustments for contrast and brightness should
be performed before the LSM and a subsequent correlation in the subpixel region be
performed using Equation (6.8-3).
Exercise 6.8-1. Linearize Equation (6.8-6) and repeat the numerical example. Hint:
The unknown 6 should be replaced by the unknown 6 = 6 c. (Answer: c = 0.69,
d = 0.81, b = 3.1 μπι, 6 = 4.5 μπι.)
The formulation for LSM correlation in the subpixel region can very simply be ex-
tended for the two-dimensional correlation case. In this case the process is limited to
the two translations 0ς and bn. Adjustments for contrast and brightness for the two
image regions to be correlated are, therefore, abandoned. Following the example of
Equation (6.8-3), the correction equations for the two unknowns bξ and bv become:
v b b
= ( % ) t + ( Ψ ) " " ^(ξ,ν) ~ 9ν{ξ,η)) (6.8-7)
v ' (ξ,ν) V αη ' (ξ,ν)
From this the following normal equation matrix for two-dimensional correlation is de-
rived:
2
N = / Udgrm E(dgrm(dgr/dv)
(6.8-8)
~ \ E ( d 9 r m ( d g r / d V ) j:(d9r/dV)2
In this matrix the summations are over all pixels in the particular correlation window.
Equation (6.8-8) is the basis for what is known as an interest operator (Section 6.8.1.3)
and for feature-based matching which is dealt with in Section 6.8.1.4. Before that,
some general observations should be made concerning correlation:
(a) the accuracy of the positioning depends decisively on step changes in grey lev-
els in the images. Large changes generate very large elements on the principal
diagonal of the matrix Ν of the normal equations (Equations (6.8-4), (6.8-5) and
(6.8-8)); large elements on the principal diagonal of the normal equation matrix
lead to small weight coefficients (Equation (6.8-5)) and therefore also to small
standard deviations of positioning.
(b) enlargement of the reference matrix raises the accuracy if it means that additional
step changes in grey level are thereby introduced. This increase in accuracy can
also be estimated from the inverse of the normal equation matrix.
(c) independent of its individual position in the digital image, a reference image
always results in the same inverse N _ 1 of the normal equation matrix (Equa-
tion (6.8-8)); therefore, in order to recognize the same pattern in different places
in the image, one has to evaluate N _ 1 only once. This characteristic is important
in the processing of reseau photographs.
(d) the correction equations (6.8-7) can be weighted differently; for example the
central picture element can be given a higher weight. This brings to mind the
higher sensitivity of the fovea in the human eye.
(e) in the basic equation (6.8-2) aside from a translation b, a scale factor in the ξ
coordinate of the reference image can also be introduced. Such an extension is
necessary, for example, when template and search window have different pixel
sizes. In the two-dimensional case the parameters of a similarity transformation
or an affine transformation (Section 2.1.1) can be introduced.
(f) in a more complex implementation the LSM usually requires a number of iter-
ations. Between individual iterations a resampling (Section 2.2.3) of the ξη raster
of one of the two images is carried out.
Exercise 6.8-2. How can Equation (6.8-1) be re-arranged in order to economize on

computing time? Repeat the example of Table 6.8-1 only with use of the covariance
a r s as the measure of similarity. Is it adequate to enlist the covariance as similarity
measure? (Answer: Equation Β (6.1-1), Volume 2. In principle the covariance a r s can
be used as a measure of similarity. All the same, for the introduction of threshold values
which in practice are important, the correlation coefficient, which varies between zero
and one, is significantly better suited.)
Exercise 6.8-3. Repeat the whole of the exercise of this section under the assumption
that the reference image comprises nine pixels which have the values g\ = gi = =
g9 = 1,03 = 37 = 4,34 = g5 = g6 = 9. (Result: ξ = 76.8 ± 2.4 fim). Why does a
reduction in accuracy now occur? (Answer: The sharpness of the grey level gradations
has become less pronounced.)
Exercise 6.8-4. Repeat the whole of the exercise of this section under the assumption
that the pixels are not 10 μπι but 5 μπι in size. This results in a considerable rise in
accuracy. (Result: σι, = ±1.06 μπι.)
Exercise 6.8-5. Reduction in pixel size usually results in an increase in the noise com-
ponent in the digitized image. As a result the increased noise component is carried
into the computation in which some of the grey levels in the search window are arbi-
trarily changed to g^ — 2, <77 = 9,g\ \ — 3 in the 10 μπι image. How accurate is the
positioning of the template now with the 5 μπι pixels in the search window? (Result:
&b = ±1.11 /im.)
Further reading: Ackermann, F.: Schriftenr. d. Inst. f. Photogr. d. Uni Stuttgart, Heft 9,
pp. 231-243, 1984. Rosenholm, D.: The Photogrammetric Record, 12(70), pp. 493-
512, 1987. Trinder, J.: IAPR 27(B3), pp. 784-792, Kyoto, 1988. Luhmann, T„ Rob-
son, S., Kyle, S., Harley, I.: Close Range Photogrammetry. Whittles Publishing, 2006.
Schenk, Τ.: Digital Photogrammetry, TerraScience, 1999. See also Sections Β 6.1.2
and C 2.2.1 in Volume 2.
6.8.1.3 Interest operators
Since the quality of the correlation depends strongly on the texture of both of the cor-
responding image patches it is to be recommended in many cases, before the actual
correlation, to choose image positions especially suitable for correlation. Interest oper-
ators, as they are called, are suitable for this purpose. In this and the following section
we assume two images overlapping each other. The reference matrix is thus a (smaller)
excerpt from one of the two images.
A very well known interest operator uses the accuracy measure for the two transla-
tions ϋξ and bri of the LSM from Equations (6.8-7) and (6.8-8). For this purpose one
forms the grey level differences ( d g / d ξ ) := g^, 77)1+1,j — g^, v)i-i,j and (dg/Οη) :=
an
v)i,j+1 ~~ v)i,j-i d establishes, in the neighbourhood of one of the pixels, for
example for a 7 χ 7 window, the matrix Ν of the normal equations (Equation (6.8-8)).
The inverse Q of the normal equation matrix is directly usable for the analysis of
Figure 6.8-4: Marked pixels with a small value of t r Q
accuracy. The trace of the matrix Q, denoted by "tr Q", should be as small as possible.
The central pixels of such a window in which trQ is less than a specified threshold
value are marked in Figure 6.8-4.
Some markings lie on straight-line grey level edges, where an accurate correlation is
possible only in a direction perpendicular to the edge; in the direction along the edge no
accurate correlation is possible. On this ground the criterion for this interest operator is
extended insofar as not only a small value for tr Q is sought but also a good distribution
of edges in all directions. This second criterion can also be tested using the matrix Ν
of the normal equations. Details can be found in Section Β 6.4.1.1, Volume 2.
Interest operators can, for example, be applied in both images of a stereopair, indepen-
dently of each other. If the criteria for the interest operators are met in the correspond-
ing places in both images, it is guaranteed that an accurate correlation of the two image
patches will be possible. The inclusion of interest operators makes automated process-
ing markedly superior to operator processing. Nevertheless, the good operator would
still be inclined, even with poor texture in both images, to measure stereoscopically.
Förstner was the first to report the interest operator outlined above which analyses
the inverse of the normal equation matrix (6.8-8). In the literature it is known as the
Förstner operator. Another interest operator comes from Moravec. For this interest
operator the variance of the grey level differences of neighbouring pixels in a window
is computed, and compared with a threshold value, along both the rows and the columns
of the image as well as along both diagonals. Large variances offer good prospects for
correlation.
6.8.1.4 Feature based matching
An interest operator extracts features in a digital image; they are more or less the
intersection points of grey level edges. These points are positioned with single pixel
accuracy in the image matrix, in accordance with the Förstner operator introduced in
Section 6.8.1.3. But, as a secondary product, the values for the chosen interest criterion
also come up for the neighbouring pixels. As a result the possibility arises of finding,
from all interest values, a continuous function over a surrounding area and of setting its
first derivative to zero. In this way the position of a feature point in the subpixel region
is found (an example of such a positioning in the subpixel region is to be found at the
beginning of Section 6.8.1.2).
In the processing of a stereopair of images the extraction of features in the subpixel

region is carried out in both images independently of each other. Subsequently, corre-
sponding features in the two images are assigned to each other; this is yet to be spoken
of (see, for example, Section 6.8.3.2). If two features are found to correspond the task
of correlation is solved as well. Correlation through extracted features is known as
feature based matching as opposed to area based matching. LSM, for example, is an
area based matching correlation since within a region, or an area, the pixels from the
reference image and the search image are directly connected with each other.
Some points to note: The accuracy of feature based matching, lying somewhere around
1 / 4 pixel, is distinctly worse than that of LSM. Feature based correlation is, as a rule,
more robust than LSM. A substantial disadvantage of LSM is that the approximate
positioning required before the LSM step must be relatively accurate. Feature based
correlation is rotation invariant, that is, the two images can be rotated relative to each
other. Two-dimensional correlation using correlation coefficients (Section 6.8.1.1) can-
not be applied to images rotated with respect to one another.
Literature: Förstner, W., Gülch, Ε.: Proc. Intercommission Workshop on "Fast Process-
ing of Photogrammetric Data", pp. 281-305, Interlaken, 1987. Baltsavias, E.: Disser-
tation, ΕΤΗ Zürich, Institut für Geodäsie und Photogrammetrie, Nr. 49, 1991.
6.8.1.5 Simultaneous correlation of more than two images
The human being can correlate only two images. More than two images can be simulta-
neously correlated using machine vision. For the simultaneous correlation of the three
images from a three line camera, or three line scanner (Figure 5.5-2), an extension of
correlation from two to at least three images is necessary. In Figure 6.8-5 three image
excerpts are sketched, as they relate to each other after a coarse fitting. It is very clear
that a fitting in the subpixel regions will bring about an improvement. Equation (6.8-7)
can be used for this kind of subpixel fitting. Thus any possible simultaneous contrast
and brightness fitting is abandoned; it suffices to use only translations 6ξ and bri while
transformation with additional parameters (as is given in Section 6.8.1.2) is discarded.
It is possible to fit image 2 with image 1 and image 3 with image 1 more accurately
using Equations (6.8-7). If image 1 is shifted during these two fittings, two different
positions for image 1 will have been found. If, during both the first positioning and the
second, image 1 is held fixed, this shortcoming can be remedied: image 2 is shifted in
the first case and image 3 is shifted in the second case. Up to this point the remain-
ing possible pairing, image 3 with image 2, has been disregarded. Separate pairwise
Figure 6.8-5: Multi-image correlation
fittings thus have the disadvantage that not all possible pairings, in our case three, are
simultaneously looked at and that multiple positions are possible.
A possible variant is that one of the three images should be designated as master; we
choose image 1. A subpixel fit of image 2 to image 1 can be carried out using the
following equations, which result from relevant substitutions in Equations (6.8-7):
V = (dgi/dOh,21 + (a52/Ö7?)6r),2i - (<?ι (ξ, η) - 92(ξ, η)) (6.8-9)
Similarly, for the subpixel fit of image 3 to the (master) image 1, the following equa-
tions are derived:
W = (0WÖO&£,31 + (^3/077)^,31 - (9ι(ξ,η)-93(ξ,η)) (6-8-10)
The formation of the normal equations from Equations (6.8-9) together with Equa-
tions (6.8-10) leads to a system of normal equations in the four unknowns, which de-
compose into two smaller systems of normal equations, in one case in the unknowns
6ξ,2ΐ and brh2\ and in the other case in the unknowns Ις^ι and 6^31. While these are
indeed joint equations involving the grey levels of the three images, the grey level
differences between image 3 and image 2 are not, thus far, minimized.
Additional equations are, therefore, sought of which the absolute term contains the
r
grey level difference (<&(£, η) i)) and in which no further positioning unknowns
appear. The four positioning unknowns in the Equations (6.8-9) and (6.8-10) actually
define the geometric problem clearly because 6^,32 appears in the difference (ί>ξ,2ΐ —
6ξ,3ΐ) and bO,32 appears in the difference (6η,21 — ^,3i)· The extra equation which is
being sought comes from the difference of the two Equations (6.8-9) and (6.8-10):
dgi, , dg2
(6
d 3 ö53 '8"11}
The normal equations formed from Equations (6.8-9), (6.8-10) and (6.8-11) lead to a
simultaneous correlation of three images. In the case of four or more images corre-
sponding equations are added. The LSM algorithm outlined can be called multiple-
patch matching.
Exercise 6.8-6. Consider the correction equations for a four image correlation. (An-
swer: There are six unknown translations; the absolute terms read as follows:
(5i - 92), (9i ~ 93), (9i ~ 54), (53 - 92), (94 ~ 92) and (g4 - g3).)
A more general method, using a fictitious (master) image was given in Krupnik, A.
(PE&RS 62, pp. 1151-1155, 1996), in which references to further literature can be
found.
6.8.2 Automated interior orientation

In the case of photographs taken with a digital metric camera, the principal distance
and the position of the principal point are known. The interior orientation of digital
metric cameras is directly available for use.
If, on the other hand, the photographs were taken with a film-based camera and the
digital metric image originates in a subsequent digitization (Section 3.4), then the
processing begins with the location of the individual fiducial marks. The same kind
of task arises when the images are taken with reseau cameras (Section 3.8.4) or when
digitizing of a reseau plate is utilized (Figure 3.4-3).
What follows concentrates on the automated location of the fiducial marks. A reference
matrix is produced for the particular fiducial mark. This is therefore a typical exercise
in pattern recognition.
In order to minimize the time and cost of the search process:
• image pyramids are produced from the photographs (Section 3.6.1) and
• relatively small search regions are defined in the vicinity of the standard positions
of the fiducial marks (as, for example in Figure 3.7-7; it is required that when the
image is laid in the scanner it should be rotated and shifted as little as possible
from a standard position).
For the different levels of pyramid one requires different reference templates to match
different extracts from the image of the fiducial marks. Figure 6.8-6 shows these ex-
tracts for a typical fiducial mark. The pixel size is reduced in accordance with the
Figure 6.8-6: Negative image of a fiducial mark with three different extract sizes
reduction in size of the extract. In the step using the coarsest pyramid the largest ex-
tract, with a wealth of detail, is required for recognition. At the level of this pyramid
the correlation is performed with only the correlation coefficient as a measure of sim-
ilarity (Section 6.8.1.1). At the next pyramid level the result from the coarsest level is
invoked as an approximate position so that a relatively small search region can be cho-
sen. Advancing in this manner is continued until the finest pyramid level. A subpixel
correlation (Section 6.8.1.2) is always carried out at the level of the finest pyramid. In
practice, at the finest level, the achievable accuracy lies between 1/10 and 1/5 pixel.
Accurate location of the fiducial mark in the digitized metric image is one task; the
other task is to establish which particular fiducial mark has been located. This is not a
trivial task since the image may have been placed in the scanner right way round, upside
down, rotated through 90 or whatever. It would be best if each individual fiducial had
its own unambiguous identification markings, for example in the form of a pattern of
lines. Another possibility which includes an additional, asymmetric fiducial mark may
be mentioned with reference to Figure 3.7-7.
Section Β 6.1, Volume 2 contains information on those forms of fiducial marks which
are favourable for automatic measurement, on efficient computation for correlation
coefficients, on the use of binary images instead of grey value images, etc. Examples
of further literature on these subjects: Heipke, C.: ISPRS-Journal 52, pp. 1-19, 1997.
Schenk, Τ.: Digital Photogrammetry. TerraScience, 1999.
6.8.3 Automated relative orientation and automated determination of

tie points
The relative orientation of two overlapping metric images is a central task both in pho-
togrammetry and in computer vision. The mathematical bases of relative orientation
were set out in Section 4.3. It remains unsaid, however, how the image coordinates
of corresponding points in both digital images can be found. In recent years the au-
tomation of this measuring process has been developed successfully. The strategies
that have been implemented for its solution in software packages depend above all on
the configuration of the camera stations and the orientations of the images as well as
on the form of the object.
6.8.3.1 Near-vertical photographs with 60% forward overlap taken over land
with small height differences
Near-vertical photographs with 60% forward overlap are discussed first. Figure 6.8-7
shows such a stereopair. For correlation, reference matrices are chosen in the first
image and search regions are defined in the second image. In order to get a good dis-
tribution of orientation points the system chooses the reference and search matrices
in the standard positions, some times called the Gruber points (Section 4.3.4.2). This
choice takes place at the coarsest pyramid level. In order to increase redundancy in
the correlation, relatively large reference matrices are chosen. In highly distorted pho-
tographs, however, problems can arise with reference matrices that are too large; this
can happen when there are large height differences in relation to the camera-object dis-
tance. The correlations in this first step can be carried out using correlation coefficients
(Section 6.8.1.1); an estimate with subpixel accuracy is not necessary.
V
Image 1 •
•
> • • <
A
/
H
• •
•
Figure 6.8-7: Reference matrices in the first image and search matrices in the second
image for the coarsest pyramid level; additional reference matrices, containing interest
points, and corresponding search matrices for a finer pyramid level in the neighbour-
hood of standard point number 5.
The corresponding correlation windows found in this way are very good approxima-
tions to be used for correlations at a finer pyramid level. As indicated in Figure 6.8-7
for standard position number 5 in the first image, a number of small reference matrices
can be selected inside such a window. Feature points, which may be found using an in-
terest operator (Section 6.8.1.3) and around which the reference matrices are extended,
can be of great service in making this choice. (Incidentally, the choice of the reference
matrices in the first image at the coarsest pyramid stage can also be supported with
interest operators.) As illustrated in Figure 6.8-7, the search window in the second im-
age can likewise be chosen to be relatively small, but clearly larger than the reference
matrix.
It is sufficient for the relative orientation to run through three or four pyramid levels.
Only at the level of the finest pyramid should subpixel estimation be performed, for
example using LSM (Section 6.8.1.2), after matching using correlation coefficients.
The result is six homologous locations in each of which, in general, a number of ho-
mologous points lie. As a result of the high number of orientation points provided by
the automated procedure, the accuracy of the orientation elements is as a rule distinctly
better—on average by a factor between two and three—than the accuracies given in
Table 4.3-1 when using six orientation points.
6.8.3.2 Near-vertical photographs with 60% forward overlap taken over land
with large height differences
With big height differences in relation to the camera-object distance the strategy out-
lined in Section 6.8.3.1 can fail, especially when there are large areas which are not
visible in one or the other image. A first countermeasure consists of choosing search
regions in the second image which are perceptibly larger, above all in the direction of
the base. On account of the severe difference in scale which stems from the large height
difference, correlation coefficient and LSM methods can also fail. For this reason fea-
ture based correlation (Section 6.8.1.4), which is independent of scale and rotation, is
used; that is, in each image feature points are extracted with an interest operator and
corresponding image points are subsequently sought. At the level of the coarsest image
pyramid the properties of near-vertical photographs, which display only small differ-
ences in η coordinates between left and right images, are exploited in support of these
correspondence analyses. The results from the coarser pyramids are good approxima-
tions, in both coordinate directions, for the finer pyramid levels. Relatively many errors
are made in the process of finding matching points. The computation of the relative
orientation must, therefore, be combined with gross error analysis. Data snooping and
robust estimation, which are dealt with in detail in Section Β 7.2, Volume 2, are suitable
for this purpose.
Literature: Heipke, C.: ISPRS-J 52(1), pp. 1-19, 1997, which contains an extensive
literature list, is relevant to Sections 6.8.3.1 and 6.8.3.2.
6.8.3.3 Arbitrary configurations of photographs and objects with very complex

forms
In close range work, especially, many of the photographic configurations used (multi-
image photogrammetry) would exemplify the above heading. Automated relative ori-
entation in such a photogrammetric network is very difficult. First, using interest op-
erators, feature points are extracted, independently of each other, in both images. The
so-called epipolar geometry, which is not considered more closely until Section 6.8.5,
plays a central role in the subsequent process of allocating corresponding feature points
to each other. To be sure, epipolar geometry requires approximate values for the ori-
entation elements; it may be possible to recover these by directly observing and identi-
fying at least five homologous points and subsequently computing relative orientation
using the methods of Sections 4.3.2 or 4.3.3. In situations where the orientation must be
carried out without human intervention, the analysis of correspondence may be placed
on a good foundation by the introduction of object recognition. Here, for example,
feature points extracted using an interest operator are compared with each other and
topologically similar patterns are sought in the first and second images.
Literature: Haralick, R., Shapiro, L.: Computer and Robot Vision. Addison-Wesley,
Reading, USA, 1992. Van Gool, L., Tuytelaars, T., Ferrari, V., Strecha, C., Vanden
Wyngaerd, J., Vergauwen, M.: IAPRS 34(3A), pp. 3-14, Graz, 2002.
6.8.3.4 Line-based (edge-based) relative orientation
Prompted by the fact that human stereoscopic vision uses more lines and fewer points,
Schenk 12 proposed a method for automated relative orientation which uses extracted
edges for the correlation. To this end, edges are first extracted independently in both
images (Figure 6.8-21 shows an example in which relevant edges still have to be se-
lected and extracted). The form of the edges, for example the direction of the tangent as
a function of the edge length, is then analysed and similarities in the lines of both im-
ages are selected. When corresponding edges have been found either of two procedures
can be followed:
• either, distinctive points along corresponding edges, for example corners, can be
found and the relative orientation can be computed using these points following
the algorithm of Section 4.3
• or, a method based on associated lines can be used for the relative orientation.
Formulae for relative orientation based on straight lines can be found, for example, in
the work of Schwermann and Luhmann 13 .
For relative orientation using only straight lines, it is necessary to observe three images
together. An explanation: a straight line in the first image, together with the perspective
centre of that image, defines a plane; the corresponding straight line in the second
image, together with the perspective centre of the second image, defines a second plane,
which intersects the plane from the first image in object space. This is a necessary, but
not a sufficient constraint for relative orientation, since all such pairs will intersect in
space, independent of the relative orientation of the two images. Therefore, no solution
to relative orientation is possible, even with many pairs of corresponding straight lines
in two images. A third image with a straight line corresponding to the lines in the first
two images provides additional lines of intersection in object space; for correct relative
orientation all of these lines must coincide. This can be used for a combined relative
orientation of an image triple.
In the future line photogrammetry, merely touched upon in this section, will gain in
importance because, in the extraction of features, lines are more suitable than points.
12
Schenk, T.: Digital Photogrammetry. TerraScience, 1999.
13
Schwermann, R.: Veröffentlichung des Geod. Inst, der RWTH Aachen, Nr. 52, 1995. Luhmann, T.,
Robson, S., Kyle, S., Harley, I.: Close Range Photogrammetry. Whittles Publishing, 2006.
6.8.3.5 Tie points for automated aerotriangulation
Up to this point in Section 6.8.3 only the automated establishment of corresponding

points for relative orientation has been dealt with. The determination of tie points
for a bundle block adjustment (Section 5.3) proceeds in a similar fashion. The posi-
tions of the reference matrices and the definitions of the search matrices for a block
of near-vertical photographs with 60% forward and 20% side overlaps are sketched in
Figure 6.8-8; it is an extension of the pattern of Figure 6.8-7 from two metric images
to many.
V V V
• • • • • •
•
•
> • • < > • • • < • • •
• • • • • • • •
Λ Λ Λ
V V —Τ-
• • • • • • • •
> • • < > • • • < > • • • <
• • • • • • • •
Λ Λ Λ
~V
• • • • • • • • Γ
Figure 6.8-8: Upper-left portion of a block of photographs with 60% forward and 20%
side overlaps showing reference matrices and search matrices
Using the strategies and algorithms given in Sections 6.8.3.1 and 6.8.3.2 one could
establish the tie points pairwise, including the tie points between the strips. Of course
in this way one would get only points that tied together the two images concerned. The
accuracy of a bundle block adjustment is, however, increased in a fundamental way
if it is possible to find unambiguous points lying in the overlap zones of all affected
images. As can be gathered from Figure 6.8-8, that point which lies in the lower row
of the second image of the first strip connects six images together. To determine a tie
point in this area, therefore, the six-image correlation is to be applied (Section 6.8.1.5).
Literature: Fritsch, D., Tsingas, V., Schneider, W.: ZPF 62, pp. 214-223,1994. Schenk,
Τ.: ISPRS-J 52, pp. 110-121, 1997.
6.8.4 Automated location of control points

For indirect georeferencing, which also implies indirect sensor orientation, as well as
for verification either of direct georeferencing or of direct sensor orientation, control
points are required in the object coordinate system; generally speaking these are given
in the national coordinate system (see the introduction of Section 4.1.1 and that of
Section 5.4). Location of control points in digital photographs is required at different
stages of the photogrammetric process; in a sequence based on their importance in
practice these are:
• in metric photographs to which a bundle block adjustment (Section 5.3) is to be

applied
• in each of two metric photographs constituting a stereomodel which is to be

included in a block adjustment using the method of independent models (Sec-
tion 5.2)
• in a single metric photograph which has to be oriented using three-dimensional

resection (Section 4.2.1), for example in the production of orthophotos (Sec-
tion 7.3)
• in each of two metric photographs for their combined, single-stage orientation

(Section 4.2.2)
• in both of the two metric photographs constituting a stereomodel for which ab-
solute orientation is required (Section 4.4)
Finding the control points in the images depends overwhelmingly on whether targeted
or natural control points are available. The location of targeted (signalized) points
is a pattern recognition task, comparable to the automated location of fiducial marks
(Section 6.8.2). Undeniably, the location of targeted control points is considerably
more difficult than the location of fiducial marks for the following reasons:
• in the reference pattern, generally speaking, there is no supplementary geometri-

cal information, such as plays an important role in the recognition of targets.
• as a rule, the background to the target is not uniform; there is disturbing infor-
mation present.
• getting approximate values for the search window is not easily possible.
• the targets are usually portrayed at different sizes in the image.
For aerotriangulation, which is the standard procedure for the determination of the
orientation elements of metric photographs, only a small number of control points are
necessary (Sections 5.2.3 and 5.3.5); for this reason the approximate positioning and
identification are frequently undertaken by the operator and only the fine measurement
is entrusted to the computer.
Further details are to be found in Section Β 6.2, Volume 2 and in the literature: for
example Gülch, Ε.: IAPR 31(B3), pp. 279-284, 1996. Rottensteiner, F., Prinz, R.:
VGI 84(2), pp. 189-195, 1996.
Targeted control points are to be found almost exclusively in close range photogram-
metry. In aerotriangulation the control points are usually natural points, such as corners
of buildings, corners of the boundaries of agricultural land, and so on, but also larger
ensembles of topographic elements. The recognition process for such control points is
exceedingly difficult, especially in obtaining suitable reference matrices (further details
and some sources of literature are to be found in Section Β 6.3, Volume 2). Höhle 14
extracted such a reference image from the topographic database of a geographic infor-
mation system (GIS) in conjunction with an obsolete, digital orthophoto. It goes with-
out saying that relatively many erroneous identifications resulted, but the high number
of "GIS control points" in general permitted very reliable error removal.
6.8.5 Inclusion of epipolar geometry in the correlation

With the aid of epipolar lines, a two-dimensional correlation for finding corresponding
points in two relatively oriented images can be reduced to a one-dimensional corre-
lation problem. Such a one-dimensional correlation is illustrated in the stereopair of
Figure 6.8-9. Points are selected in the left-hand image, for example with an inter-
est operator (Section 6.8.1.3). The corresponding point in the right-hand image can
lie only on the particular corresponding epipolar line. The reference matrix of the
left-hand image is therefore shifted along the epipolar band in the right-hand image
until the optimal correlation is reached. The saving in computer time arising from the
one-dimensional correlation is of great significance in practical photogrammetry. Not
only is there a great saving in computer time connected with the use of epipolar lines,
however, but the reliability of the correlation is also much improved.
Figure 6.8-9: Stereopair with four points in the left image and corresponding epipolar
bands in the right image
14
Höhle, J., Potuckova, M.: PFG 6/2001, pp. 397-^04, see also Morgado, Α., Dowman, I.: ISPRS-J
52, pp. 169-182, 1997.
6.8.5.1 Epipolar geometry after relative orientation using rotations only
Epipolar geometry is illustrated in Figure 6.8-10. An epipolar line is the line of inter-
section of the plane of the photograph and an epipolar plane or basal plane. A basal
plane is a plane containing the two perspective centres 0\ and O2 and a relevant object
point P. All epipolar rays in a photograph intersect in the epipole K. The epipole is
the point of intersection of the base Ο1Ο2 and the plane of the photograph. An epipole
in one photograph is therefore the image of the other perspective centre. Thus a stere-
opair has two epipoles, the points K\ and K2 in Figure 6.8-10. Corresponding image
points lie on the two epipoles. In the "normal case" of photogrammetry, the epipoles
are imaginary points, coinciding in a single point at infinity in the direction of the base.
For this reason it is probably preferable to use the term basal plane.
Figure 6.8-10: Original images, normalized images and the epipolar rays for two cor-
responding image points, radiating from the epipoles K\ and K2
When the relevant epipolar rays have been found in both images, a one-dimensional
correlation can be carried out along these two epipolar rays. (The difficulties which
arise from the diagonal course of the epipolar rays in the matrices of the original images
should not be entered into here.) The obvious points from which to determine the
epipolar rays are the epipoles K\ and K2 (see Figure 6.8-10). The image coordinates
of Κ ι, for example, are obtained by substituting the model coordinates (6,0,0) of the
perspective centre O2 in the collinearity equations (2.1-19) for the first image (see
Equations (6.8-12)):
_ ri\b + Γ2ΐ0 + Γ3ΐ0 _ r\2b + rÔ + rÔ
ξΚι
~~ ~Crl3b + r 2 3 0 + Γ33Ο VK
' ~ ~Cn3b + r230 + r330
Or, for the general case:
(6.8-12)
Exercise 6.8-7. Derive the coordinates of the epipoles of both of the photographs of
Section 4.3.1 using the orientation angles from Section 4.3.2. Answer: The rotation
matrices R i and R2 must first be evaluated from the orientation angles:
/ 0.999615 -0.026817 -0.007148 \

R, = 0.026818 0.999640 0.000000
\ 0.007145 -0.000192 0.999974/
/ 0.999912 0.013170 -0.001502 \

R2 = -0.013200 0.999676 -0.021780
\ 0.001215 0.021798 0.999762 J
The image coordinates of the epipoles, in [mm], are hence derived as: £«·, = 21351,
ηΚχ = - 5 7 3 , ξΚι = 101613, ηΚι = 1338.
6.8.5.2 Epipolar geometry in normalized images
The epipolar geometry of normalized images is especially simple. Normalized images

are indicated in Figure 6.8-10 beneath the perspective centres. They correspond to
the "normal case" of stereoprocessing (Section 2.1.5). In Section 2.1.5 the fact was
recorded that in such an image pair no η parallaxes appear but only ξ parallaxes. This
statement can also be put in terms of epipolar geometry: in the "normal case" the
epipoles are infinitely distant so that in every instance two corresponding epipolar rays
in the two images exhibit the same η coordinate (Figure 6.8-10).
In human vision (Section 6.1.1), also, there are no η parallaxes (vertical parallaxes) but
only ξ parallaxes (horizontal parallaxes). With such an image pair a person correlates
the two images received on the retinas and from this derives the succession in depth of
visible objects. Hence, in machine vision, correlation of normalized images emulates
natural human stereoscopic vision. The ξ parallaxes in normalized images constitute
what, in computer vision, is referred to as a disparity map.
Normalized images are obtained in photogrammetry if the photographs are taken as
"normal case" images, for example with a stereometric camera (Section 3.8.2). But
normalized images can also be produced from arbitrarily configured photography. It is
a precondition that the elements of orientation are known for both images. It is sup-
posed, first of all, that relative orientation by means of rotations only has previously
been carried out (independent relative orientation, see Section 4.3). The connections
are represented in Figure 6.8-10. ξι,η\ and ξι,ηι are the image coordinates of the orig-
inal metric image, ξ[,η[ and ξ^ 7 ^ those of the equivalent normalized images. The
principal distance of the normalized images is c'. The origin of the local object co-
ordinate system, or model coordinate system, is set at 0 \ . The χ axis is taken in the
direction of the perspective centre O2. The ζ axis becomes, as it were, the camera axis
of the first normalized image. The y axis is perpendicular to the xz plane.
The mathematical relationship between the image coordinates ξ and η of one of the
original metric images and the image coordinates ξ' and η' of the equivalent normalized
image follows from the collinearity equations (2.1-19), in which
• the specific position of the object coordinate system is noted
• the (negative) principal distance c' is substituted in place of (Z — Z())
• ξο and ηο are set to zero
ri2? + riirj - ruc'

η = —c
+ T23T]' ~ i"33c'
Equations (6.8-13) may be re-arranged to give expressions for ξ' and η', corresponding
to the Equations (2.1-20):
_ + τηη - r}3c
Γ?ΐ£
ς + fi2V — 7*33 C
' (6.8-14)
, ,Γ2ΐξ + T22V - r23C
V = - c — 7 —
ί·3ΐξ + T32V - r33C
Tik . . . are elements of the three-dimensional rotation matrix R derived from

the angles of relative orientation performed using rotations only; that
is, using u)\ = 0, ψ\, κ\ for the first image and ω 2 ,φ2, «2 for the sec-
ond image. (The solution for Exercise 6.8-7 contains an example.)
The conditions for the transformation of the original digital metric images into normal-
ized digital images are set out in these equations. A new image matrix is defined in the
normalized image (Figure 6.8-11). In order that no pixels will be lost from the origi-
nal image, the principal distance c' is chosen to be somewhat larger than the original
principal distance c. Taking this into consideration we are starting from the assumption
that the pixel sizes in the original and in the normalized images are chosen to be the
same. For every one of the pixels in the normalized image, with coordinates ξ', η', c',
the required position in the original image can be found (Equations (6.8-13)). There a
grey level interpolation over the neighbouring pixels can be carried out, as explained
in Section 2.2.3. The result of this resampling are two normalized digital images, and
therefore a normalized digital stereopair.
With a normalized digital stereopair the complete stereomodel can be processed using
one-dimensional correlation. Corresponding image points in the two normalized im-
ages lie on two lines in the images with the same η coordinates (η\ — r^). Reference
matrices, or more accurately reference vectors, are best chosen in one image in the
-H-
X
Figure 6.8-11: Original digital image and digital normalized image
neighbourhood of intersection points of grey level edges with the η coordinate lines.
The larger search matrices, or more accurately search vectors, should be chosen in the
other image, shifted by the base length (at image scale) along the η coordinate. In
cases where the orientation elements still contain significant errors the search should
be made along bands of lesser or greater width lying along the epipolar lines instead of
along the epipolar lines. A small rectangle is used as the reference matrix; the search
matrix is also defined as rectangle, but one that is somewhat wider and longer.
6.8.5.3 Epipolar geometry in original, tilted metric photographs
The transformation of tilted photographs into normalized images is relatively costly.

When correlation is necessary for only a few points, the correlation should be car-
ried out in the original, tilted photographs along the epipolar rays. For this purpose a
method must just be given for finding the epipolar rays in tilted metric photographs.
The first important step, determining the epipolar points Κ ι and K 2 , has already been
explained in Section 6.8.5.1. Then a point Pi is arbitrarily chosen in the first origi-
nal image. The points Pi and Κ ι define the epipolar ray in the original image (Fig-
ure 6.8-10). Subsequently the point -Ρι(ξι,τ?ι) is transformed into the first normalized
image using Equations (6.8-14). The resulting point is f / ^ J , η [ ) . This point can be car-
ried over directly into the second normalized image as P i : ξ ' 2 = ; ή 2 = η[ . A constant
could be added to the ξ coordinate correspondingly approximately to the base in the
photograph. Finally the point Ρ 2 ( ξ 2 , V 2 , c') is transformed into the second original im-
age using Equations (6.8-13). The points P 2 and K 2 in the second original photograph
fix the epipolar ray in that photograph; it corresponds to the epipolar ray through the
points P\ and K\ in the first original photograph.

Exercise 6.8-8. Taking the image coordinates ξι and η\ in the first original image as
starting point, find the epipolar rays for the eight points of the numerical example in
Section 4.3.1. The rotation matrices of Exercise 6.8-7 should be used for this. (Check:
Ignoring accidental errors, the points with the image coordinates £2 and r/2 must lie on
the epipolar rays in the second original image.)
6.8.5.4 Derivation of normalized images using the elements of exterior

orientation
In Figure 6.8-10 it is assumed that the correlation has been preceded by a relative
orientation using rotations only. In many cases, for example with direct orientation of
the sensor using GPS and IMU (Section 4.1), this assumption does not apply. Either
the formulae used above are re-arranged in terms of the elements of exterior orientation
or, using a suitable transformation, the relationships of Figure 6.8-10 are established.
Here, the second route is taken.
Figure 6.8-12: Normalized images using the elements of exterior orientation
Two metric images, after determination of the elements of exterior orientation, are
illustrated in Figure 6.8-12. The camera axes (rays from the principal points through
the perspective centres) of both normalized images should be normal to the base Β and
parallel to each other; the principal distances are c'; these requirements determine the
angles Ki and Φ2. The Ω direction of the camera axes (the parallel camera axes of the
normalized images) can be set more or less arbitrarily. A good choice is the mean of
the angles ω\ and u>2 from the exterior orientation:
(6.8-15)
In the transition from the original images to the normalized images, the angles Κ and
Φ are to be taken into account too; they are derived from the base components (Fig-
ure 6.8-12):
Κ = arctan (6.8-16)
Bx
Β,
Φ = arctan — j = L = (6.8-17)
\jBl + Bl
A 3 χ 3 matrix representing a single rotation can be formed for each of the three an-
gles Ω, Φ and Κ (Appendix 2.1-1). The three rotation matrices R n , Κφ and R k must
be combined with the known rotation matrices from the exterior orientation; these
are, for the first photograph, the rotation matrix Ri(wi,v?i,ki) and, for the second,
R2(ωό, ψ2, κι)· In so doing, the correct order of the matrices must be observed. The
object coordinates X and the image coordinates ξ, from the original image 1 are re-
lated as follows through the rotation matrix R i of the exterior orientation (from Equa-
tion (2.1-2-5) of Appendix 2.1-2; in the present instance the translations and the scale
factor of Equation (2.1-2-5) have no significance): X = R i ^ . The image coordi-
nates in the normalized image 1 are found from the X coordinates after rotations
through angles Κ (primary), Φ (secondary) and Ω (tertiary); taking note of the correct
order of multiplication of the matrices (Appendix 2.1-1) the following result is ob-
tained: £1 = RkRxjjRaX After substitution of X = Ri£ 1 ? the following relationship
between the coordinates of original image 1 and those of the normalized image 1 is
obtained (the corresponding relationship for image 2 is added):
=
£'1 = RicR<i>RnRi£i = R i £ i £2 — RicR<i>RnR2£2 (6.8-18)
The relationship between the ξη system of the original image and the ξ'η' system of the
normalized image is given in detail in Equations (6.8-13) and (6.8-14). The elements
rik appearing in these equations are to be replaced by the elements π ^ from the rota-
tion matrix R i for image 1 and by the elements rj^k from the rotation matrix R2 for
image 2. As is evident from Equations (6.8-18) the rotation matrices R j and R2 arise
from multiplication of the relevant four rotation matrices.
Exercise 6.8-9. If the angles Φ and Κ of Figure 6.8-12 are chosen as Φ =
a r c t a n ( B z / B x ) and Κ = a r c t a n ( B y / ^ B l + BD), what is the relationship between
the original image and the normalized image? (Answer: For image 1, for example,
= R<j>RKRnRi£i·)
6.8.5.5 Epipolar geometry in images which have been oriented relatively using
projective geometry
In Section 4.3.3 acquaintance was made with an alternative method of relative orienta-
tion with its origin in projective geometry. It is suitable, above all, for photographs of
which the interior orientation is unknown. With at least eight corresponding points
in both original images the eight unknowns fa of this relative orientation can be
found. For one-dimensional correlation in such non-metric images the corresponding
epipolar rays are necessary. They have already been found: in connection with Equa-
tion (4.3-21) it was realized that, for a point in one image, there is a line in the other
image on which its corresponding point must lie. Starting from Equation (4.3-22), the
equation of the epipolar ray in the first image corresponding to a point with image
coordinates V2 in the second image is found to be:
αξί + bV\ + c = 0 (6.8-19a)
where 15
a
/ \ / 6 / n + mfn + f x A / 6 \
Μ = 6 / 2 1 + m f n + /23 from (4.3-23) hence = F % (6.8-19b)
w V & / 3 1 + mf32 + l J \ 1 J
In a similar way, using F T , the transpose of F, the epipolar ray in the second image
can be found corresponding to the point £ι, η\ in the first image.
An insight: the epipolar rays can be defined directly using the F matrix, without re-
course to the epipole. This insight can also be applied to the task of correlation in
metric images. In Equation (4.3-20) it is stated how the fundamental matrix may be
arrived at in terms of the elements of interior orientation (matrices Ci and C2, Equa-
tions (4.3-19)) and the elements of relative orientation (matrices R i and R2, Equa-
tions (4.3-9)). As a result the epipolar rays, which are important for correlation in met-
ric images, may be defined by means of the Equations (6.8-19). Although the epipoles
are not needed for this, relationships should be stated from which the coordinates of
the epipoles may be found: ξ κ ι , ηκι of K\ in the first image and ξκ2; νκ2 of K2 in the
second image:
(6.8-20)
Finding the coordinates of three-dimensional object models using the fundamental ma-
trix can be understood from the relevant literature already given at the end of Sec-
tion 4.3.3.
Exercise 6.8-10. In Exercise 4.3-6 the F matrix was derived from 8 corresponding
points in two photographs. Using that F matrix and the relationships (6.8-20) find
the epipoles. (Answer: ξΚί = 6406 mm, ηΚί = - 2 0 6 mm, ξΚι — 8456 mm, ηΚι =
62 mm.) Compare these results with those of Exercise 6.8-7. Comment: The large
difference between the two results stems, among other things, from the fact that the
15
This relationship fails if either or both of £2 and 772 are infinitely large, that is if the image point lies
at infinity. This insight has already been gained at the end of Section 4.3.3. It was established there that,
in the exact "normal case", relative orientation is not possible using Equation (4.3-22).
elements of interior orientation were used in one case and not in the other. As well,
the tilts of the photographs are very small so that in each case the coordinates of the
epipole are very unreliable.
6.8.5.6 Epipolar geometry in three images
Epipolar geometry can also be used efficiently for identification of points. Assume that
on a surface, for example the body of an automobile, many small circular object points
are available (for example a grid of projected points, Section C 2.3.1.2, Volume 2).
Photographs of the surface should be taken in such a way that every object point is
captured on at least three images. These points can be successfully found in a partic-
ular photograph with a circular reference matrix, as explained in Sections 6.8.1.1 and
6.8.1.2. The weighted centroid method has proved itself for such a task: within a win-
dow inside which the circular disk lies, the coordinates and ηΜ of the centre of the
circle are calculated as the weighted arithmetic mean of the ξ and η pixel coordinates
using the grey levels gij as weights:
ξΜ = Τ^·>!Ι'ΐ ηΜ = (6.8-21)
L,9n L· 9ij
The summations in Equations (6.8-21) run over all pixels within the chosen window. In
this calculation those pixels are removed for which the grey value lies below a suitably
chosen limit (for example all pixels with a grey level smaller than 10, if the circles
are light coloured against a dark background). The problem of positioning the window
within the neighbourhood of a particular circular area still remains open. One solution
consists of moving the window one pixel at a time left or right and up or down until
the sum of the grey values g tJ reaches a maximum. This solution leads to success,
though, only when the size of the window is chosen to be less than the separation of
neighbouring circular areas. If the circular areas are not too small and are against a
dark background the standard deviation of this method lies around ±0.05 pixel, thus a
very high accuracy, comparable to that of LSM (Section 6.8.1.2).
In the following paragraph it is assumed that the circular points in all three images
have been found. Epipolar geometry offers an interesting solution to the problem of
identifying corresponding points, illustrated in Figure 6.8-13.
A point Ρ is chosen in image 1; the corresponding point is sought in image 2. The
epipolar ray Κ\ —> 2, in image 2, cuts down the set of points to a one-dimensional
group; when there are many points, no unambiguous solution is reached, however;
three points which come into consideration are marked in Figure 6.8-13. Certainty can
be restored with the help of image 3. For this purpose epipolar rays Kt2 —> 3 are
established in image 3 corresponding to all candidate points in image 2; in the example
of Figure 6.8-13 there are three such candidate points. Finally the epipolar ray —> 3
is found in image 3, corresponding to the point Ρ chosen in image 1. The required point
Ρ in image 3 lies on an intersection of the epipolar rays Ki 2 —> 3 with the epipolar
ray Κ ι —• 3. From this the corresponding point is also found in image 2. Should the
points lie very close to each other or if the orientation elements are not very accurately
Figure 6.8-13: Epipolar geometry for the identification of the point Ρ in three images
known the inclusion of a fourth photograph in the epipolar geometry system can be
very valuable.
Exercise 6.8-11. How does the identification method explained in this section perform
if the three camera stations are arranged along a straight line? (Answer: Since the
epipoles are the images of the perspective centres of the other images, in this case two
epipoles coincide. Therefore identification of corresponding points is not possible with
this method.)
Literature: Dold, J., Maas, H.-G.: IAPR 30(5), pp. 65-70, Melbourne, 1994.
6.8.6 Automated recording of surfaces

In cases where there is an artificial pattern of points on a surface, for example points
marked by circles as was discussed in Section 6.8.5.6, the surface can be recorded
fully automatically. If the surface exhibits no such texture, another strategy must be
followed. The recording of the surface of the Earth is central to the following discus-
sion, although the same solution can also be applied to many kinds of tasks in close
range photogrammetry.
Characteristic of this solution is that, instead of finding homologous areas or points
of one image directly in the other image and then performing object reconstruction
using three-dimensional intersection, the establishment of the relationships among the
images and the reconstruction of the object are completed in a single, unified process.
Figure 6.8-14 illustrates the simplest procedure of this kind, known as the VLL relation
(vertical line locus) which starts after the exterior orientation of the two photographs
has been completed. The X Y coordinates are set in advance, preferably over a regular
grid (Figure 6.6-5). Beginning with one of the grid points a series of equally spaced
Ζ coordinates is defined along the vertical line through that point. The strategy begins
with points at particular Ζ levels being transformed into both photographs using the
collinearity equations (2.1-19). In this way candidates for homologous windows are
found in both photographs. Taking excerpts of equal size in both images, a measure of
similarity, for example the correlation coefficient (6.8-1), is computed for each window
pair. The required object point and its X Y Z coordinates have been found when the
maximum value of this similarity measure is reached.
This raster-driven correlation method gives only a point cloud and no break-lines on the
ground. These have to be obtained by an operator and, together with the ground points
obtained by correlation, brought together in a digital terrain model (DTM). Contour
lines derived in this way have the appearance portrayed in Figure 6.6-4.
The VLL procedure delivers satisfactory results only under the following conditions
(Section C 2.2.1, Volume 2):
• the image regions to be correlated must exhibit a good natural texture.
• the ground must be open (without buildings and not forested).

• the slope of the terrain may not be great (the maximum rotation of the corre-
sponding area of the image is 30° and the maximum scale difference of the cor-
responding excerpts of the images is 25%).
The VLL procedure can be refined in various respects:
• from the different correlation coefficients r and the corresponding Ζ values, a

continuous function r(Z) can be found; if the first derivative of this function is
set to zero the desired Ζ value is found with sub-pixel accuracy.
• the predetermined X Y raster points can lead to pairs of windows in the pho-
tographs in which no suitable correlation is possible (too great a slope of the
terrain or too much vegetation); at other X Y raster points where a perfectly sat-
isfactory correlation can be made the measured point may lie above the ground,
as for example on a house roof; at such points still more "intelligence" must be
built into the processing so that the correlator reacts in a similar way to a human
operator (see the end of Section 6.6.2) and bypasses the raster point or chooses a
substitute point nearby; the elimination of points on roofs can also take place in
post-processing (Section 8.1.2.2).
• in place of the horizontal square element in the object (Figure 6.8-14) an inclined
object element can be used; the slope can be estimated either from additional
geometrical parameters in the LSM (Section 6.8.1.2(e)) or from the X Y Z coor-
dinates of neighbouring points already found through correlation.
• instead of isolated elements at each raster point the surface can also be approx-
imated by continuously connected facets and in a computational procedure the
correlations and all parameters of the facet structure be derived.
The object sends out not only geometric information but also additional radiometric
information. In the case of digital photography we must consider not only geomet-
ric rays emitted from an object point, but also their electromagnetic properties which
produced the grey, or colour, values in the photographs. In the reconstruction of the
object surface the homologous geometrical rays must intersect in the object point and
the homologous grey or colour values must produce identical radiation values in object
space. To this end radiation models which take the line of sight into account should,
above all, be introduced into the processing.
Literature: Grün, Α., Baltsavias, Ε.: PE&RS 54, pp. 633-641,1988. Heipke, C.: DGK,
Reihe C, Heft 366, München, 1990 and PE&RS 58, pp. 317-323, 1992. Helava, U.:
PE&RS 54, pp. 711-714, 1988. Wrobel, B.: IAPR 27(B3), pp. 806-821, Kyoto, 1988,
and BuL 55, pp. 93-101, 1987. Zheng, Y.-J.: PE&RS 59, pp. 489-498, 1993. Nevatia,
R.: IAPR 31(B3), pp. 567-574, 1996. See also Section C 2.2.4, Volume 2.
6.8.7 Semi-automated processing for plan

Fully automatic analysis of a scene captured on two or more photographs places the
highest of demands on recognition and modelling processes. In computer vision such
complex image analyses are performed for many problems. These processing methods
from computer vision can frequently be transferred to close range photogrammetry. For
complete processing in plan from aerial photographs, on the other hand, there are still
few reliable and accurate solutions. That is the reason why photogrammetric process-
ing in plan is solved semi-automatically, the recognition part of the task rendered by
man (Section 6.6.1) and the measurement and modelling part automated. This sec-
tion is limited to determination of the boundaries of different kinds of land use, to the
transport network and so on; the independent Section 6.8.8 is dedicated to the semi-
automated processing of buildings.
6.8.7.1 Active contours (snakes)
Semi-automated processing in plan can be put into practice using computed curves
known as snakes. They are explained with reference to Figure 6.8-15: in the stere-
omodel the operator selects the forest boundary and sets the floating mark on a few
points (Section 6.5). Then he initiates an automatic measuring process which, starting
out from a coarse polygonal line, lays a curve along the forest boundary. This adapta-
tion can take place in both photographs independently of each other. Intersection with
epipolar rays produces the ξη coordinates of corresponding points in both images so
that the XYZ coordinates of points along the three-dimensional curve in the object
model can be derived16. Other strategies are conceivable which, for example, carry out
adaptation to the image curves in both photographs simultaneously, while also taking
the 3D spatial curve into account. (Details are to be found in the literature given in this
section.)
Adaptation of the coarse polygonal line to a curved grey level edge in a photograph
still needs closer consideration. The curved edge may be modelled for example by a
spline function s(t); s is vector valued,i.e. there is a spline function for each of the two
image coordinates £(£) and η(ί). The parameter t can be taken as curve length from
the starting point.
The spline function s(f) should, in our example, be smoothed as much as possible. This
geometric constraint is arrived at by minimizing the first and second derivatives, that
is by minimizing s'(t) and s"(t). The appropriate corrections for the corresponding
observation equations in the sense of a least squares estimation17 are denoted by v S '
and v s ».
16Should the particular epipolar ray make only a very shallow intersection, or even meet the curve
tangentially, the determination of the XYZ object coordinates becomes very inaccurate or, in some
cases, indeterminate. Details of the problem are to be found in the third item in the comment list of
Section 6.8.8.
17 Instead of the least squares solution used in this section the optimization of snakes can be arrived at
by calculus of variations, by dynamic programming, by level-set formulation (see, for example, Osher,
St., Paragios, N.: Geometric Level-Set Methods in Imaging, Vision and Graphics. Springer, 2003) or by
simulated annealing (van Laarhoven, P.J.M.: Simulated annealing—Theory and Applications. Reidel,
1992).
Figure 6.8-15: Three operator-placed starting points (left); interim result (middle); and
final position (right) of the curve as forest boundary
The corrections v s / and v s » are internal geometric constraints on the desired outline.
An external constraint is that the boundary line should run as close as possible to the
digitized polygon points. The corrections connected with this in a least squares esti-
mation are denoted by v e x t . (Details of the mathematical formulation are to be found
in Sections Β 3.5.4, Volume 2.)
The third group of corrections v p h o relates the curve to the image content. The LSM
scheme (Section 6.8.1.2) can be invoked. A reference matrix, in this case a reference
matrix for a forest boundary, is defined along the approximate curve. To the left of
the approximate curve a dark grey level is used and to the right of it a light grey level.
Both grey values will be taken from the mid-range of values in the given photograph.
A two-dimensional LSM formulation (6.8-7), which allows not only two translations
6ξ and bri but also permits a deformation of the reference image according to the spline
function chosen above, will improve the curve in a combined adjustment. (With longer
curves the reference image is broken into shorter pieces.)
Applying appropriately chosen weights P , the combined adjustment minimizes the
following function:
V T P v = (vJ/P s 'V s < + v]"„P s »v s ") + vJxtPextVext + VphoPphoVpho

(6.8-22)
=Eint + Eext + Epho
In the terminology of snakes, the function to be minimized is known as an energy

function with the terms E i n t for the internal energy, E e x t for the external energy and
Epho for the photometric energy.
The snakes strategy introduced here for semi-automatic photogrammetric processing
follows the example of the publication by Grün and Li, which also contains instances
of different applications and an extension to the simultaneous determination of curves

in several overlapping photographs.
This section should conclude with further indication of the flexibility of the snakes
formulation by mentioning some of the different ways in which the E p h 0 terms may be
framed. In the above formulation, different grey levels left and right of the provisional
curve were used. Frequently, however, differences in grey levels along a forest border
are not very meaningful, so that texture ought to be invoked in the place of grey levels.
The interior of a forest is highly textured whereas outside the forest there is usually
little texture. From the photographs in question a representative measure of texture
is made to the left of the provisional curve and likewise a representative measure of
texture is made to the right. Texture measures can be determined with the help of the
Förstner operator or the Moravec operator (Section 6.8.1.3).
Additional Literature: Kass, M., Witkin, Α., Terzopoulos, D.: International Journal of
Computer Vision 1(4), pp. 321-331, 1988. Kerschner, Μ.: IAPR 32(3/1), pp. 244-249,
Ohio, 1998.
6.8.7.2 Sequential processing
After the operator has chosen the particular kind of line to be used in plan and has
digitized the starting points, the snake solution runs through smoothly in one go, so
long as the operator does not intervene because he considers that the lines in the pho-
tographs look unsatisfactory. In the present section this closed solution described above
is compared with another method which consists of a number of sequential steps.
The key phases of this semi-automatic, three-dimensional processing method are as
follows:
• in each of the two photographs, independently of each other, the lines are ex-
tracted.
• then corresponding lines are matched to each other; here epipolar geometry pro-
vides very valuable assistance.
• the matching of corresponding lines is supported by means of human stereo-

observation. This gives the operator an opportunity to prescribe the meaning of
individual lines from a catalogue of object types.
• the image coordinates of homologous points come from epipolar geometry. If

the particular epipolar line just grazes, or barely cuts, the particular plan line, the
image coordinates are very unreliable; the best remedy is the inclusion of a third
metric image.
• with the image coordinates of the homologous points and the known elements
of exterior orientation of both, or in some cases more, metric images the X Y Z
coordinates of a succession of points in the 3D object model follow.
Operator involvement in the matching of corresponding lines can be very costly; by

accepting a certain loss of quality, this cost can be reduced when choosing the following
work programmes:
• capturing and modelling of the object surface, that is for example a terrain model
from aerial photographs (Section 6.8.6)
• production of a digital orthophoto from one of the two metric images (Sec-
tion 7.3)
• line extraction in the two-dimensional digital orthophoto; the result is the X Y

coordinates for the succession of points in the two-dimensional object model
• in cases where a three-dimensional object model is called for, the Ζ coordinates,

heights corresponding to the X Y coordinates of the DTM points, can be ex-
tracted by means of interpolation.
Line extraction requires still closer examination. In the section "Filtering in the spatial
domain" (Section 3.5.2.1) readers have already become acquainted with the basic ideas
of edge extraction which, in the case of processing in plan, is indicated as better than
line extraction 18 . A typical line-extracting operator is the Laplace operator (3.5-14).
Figure 6.8-16 is the result of a convolution of the right-hand matrix of Figure 6.8-1
with the Laplace operator. The lines in the original image come out as zero-crossings in
the convoluted image. Through linear interpolation between the negative and positive
pixel values along the rows and along the columns, the lines are found with subpixel
accuracy. (The achievable high accuracy can be diminished, however, because of the
shift in position connected with compression of the photographs, Section 3.6.2). In
Figure 6.8-16 the outline of a cross is obtained.
But isolated zero lines also occur; they are caused, for example, by noise in the original
grey values. In the original Figure 6.8-1 the grey value differences at the isolated
points were around 1; in Figure 6.8-16, which is formed through convolution with the
Laplace operator, they are around 5. This statement can be generalized: the Laplace
operator and comparable operators, which are based on second differences, reinforce
the grey value differences between neighbouring pixels (in our example the maximum
grey value difference in the original image is 8 ( = 9 — 1) and in the convoluted image
it is 32 ( = 8 + 24).
18
In the image-processing literature there is frequently a differentiation between edges and lines. At
an edge an abrupt transition occurs from one grey level to another grey level. Along a grey level profile
perpendicular to the edge, the position of the edge is at a maximum of the first derivative or at a zero
of the second derivative of the profile function. At a line, coming from either side, there is a relatively
more abrupt transition from one grey level value (possibly, indeed, from two different grey value levels)
to one lying in the middle of a common grey value level. The band of pixels with the common grey value
level, the actual line, is very narrow. If one imagines a grey value profile perpendicular to the line, the
position of the line is at a zero of the first derivative of the profile function. An operator which extracts
edges gives the envelope of a line. The thinner the line the more merged together the envelope. The edge
extraction operator can thus be used in a similar manner for extraction of lines.
ο;ο ο; ο , 1 C D 1 j ο ioii
0 11 00 00 00
0 1 00
1 ϊ Γ ϊ 1<(34
i 0 )
1C3>1 0 0 1 0 0 0 1
0 1 0 0
1 4 1 0 0 0
0 0 0 0
8 9 0 0 0
0 0 0 7 1 6 1 2 9 1 0
0 0 7 C20 - 1 7 - 1 5 -23)4 1
0 0 1 7 1 6 1 6 8 1 0
0 1(7)1 8 8 0 0 0
0 0 1 0 0 8 0. 0 0 0
Figure 6.8-16: Convolution of the right-hand matrix of Figure 6.8-1 with the Laplace
operator (3.5-14) (the convoluted matrix is smaller than the input matrix by two
columns and two rows)
The troublesome consequences of noise in line extraction with the Laplace operator
can as a result be removed to a large extent if the original image is first subjected to
a low-pass filter (for example a Gauss filter (Sections 3.5.2.1 and 3.5.2.2)) before the
Laplace operator.
Both of these processes are combined in the LoG operator, the Laplacian of Gaussian.
The LoG operator filters out the noise (a low-pass filter) and simultaneously provides
an image in which the zero crossings represent the lines. This operator is obtained
when, in imitation of the differentiation of the Laplace operator (Section 3.5.2.1),
the Gaussian bell curve, with which the elements of the Gauss filter are fixed, is
twice differentiated and the second derivatives in both coordinate directions—the total
differential—are combined in one function.
If one chooses the determining parameter σ of the Gaussian low-pass filtering to be the
same in both coordinate directions, that is σ — σζ = ση, and assumes independence
between the two coordinate directions, that is σξη — 0, the two-dimensional normal
distribution is obtained (Figure 6.8-17)19:
r2 = ξ2 + η2 (6.8-23)
Differentiating with respect to ξ and with respect to η:
(6.8-24a)
Differentiating again with respect to ξ and with respect to η:
(6.8-24b)
19
For example: Mikhail, E.: Observation and Least Squares. IEP-A Dun-Donnelley Publisher, New
York, 1976.
LoG(r)
vJ
Figure 6.8-17: Gaussian bell

curve for the two-dimensional Figure 6.8-18: Two-dimensional
normal distribution LoG function
As in the derivation (3.5-13) of the Laplace operator from the second differences
(3.5-12), here the second derivatives are added and are designated the LoG function:
e + ri2~2a2
LoG:= (V2H^,7?)) = ~—ζ6 e 2σ (6.8-25)
2πσ
The LoG function has the shape of an inverted sombrero (Figure 6.8-18). The one and
only parameter that is free to be chosen is σ. A small σ results in many zero points. A
large σ suppresses the pseudo-lines, which are caused above all by the noise; it is more
or less the case that only the relevant lines in the original image are extracted. The LoG
convolution operator can be derived from the LoG function (6.8-25). For σ = y/2, for
example, the following convolution matrix is obtained; it is limited to a 9 χ 9 matrix;
the values lying outside this 9 x 9 matrix, which are small, are not shown:
/Ο 1 2 4 4 4 2 1 0\
1 3 7 10 10 10 7 3 1
27 11 6 0 6 11 7 2
4 10 6 - 2 4 - 4 6 - 2 4 6 10 4
LoG = 10~ 3 χ 4 10 0 - 4 6 - 8 0 - 4 6 0 10 4 (6.8-26)
4 10 6 - 2 4 - 4 6 - 2 4 6 10 4
27 11 6 0 6 11 7 2
1 3 7 10 10 10 7 3 1
yO 1 2 4 4 4 2 1 0 /
The advantage of the LoG operator over the Laplace operator for the extraction of lines
is demonstrated with the aid of a practical example. Figure 6.8-19 shows the original
Figure 6.8-19: Original image
Figure 6.8-20: With lines ex- Figure 6.8-21: With lines ex-
tracted using the Laplace operator tracted using the LoG operator
image. Figure 6.8-20 reproduces the image after convolution with the Laplace operator
(3.5-14); Figure 6.8-21 is the result of a convolution with the LoG operator (6.8-26).
The advantage of the LoG operator is clearly recognized; the many pseudo-lines are
suppressed and the relevant lines are clearly emphasized.
Before leaving this section the problem of the shift in position of lines extracted with
the LoG operator should be touched on. Since the LoG operator is an edge-extracting
operator, the sum of its elements should be zero (Section 3.5.2.1). Checking this for
the matrix (6.8-26) though gives a sum of —0.008. This defect leads to a shift of the
extracted lines that must be taken into account when high accuracy is demanded. The
shift is dependent both on the deficiency in the edge operator and on the contrast at the
particular edge (Section C 2.2.2,Volume 2).
Incidentally, similar shifts also appear on application of the Förstner operator to edge
extraction (Section C 2.2.2,Volume 2). The Förstner operator was introduced in Sec-
tion 6.8.1.3 as the interest operator for finding well correlated image points. With this
operator points in the image are found at which a high contrast is present and a good
distribution of grey level edges in all directions.
For the extraction of lines the Förstner operator must be modified; patches with high
contrast are indeed sought but they should also have a very limited distribution of edges
in all directions, in the ideal case only a line. The Förstner operator modified in this
way delivers points and the directions of the lines at those points. From such a sequence
of points and associated tangents a complex curve may be produced.
Exercise 6.8-12. Develop the LoG operator for σ = 2. Result: The central element
is -0.0199; the element which is distant two pixels from the centre is -0.0037; the
element a distance of 2y/2 from the central point is zero; and so on.
Exercise 6.8-13. Establish the one-dimensional LoG operator. Take, as starting point,
the one-dimensional normal distribution, = 1/(σ y/ϊϋ) e x p ( - £ 2 / ( 2 σ 2 )). σ is the
abscissa of the inflection point in the Gaussian bell curve and also the abscissa of the
zero point in the LoG operator. Result: LoG = (ξ 2 - σ2)/(σ5\/2π) exp(—ξ 2 /(2σ 2 )).
Literature: Schenk, T„ Li, J.-C., Toth, C.: PE&RS 57, pp. 1057-1064, 1991. Schenk,
Τ.: Digital Photogrammetry. TerraScience, 1999. Couloigner, I., Ranchin, T.: PE&RS
66, pp. 867-874, 2000. Baumgartner, Α., Hinz, S., Wiedemann, C.: IAPR 34(3B),
pp. 28-31, 2002.
6.8.8 Semi-automatic measurement of buildings

The semi-automatic measurement of buildings usually begins with the operator making
topology-supported measurement, as discussed in Section 6.6.3. The corner points of
the buildings are only roughly set; the fine measurement, which is the subject of the
following passages, is taken over by the computer.
The starting situation for the automatic fine measurement is portrayed in Figure 6.8-22.
The buildings have already been reconstructed with moderate accuracy. Using the
known elements of interior and exterior orientation, these relatively inaccurate building
edges are transformed into the relevant metric images. Figure 6.8-22 is limited to the
roof ridge in two photographs.
Before this transformation as illustrated in Figure 6.8-22, line extraction was carried
out in the photographs involved (Section 6.8.7.2). The points used in extracting the
lines are illustrated in Figure 6.8-23, as well as the transformed straight lines of the
roof ridge, accentuated in Figure 6.8-22. In both of the photographs an ε frame (i.e. a
tolerance band) has been laid around the transformed (approximate) roof ridge. Subse-
quently all points used in extracting the lines inside the particular ε frame are chosen
and an improvement of the straight-line roof ridge is brought about in the object model.
Two linearized observation equations can be set up for each point of the extracted line
in each of the two metric photographs (in image 1 they are the points 1, 2, 3, 4 and, in
image 2, the points 5, 6, 7, 8, 9). These equations correspond to Equations (4.1-3) for
Figure 6.8-22: The approximate building model, transformed into and inserted in, two
photographs from which line extraction is planned. The area surrounding the roof ridge
has been highlighted (after Rottensteiner, F.: Geow. Mitt, der TU Wien, vol. 56, 2001).
Figure 6.8-23: The approximated roof ridge in both metric images and the image
points found during line extraction inside an ε frame
three-dimensional intersection but, since there are no homologous points as in three-

dimensional intersection, for each image point there are three unknown object coordi-
nates Χ, Υ, Z. In the example of Figure 6.8-23 this means that, as against 18 ( = 9 χ 2)
observation equations, there are 27 ( = 9 χ 3) unknown X Y Z coordinates.
In order to make the adjustment equations soluble and to achieve straightness in the
roof ridge, additional condition equations must be introduced into the least squares
estimation. In order that three points with indices i,j,k should lie on one straight line,
the following two conditions should be met:

y V" y
In the X Z coordinate plane: —!- —
- i = —^ - i (6.8-27)
•^fc Zi Zj — Zi
In the Y Z coordinate plane: = ^ ~ ^ (6.8-28)

Zk — Ζι Ζj — Zi
A third, corresponding, equation in the X Y coordinate plane could indeed be formu-
lated but it is redundant, for it could be obtained by division of Equations (6.8-27) and
(6.8-28). The next point, and each further point, provides two condition equations of
the form of (6.8-27) and (6.8-28). Since 9 points occur in this example, there are in total
14 ( = (9 — 2) χ 2) condition equations, so that for the determination of the 27 unknown
X Y Z coordinates mentioned above a total of 32 ( = 18 + 14) equations is available.
A discussion of this type of estimation problem, adjustment with indirect observations
and condition equations, can be found in textbooks 20 on least squares estimation.
This adjustment results in a set of points, 1 - 9 , with their X Y Z coordinates, all lying
along the straight line of the 3D roof ridge. The observed image coordinates ξ and
η in the photographs receive "corrections", the sum of the squares of which has been
minimized.
This section on the refinement of the solution for the roof ridge should conclude with
some comments:
• the conditions corresponding to the Equations (6.8-27) and (6.8-28) should be es-
tablished for those pairs of coordinate planes in which the coordinate differences
are greatest.
• in order to avoid numerical problems in the least squares estimation, the separa-
tions of the points along the extracted lines should not be very small. In every
condition equation the first and last points of a line should be included; points to
be replaced should always be confined to interior points.
• if the roof ridge runs parallel to the camera base the bundles of rays from both
photographs lie in the same plane; therefore the intersection conditions for points
along the 3D line in the object will be unsatisfactory. If the roof ridge portrayed
in the photographs runs along the epipolar rays in both images, a similar critical
situation can also arise. The best remedy is the inclusion of a third image. The
use of more than two photographs also has an additional advantage; in cases
where part of a building is hidden in one photograph that problem, too, can be
overcome.
• likewise, with robust estimation and data-snooping (Sections Β 7.2.1.4 and
Β 7.2.1.5, Volume 2) wrongly matched image points can be reliably eliminated.
• analysis of the "corrections" can also lead to the realization that the roof ridge is
not a straight line but should be approximated by a 3D polygonal path or a 3D
20
E.g. Mikhail, E.: Observations and Least Squares. IEP-A Dun-Donnelley Publisher, New York,
1976.
curve. In this case, the 3D straight line should be replaced with a 3D polygonal
path or curve.
• the use of image pyramids from the photographs (Section 3.6.1), which can be
generated with low-pass filtering, is recommended for the automatic refinement
process. The coarsest pyramid is used to start with; extraction at this level finds
only prominent lines, so there are no disturbing points from lines of minor im-
portance. When passing to finer levels the ε frame is progressively reduced; to a
large extent, errors in identifying corresponding points are excluded in this way.
• other a priori knowledge besides straightness of lines is brought into the refine-
ment, for example the horizontality of roof ridges in new buildings. In this case
the condition equation
Zj -Zi= 0 (6.8-29)
is additionally introduced in the least squares estimation. The second and further
points each give one such condition equation. Thus for our example from Fig-
ure 6.8-23, eight further condition equations (6.8-29) appear; this means that 40
equations are available for the determination of the 27 unknowns.
• instead of the X Y Z point sequence along the adjusted 3D straight line, the para-
meters of the straight line can be used directly in the automatic refinement. Such
an expression has few unknowns; its formulation is didactically more demand-
ing.
• automatic refinement should not be applied for each 3D line in isolation, but
simultaneously for all edges of a whole building (Figure 6.6-6 shows an example)
or at least for a part of a building which is indicated as a building primitive 21 . The
advantages of a simultaneous treatment are obvious; especially that extensive a
priori knowledge can be elegantly introduced in the form of numerous kinds of
conditions (building walls perpendicular to each other, roof planes with equal
inclination, and so on). The example shown in Figures 6.6-7 and 6.6-8 resulted
from this kind of complex automatic refinement.
Additional Literature on semi-automatic processing of buildings: Gülch, E., Müller,

Η.: in Baltsavias, E., Grün, Α., van Gool, L. (Eds.): Automatic extraction of man-
made objects from aerial and space images (III), pp. 103-114, Balkema Publishers,
Lisse, 2001. Grün, Α., Wang, X.: ibid., pp. 93-101, 2001. Zhou, G„ Li, D.: PE&RS
67,pp. 107-116, 2001.
6.8.9 Accuracy and reliability of results obtained by automated or

semi-automated means
There is as yet no long tradition of semi-automated and automated measurement pro-
cedures in photogrammetry. It is not surprising, therefore, that there are no rules of
thumb for their accuracy.
21
Rottensteiner, F.: Geow. Mitt., vol. 56, 2001.
Certainly semi-automated and automated measurement methods are displacing human

operators more and more. These modern methods of processing, however, will be
accepted only if they achieve at least the same quality of the work as for that carried
out by human operators, which was reported in Section 6.7.
For project planning, therefore, the rules of thumb given in Section 6.7 can be used,
independently of whether the work is done semi-automatically, automatically or by a
human operator. But it is necessary to see to it that the prerequisites for good semi-
automated and automated measurement procedures are present, for example relatively
high resolution in digitizing of the photographs, correlation algorithms which give sub-
pixel results and so on.
From interim published accuracies achieved in specialist projects it can be concluded

that semi-automated and automated procedures have a higher potential accuracy than
that of work done by human operators. When it comes to single point measurement,
it may indeed be the case that the work of human operators is more accurate; the
much higher number of points and lines which are amassed in the automated proce-
dures compensate for the accuracy deficit in single-point measurement. Usually over-
compensation is made.
The reliability of results obtained automatically can be guaranteed only for simple
tasks. If the nature of the task is complex, as in capturing topographic information from
aerial photographs, a visual check should always be planned, such as by superimposing
the results on the photographs.
6.8.10 Special features of the three-line camera

The use of images from a three-line camera, or three-line scanner (Section 3.7.2.3),
involves some special features which will be touched on in this section. Since a three-
line camera is always flown with GPS and IMU support, direct georeferencing (Sec-
tion 4.1.2) can be performed. It remains an unanswered question though, how auto-
matic localization of corresponding points in the three image strips takes place.
Good conditions for automatic localization occur because the images are first normal-
ized using information from the GPS/IMU, in order to remove large distortions arising
because of the dynamics of the flight. This step is comparable with the production of
normalized images from arbitrarily configured images from conventional frame cam-
eras (Section 6.8.5.2). The mathematical relationships need not be repeated here. The
procedure corresponds to an indirect transformation of the images from the three-line
camera into an adopted horizontal plane.
Corresponding points in images from a normalized strip have almost identical η coor-
dinates. On the contrary the ξ coordinates of corresponding points differ very greatly.
On the one hand the differences come from the way in which the rows a and b in the
camera are staggered (Figure 3.3-3, right) and on the other hand from the height differ-
ences of the ground. Normalized strip images have characteristics comparable to those
of normalized images from stereophotogrammetry with conventional metric cameras
(Section 6.8.5.2): correlation can proceed along narrow bands which exhibit the same
η coordinates.
Correlation to establish corresponding points in the normalized strip image-triples for
automated and semi-automated procedures (Section 6.8) can be carried out with three-
image correlation (Section 6.8.1.5). This three-image correlation is a significant advan-
tage of the three-line camera in comparison with the conventional aerial photographs
with 60% forward overlap where, over large areas, correlation can be performed only
with two-image correlation.
After the normalization of the image-triple using the GPS/IMU information, it is un-
usual for the processing for height, plan and buildings to begin immediately; normally
the GPS/IMU orientation will first be improved. For this purpose many tie points on
the upper and the lower edges and in the middle of the image-triple are automatically
located and an integrated sensor orientation (Section 5.5) carried out.
After this improvement of the orientation there are no longer any residual η parallaxes
between the images of the strip-triple. The images can be observed in digital work-
stations (Section 6.5) and the heighting, measurement in plan and recording of the
buildings can be carried out by an operator, with computer assistance (Section 6.6).
The operator has the possibility of choosing any particular pair of strips and observing
their image-triples stereoscopically.
This choice between normal angle, wide angle and intermediate angle (Section 3.7.2.3)
depends on the desired accuracy and the visibility on the ground. Work-stations that are
set up for the processing of images from a three-line camera allow a rapid switchover
between particular pairs of images of a triple. Such workstations take advantage of the
fact that in flight direction there are no stereomodel boundaries.
Literature additional to that given at the end of Section 5.5: Grün, Α., Li, Ζ.: PFG
2003(2), pp. 85-98. Sandau, R. et al.: IAPRS 33(B1), pp. 258-265, Amsterdam, 2000.
Chapter 7
Orthophotos and single image analysis
As is well known, a single image is not sufficient for reconstructing a spatial object.
However, methods have been developed which only use one image. Such single image
analyses are conditional on knowing the geometric shape of an object; for example that
the situation is dealing with an object plane or that a digital surface model is available.
Single image analysis frequently takes place in two stages. In the first stage an analogue
or digital photograph, which is normally distorted, is converted into a geometrically
correct photograph or orthophoto. In a second stage this orthophoto is evaluated by
analogue, analytical or digital means.
In many cases, by adding graphical elements, such orthophotos are converted to or-
thophoto maps, in particular those based on aerial photographs, and known simply as
photomaps. In undeveloped regions, such photomaps are significantly more advanta-
geous than line maps. Archaeologists, soil scientists, foresters, agriculturalists, geog-
raphers, geologists, planners and ecologists often do not find, in a line map, the details
import to them. Maps which provide the full content of analogue or digital photographs
are a better solution in this case. Even in developed areas, however, there is a demand
for the image content in the photos. Orthophoto maps are also attractive because they
are significantly faster and cheaper to produce than line maps.
Increasingly, digital orthophotos are employed as data in topographical information

systems. To a considerable extent, orthophotos bring the real world into a Geographic
Information System (GIS), in which the topographic information system is a subsys-
tem. The digital orthophoto is therefore an excellent orientation aid for the GIS user.
(This opinion was expressed in Section 6.6.3 for photogrammetric vector analysis.) A
GIS user coming from one of the related disciplines mentioned above can personally
undertake photo interpretation on the basis of the digital orthophotos and digitize the
results in vector mode. In this way these disciplines can build up their own thematic
data base.
Digital orthophotos play a large role in the visualization of both developed and unde-
veloped regions. For this purpose the orthophoto pixels are projected onto the digital
terrain model and viewed obliquely. These three-dimensional visualizations can also
be presented as animations (Section 7.6).
However, the orthophoto as texture source reaches its limits with large-scale visual-
izations, especially for urban landscapes. Here the orthophotos from aerial surveys
contain no photo texture for the vertical fagades, etc. For consistent three-dimensional
Section 7.1 Perspective distortion in a metric image 367
visualizations the orthophoto must be developed further into a photo model. This
is obtained by transferring the texture in a photograph onto the surface of a three-
dimensional object model, often a CAD vector model (Section 7.5).
Chapter 7 is mainly concerned with the production of digital orthophotos and their
accuracy. It will be seen that professionally created, digital orthophotos are a very
accurate source of data. Before describing the production process, the distortions in
the metric image (central projection) compared with a map (orthogonal projection) are
discussed.
7.1 Perspective distortion in a metric image

Distortions in mappings from one surface to another are described in differential geom-
etry by means of the Tissot indicatrix 1 . This also includes perspective images. This
theory can be applied, however, only if the surface of the object can be described by
smooth analytical functions. The surfaces of the ground and other objects can generally
not be so described. More often the ground surface is represented in photogrammetry
by digitizing a large set of individual points on the surface. These are then approxi-
mated by small (flat) surface elements which at their joins have no discontinuities in
height but discontinuities in slope. (In geometry such surfaces are characterized by the
C°-continuity.) A different method, based on the XFZ-coordinates of closely spaced,
discrete points, is therefore required to describe the distortions in the metric photo-
graph.
Figure 7.1-1 shows an object surface defined by the Z-coordinates of a square X Y grid.
In the sense of a map (orthogonal projection) the square X Y grid is the distortion-
free image of the ground surface. The distortions in the photograph are derived by
computing the points in the photograph corresponding to every corner of the grid on
the ground by means of the collinearity equations (2.1-19). The result is sketched in
Figure 7.1-2.
The deformation of the grid gives a very clear impression of the deformations in an
aerial photograph (Figure 7.1-3).
Numerical results for the deformation at individual points (e.g. point 1 of the square
1,2,3,4 in Figure 7.1-2) can be derived from the image coordinates ξ and η and the
ground coordinates X and Y of the point, as follows:
(a) length distortions in X and Y directions:
in which
Sik = (ξί - tk)2 + {Vi - Vk)2
'Snyder, J.P.: Flattening the Earth: Two Thousand Years of Map Projections. Chicago University
Press, 1993.
368 Chapter 7 Orthophotos and single image analysis
Figure 7.1-1: Relation between the square grid AX — AY in the XY plane and the
corresponding deformed grid in the image plane
(b) distortion ω$ο of the right angle at point 1:
ω 90 = 90° - Z2'1'3' (7.1-2)
(c) extreme length distortions Λι and Λ2 in the directions of maximum deformation,

from the two equations (see the theory of projection of two surfaces in any ap-
propriate textbook):
A[ + A2 — A^ + Ay (7.1-3)
Αΐλ2 = \ x X y COS CJ90 (7.1-4)
3 4
1 2
ξ X
Figure 7.1-2: Grid in the photograph (central projection) and in the map (orthogonal
projection)
Figure 7.1-3: Deformed grid in the aerial photograph and the square grid in the or-
thophoto (courtesy of the Mapping Division of the Austrian Federal Office of Metro-
logy and Surveying (BEV))
(d) maximum angular deformation 2ω:

λι — λ2 , .
sin ω = (7.1-5)
λ\ + λ2
(e) deformation of area:

ΔΙ'2'3'
Αι A2 = X x X y cos 0,190 — (7.1-6)
Δ123
Note: since every grid point is the corner not only of one grid square, but of four
squares (in general), the deformations at each grid point according to Equations (7.1-1)
to (7.1-6) can be computed four times. The four values will not be exactly the same
because, on the one hand, the ground surface is, in general, discontinuous at each
grid point and, on the other hand, the initial Equations (7.1-1) are derived from finite
differences and not from infinitely small differentials. The four values, which have
only minor differences, should be averaged at each grid point.
Exercise 7.1-1. Assume the elements of interior and exterior orientation of the exam-
ple at the end of Section 2.1.3 and compute from three neighbouring points Pi, P2 and
P3 the various deformations from Equations (7.1-1) to (7.1-6), referred to a photo map
at a scale of 1 : 10000. Solution:
X[m] y [ m ] Ζ Μ
1 363400 61500 575.12
2 363410 61500 576.87
3 363400 61510 578.04
Photo coordinates:
ξ [mm] η [mm]
1' -33.694 92.706
Τ - 3 3 . 7 7 3 93.893
3' - 3 4 . 8 3 5 92.859
Length distortions: λχ = 1.1896 Ay = 1.1512

Right angle distortion: ω 90 = 90° - Z ( 2 T 3 ' ) = 11°27'
Extreme length distortions: Λι = 1.2828 λ 2 = 1.0463
Maximum angular distortion: 2ω = 11°40'
Area distortion: Λ1Λ2 = 1.3
Exercise 7.1-2. Repeat the computations under the assumption of an exactly vertical
camera axis. (Solution: Ai = 1.2760, λ 2 = 1.0431,2ω = 11°32'.)
If the object surface is a horizontal plane, the deformations of the grid in the photograph
can be seen very clearly in comparison with the X Y grid in the object (Figure 7.1-4).
In this diagram, the photograph is tilted about an axis running from its lower left to its
upper right corner. The line of maximum slope of the photograph runs from the upper
left to the lower right corner.
The relation between the square grid and the deformed grid is here defined by the
collinearity equations (2.1-19) in which Ζ is taken as the height of the object plane
(parallel to the X Y plane). Instead of the collinearity equations, the projective trans-
formation (2.1-24) can be used.
Numerical results for the deformations can also be derived by applying Equations
(7.1-1) to (7.1-6) to plane object surfaces. A mathematical formulation by the Tis-
sot indicatrix is also possible in this case. Rather than formulating such a relatively
complex solution, we give below a practical approximation from which the distor-
tions in photographs of plane objects with small angles of tilt can be derived. The
Figure 7.1-4: Grid in the photograph and in the object plane (map)
development begins with Equations (4.3-3). First, the model is transformed into the
photograph by means of Equations (4.3-24). We then set άκ — 0 (άκ produces no
deformation) and dw — άφ — v/V2, i.e. ν is a tilt about a diagonal of the photograph.
With ξ = η = ν I\f2 we obtain for the image displacement Ap — \J Αξ2 + Αη2 along
the deformed image diagonal:
P2
Ap — —v (7.1-7)
c
Ap lies in the image diagonal perpendicular to the tilt axis, whereby Ap is a displace-
ment towards the tilt axis in the "higher" corner and away from it in the "lower" corner.
The magnitude of Ap shows the extent by which a semi-diagonal p, i.e. the distance
between a point Ν (image centre) and a point Ρ (image corner), is increased or re-
duced to ρ by the image tilt v. Figure 7.1-4 shows the upper half of a tilted photograph
in which ρ is reduced to p. Equation (7.1-7) can also be derived from the diagram of
Figure 7.1-5 by simple trigonometric relations and series expansion.
Table 7.1-1 shows the image displacements Ap as a function of tilt ν for three types of
cameras.
V
1 gon 2gon 5gon lOgon
(54') (1°48') (4°30') (9°)
Normal angle (c = 30 cm) 1.4 2.7 6.9 13.8 mm
Wide angle (c = 15 cm) 2.8 5.5 13.8 27.5 mm
Super wide angle (c = 9 cm) 4.6 9.2 22.9 45.9 mm
Table 7.1-1: Image displacements Ap (in [mm]) in the corners (p = 115 χ V2) as a
function of tilt u and type of camera
Figure 7.1-5: Reduction of the image length in an exactly vertical photograph to ρ in

a photograph tilted by ν (Ν is the nadir point and υ the nadir distance.)
The increase or reduction of a small length dp follows from differentiation of Equa-

tion (7.1-7):
d(Ap) = 2^dp (7.1-8)
Numerical Example. A line 10 mm in length along the diagonal in the "upper" corner
of a wide-angle photograph (c = 15 cm) tilted by 5 gon has been reduced by:
d(Ap) — 2(115\/2/150)(5 π/200)10 = 1.7mm

In an exactly vertical photograph the line would therefore be 11.7 mm long.
Exercise 7.1-3. How long would this line be in an exactly vertical photograph if it
measured 10 mm in the "lower" corner of the tilted photograph? Take the remaining
data from the example above. (Solution: 8.3 mm)
Exercise 7.1-4. Calculate the reduction (increase) of length of the 10 mm length if it
lies parallel to the ν axis in the "upper" ("lower") corner of the photograph. Take the
remaining data from the example above. (Solution: 10.8 mm (9.2 mm) in an exactly
vertical photograph.)
Exercise 7.1-5. Derive Equation (7.1-7) from the diagram of Figure 7.1-5 and show
also the neglected second-order terms. (Solution: pv1 (tan2 τ - 1), in which τ is the
angle shown in Figure 7.1-5.)
Exercise 7.1-6. Use Equations (7.1-1) to (7.1-6) to compute the image deformations
of the example of Exercise 7.1-1 under the assumption that the three points 1,2 and 3
are at a height of 500 m, i.e. the object plane lies at Ζ = 500 m.
Section 7.2 Orthophotos of plane objects 373
Solution: Object and image coordinates:
Point F[m] Z[m] ξ [mm] η [mm]

1 363400 61500 500 -32.014 88.104
2 363410 61500 500 -32.050 89.123
3 363400 61510 500 -33.030 88.070
Extreme length distortions: Λι = 1.0200 λ 2 = 1.0162

Maximum angle distortion: ,βω — 0.22gon(12')
Area distortion: Λ1Λ2 = 1.04
Exercise 7.1-7. Repeat the computation under the assumption that AX 12 = ΔΥ13 =
5 m and compare these results with those for AX 12 = ΔΥ13 = 50 m. Solution with
AXxl = AYn = 5m:
λι = 1.0208 λ 2 = 1.0164 2ω = 0.28gon(15')
Solution with AX]2 — ΔΥ13 = 50 m:
λι = 1.0202 λ 2 = 1.0169 2ω = 0 . 2 1 g o n ( l l ' )
What is the source of the differences? (Answer: In the initial Equations (7.1-1) finite
differences are used instead of differentials.)
7.2 Orthophotos of plane objects

Orthophoto production is significantly dependent on whether or not the camera axis is
perpendicular to the object plane.
7.2.1 With vertical camera axis

As already explained in Section 2.1.4, the photograph is the same as an orthophoto if
the image plane and the object plane are parallel at exposure (see Figure 2.1-11 for an
example). The square grid in the object plane is then imaged as a square grid in the
plane of the photograph.
If an (unavoidably) small tilt error ν occurs when pointing the camera axis perpendic-
ularly to the object plane or when arranging the image and object plane to be parallel,
then Equations (7.1-7) and (7.1-8) can be used to estimate the effect of this error on the
orthophoto.
Numerical Example. In producing the orthophoto shown by Figure 2.1-11, assume

that the orthogonal alignment of the camera axis to the plane of the fa$ade was possible
only to an accuracy of u — 0.5 gon(27'). The orthophoto, which has a scale of 1 : 320,
then has the following maximum error:
Error along the 100 mm long semi-diagonal (7.1-7):
1002 0.5
= 0 5 m m
157^5 63^6
Error in a 5 mm wide window in the corner of the orthophoto (7.1-8):
This simple procedure for creating photographs free of deformation is used mainly in
architectural photogrammetry. Users must be aware, however, that the images of points
outside the plane of the object will be displaced in the direction of the principal point
H. The magnitude of this radial displacement Ap of such objects from their bases (e.g.
houses in aerial photographs) can be derived from Figure 7.2-1:
Ap = AR^- = AZ—— (7.2-1)

Z0 cmB
Table 7.2-la shows the maximum image displacements for typical scales in terrestrial
photogrammetry for the Wild P31 camera (Table 3.8-2), while Table 7.2-lb shows
those for typical scales and cameras in aerial photogrammetry.
Exercise 7.2-1. Repeat the computation of the Exercise 2.1-7 with the aid of Equa-
tion (7.2-1).
Figure 7.2-1: Radial image displacement Ap of points outside an assumed object plane
(exactly vertical photograph)
AZ
mB 1cm 5 cm 10 cm 25 cm 100 cm
Normal angle (p — 83 mm, c= 202 mm)
25 0.16 0.8 1.6 4.1 16.5 mm
100 0.04 0.2 0.4 1.0 4.1 mm
250 0.02 0.1 0.2 0.4 1.6 mm
500 0.01 0.04 0.1 0.2 0.8mm
Wide angle (p = 82 mm, c = 100 mm)
25 0.3 1.6 3.3 8.2 32.5 mm
100 0.1 0.4 0.8 2.1 8.2 mm
250 0.03 0.2 0.3 0.8 3.3 mm
500 0.012 0.1 0.2 0.4 1.6 mm
Super-wide angle (ρ = 75 mm, c — 45 mm)
25 0.7 3.3 6.7 16.6 66.5 mm
100 0.2 0.9 1.7 4.1 16.6 mm
250 0.1 0.4 0.7 1.6 6.6 mm
500 0.03 0.2 0.3 0.8 3.3 mm
(a) Typical scales for terrestrial photogram-
metry (Wild P31)
AZ
mB 1 m 5m 10m 25m 100m 250m
Normal angle (23 cm χ 23 cm, c = 30 cm)
5000 0.1 0.5 1.1 2.7 10.8 27.1mm
10000 0.05 0.3 0.5 1.4 5.4 13.6 mm
25000 0.02 0.1 0.2 0.5 2.2 5.4 mm
50000 0.01 0.1 0.1 0.3 1.1 2.7mm
Wide angle (23 cm χ 23 cm, c = 15 cm)
5000 0.2 1.1 2.2 5.4 21.7 54.2 mm
10000 0.1 0.5 1.1 2.7 10.8 27.1mm
25000 0.04 0.2 0.4 1.1 4.4 10.8 mm
50000 0.02 0.1 0.2 0.5 2.2 5.4mm
Super-wide angle (23 cm χ 23 cm, c — 9 cm)
5000 0.4 1.8 3.6 9.0 36.1 90.4mm
10000 0.2 0.9 1.8 4.5 18.1 45.2mm
25000 0.1 0.4 0.7 1.8 7.2 18.1mm
50000 0.03 0.2 0.3 0.9 3.6 9.0 mm
(b) Typical scales for aerial photogrammetry
(p = 115>/2mm)
Table 7.2-1: Radial image displacements Ap [mm] in the corners as a function of the
image scale 1 : TUB, the distance AZ of the points from the assumed object plane and
the type of camera
7.2.2 With tilted camera axis

The production of orthophotos from tilted photographs, where the camera axis is di-
rected obliquely onto the object plane, is widely practised. This orthophoto production
is also known as rectification. In close range work the photographs are taken on site
without any special precautions and in a more or less arbitrary alignment, perhaps also
handheld. Metric cameras are now less often used. Increasingly, digital semi-metric
cameras and digital amateur cameras are employed (Sections 3.8.4 and 3.8.7).
Control points are necessary to rectify tilted images. Where calibrated cameras are
in use, only three control points are strictly necessary for the spatial resection of the
metric image (Section 4.2.1); to increase reliability and avoid extrapolation, however,
it is advisable to have at least four control points positioned in the corners of the im-
age. Where uncalibrated cameras are used, at least four control points are required for
every image in order to determine the eight parameters of the projective transformation
(2.1 -24). For large blocks of images, the numerous control points required for either
the central perspective or projective solution can be generated by means of a bundle
block adjustment with known or unknown interior orientation (Section 5.3) 2 .
The rectification of the distorted images is done either on the basis of a central perspec-
tive transformation (for metric images) or a projective transformation (for non-metric
images) (Section 2.1.4). The technique of digital projective rectification 3 has already
been presented in Section 2.2.3. Here the indirect rectification (transformation of ob-
ject plane to image plane) is preferred to direct rectification (transformation of image
plane to object plane). This section also explains grey level interpolation as part of the
required resampling. Figure 2.2-8 shows a practical example.
When rectifying tilted images of flat objects, the dominant source of error is the depar-
ture of the actual object surface from the assumed reference plane. Figure 7.2-2 shows
the radial displacements Ap in the tilted plane of the photograph and the corresponding
displacements Ar in the orthophoto resulting from the perspective rectification of the
tilted photograph. The displacements Ar in the orthophoto are radial from the imaged
nadir point 4 N°, which is the pole of the zenith or the vanishing point of all verticals,
in the case of aerial photographs.
Thus, if we assume a horizontal ground plane, the positions of buildings, trees etc. in
a rectified orthophoto will only be correct at ground level. Roofs, tree crowns, etc.
are displaced from their true positions. Similarly, in rectified photographs of building
facades details in front of or behind the plane of the fa9ade will be displaced radially
from the "nadir point".
2
AS a-priori knowledge, parallel and/or orthogonal straight lines, as well as much other information
about an object, can be used in place of control points to make the rectification. There are many solutions
on this topic. See Hartley, R.I., Zisserman, Α.: Multiple View Geometry. Cambridge University Press,
2001, and also Section Β 4.7.2, Volume 2.
3
Analogue rectification, which has been practised for many decades and which is based on re-
photographing the original images, no longer plays a significant role. It is therefore not discussed in
this edition; the methods and associated instrumentation are extensively presented in earlier editions.
4
In the general case, as opposed to that of aerial photographs, the term "nadir point" is used to mean
the foot of the perpendicular from the perspective centre to the reference plane.
Figure 7.2-2: Radial image displacements Ar in the orthophoto of points outside the
reference plane of the perspective rectification
The magnitude of the displacement, by analogy with Equation (7.2-1), is:
Ar = A Z ^ r (7.2-2)
ZQ
r ... radial distance of the point from N°, which is fixed in the orthophoto by
the coordinates Xo and lo of the imaging station Ο
For near vertical photographs, the most common imaging configuration in single image
photogrammetry, both formulas can be used in the same way, i.e. Equations (7.2-1) and
(7.2-2). In the first formula, Ap and ρ refer to the photo plane, in the second formula
Ar and r refer to the orthophoto which is normally an enlargement. The numerical
values in Table 7.2-1 are therefore also valid for orthophotos derived from near vertical
images. It is necessary only to take account of the enlargement factor between image
and orthophoto. This factor varies in practice between 3 and 6.
Numerical Example. An orthophoto has been generated from a normal-angle aerial

photograph. This has a photo scale of 1 : 25000 and has been digitized at a reso-
lution of 25 μτη. The orthophoto's pixel equivalent in the ground coordinate system
is 50 cm. Deviations from the assumed horizontal ground plane have a maximum of
10 m. Table 7.2-1 indicates a corresponding error of 0.2 mm in the original image. In
the 5-times enlarged orthophoto, i.e. at a scale of 1 : 5000, the corresponding deviation
is 1 mm, which is 5 m in the ground coordinate system. This error becomes particularly
graphic when expressed in pixel units: viewed on a computer screen this error amounts
to 5/0.5 = 10 pixel!
This example and Table 7.2-1, in particular, show the limits of single-image pho-
togrammetry. The upper part of the table indicates the limits for architectural surveys
and similar applications of terrestrial photogrammetry. The lower part shows the lim-
its for exposure configurations typical in aerial surveying. The table also expresses
the fact that normal-angle photographs are better suited to orthophoto production than
wide-angle photographs.
Exercise 7.2-2. A digital orthophoto is to be produced at a scale of 1 : 2500. The

orthophoto format is 50 cm χ 50 cm and an error of 1 mm can be tolerated. What is
the maximum permissible offset of ground points from the horizontal reference plane
used for the rectification? Assume near vertical photography at an image scale of
1 : 7500. There is a choice of cameras with principal distances 15 cm, 21cm and
30cm. (Solution: AZ is 3.2 m, 4.5 m and 6.4 m respectively)
Exercise 7.2-3. How large can the maximum offsets be if the image scale changes
from 1 : 7500 to 1 : 15000? (Solution: ΔΖ = 6.4 m, 9.0 m and 12.8 m respectively.)
Orthophoto production on the basis of a flat reference surface fails as soon as the
ground has small hills and depressions. For objects with curved surfaces it is appropri-
ate to apply differential rectification, as discussed in Section 7.3.
7.2.3 Combined projective and affine rectification

The problem discussed in this section is illustrated by Figure 7.2-3. An xy object
plane is tilted with respect to the X Y plane of the reference coordinate system and
an orthophoto is required in this XY plane. The xy object plane and the XY plane
of the reference coordinate system are affine with respect to one another. The general
affine transformation is described by Equations (2.1-8). The mathematical relationship
between the ξη image plane and the xy object plane can be expressed as a projective
transformation defined by Equation (2.1-24). Orthophoto production can therefore be
achieved in the following steps:
• projective rectification of the ξη image into the xy plane. Result: xy image.
• affine rectification of the xy image into the XY plane. Result: required or-
thophoto.
Both steps can, however, be combined into a single step since the ξη image plane
and the X Y orthophoto plane are also related by a projective transformation 5 . The
transformation parameters can be determined with the aid of at least 4 control points,
i.e. ξη and XY coordinates are known for at least 4 points.
The combined projective and affine rectification is of considerable practical impor-

tance: in large-scale orthophotos of developed areas, the flat (tilted) roofs should be
rectified in this way. (For completeness it is mentioned here that the principle of or-
thophoto production which takes account of buildings, as illustrated in Figure 7.3-2,
also leads to orthophotos of roofscapes which are free of error.)
5
The theory of projective transformation states that two consecutive projective transformations can
be replaced by a single (third) projective transformation. The affine transformation is also a projective
transformation in which the perspective centre lies at infinity (see Exercise 2.1-8 and its solution). The
proof of this statement is readily given: combine Equations (2.1-8) and (2.1-24), ensuring first that the
coordinate axes are identified as in Figure 7.2-3. The result is an 8-parameter projective transformation
between the ξη image plane and the X Y orthophoto plane. A more elegant derivation of this relationship
is possible using homogeneous coordinates (Appendix 2.2-1).
7.3 Orthophotos of curved objects

7.3.1 Production principle
The production of orthophotos for objects with curved surfaces is based on the theory
of perspective distortion in a metric image which was discussed in Section 7.1. To
produce a digital orthophoto of a curved object surface, a digital model of this surface
is required. For simplicity an object will be assumed which has a square grid in the X Y
plane (AX = AY) and for every grid point the Ζ coordinate is known. (The acquisition
of such models has been discussed in Sections 6.6.2 and 6.8.6.)
Orthophoto production of curved object surfaces is implemented in an indirect way, i.e.
an initially empty image matrix is created in the orthophoto plane and corresponding
matrix elements found in the reference image. This indirect transformation will be
discussed in more detail with reference to Figure 7.1-1 and under the assumption of
standard imaging conditions appropriate to aerial photogrammetry:
• first, an image matrix is defined in the X Y plane, the orthophoto plane. In or-
thophoto production from aerial photographs, the pixel spacing of the image ma-
trix is usually significantly finer than the grid spacing of the terrain model.
• in the second step, the Ζ coordinate for every pixel in the orthophoto matrix must
be determined by interpolating in the ground model. The surface of a grid square
is commonly assumed to be a hyperbolic paraboloid (also called an HP surface).
Section 2.1.3a explains in detail how the Ζ coordinate can be found at any X Y
position within a grid square. The result of this interpolation process gives the Ζ
coordinates for all X Y orthophoto pixels.
• in the third step the image coordinates corresponding to the X Y Z locations in

the orthophoto grid are calculated using the collinearity equations (2.1-19) and
the known elements of interior and exterior orientation.
• in the final step the grey value from the reference image corresponding to the ξη
coordinate pair is assigned to the corresponding X Y position in the orthophoto
matrix. Since the ξη coordinate pair will not, in general, lie at the centre of a
pixel in the reference image, a grey value interpolation is required, as outlined in
Section 2.2.3a.
This process for producing orthophotos is relatively expensive. Approximate solutions,

which have an acceptable loss of accuracy, are therefore in demand. The widely used
anchor point method is a solution of this sort. The steps outlined above, which gener-
ate the ξη coordinate pairs in the reference image corresponding to the XY coordinate
pairs in the orthophoto, are only implemented for the grid points in the ground model,
as shown in Figure 7.1-1. The results for a square mesh are shown in Figure 7.3-1:
the ξη coordinates and the XY coordinates for four points are known. (In Figure 7.3-1
the origin of the object coordinate system has been shifted to the lower left corner;
the reduced object coordinates are indicated by X and Y.) For the pixels in the or-
thophoto matrix, which are shown in the left half of Figure 7.3-1, the corresponding ξη
Section 7.3 Orthophotos of curved objects 381
Y
A
-•X
Figure 7.3-1: Bilinear transformation of a square mesh using four anchor points
coordinates can be obtained by means of a bilinear transformation:
ξ = δοι + bnX_ + b2iY + hiXY

(7.3-1)
η = bw + bi2X + b22Y + 632^
The 8 coefficients b.^ are determined from the 4 anchor points. (For each coordinate
axis, the problem corresponds to interpolation on a hyperbolic paraboloid. This is
discussed in detail in Section 2.2.3a. Exercise 3.2-2 provides the solution for a bilin-
ear transformation of fiducial marks.) The matching of corresponding points in the
orthophoto and reference image by means of Equations (7.3-1) is appreciably faster
than matching by means of collinearity equations, as described above. The economic
advantage is even greater with larger grid spacings. However, this leads to a loss of
accuracy which will be discussed in Section 7.3.2d.
Digital orthophoto production of objects with curved surfaces 6 can also be extended to
ground definitions based on both grid points and break lines. This generates orthopho-
tos which are particularly accurate on the break lines. Cadastral and land use borders
etc. often run along break lines and so such orthophotos are often in demand. Or-
thophoto production which takes account of break lines and object edges is described
in detail in Section C 1.1.2, Volume 2. The problem of Earth curvature, which must
be taken into account when producing small-scale orthophotos, is also discussed in
Volume 2, Section C 1.2.
Discussion of digital orthophoto production which can also handle buildings and other
man-made constructions is included in Volume 2, Section C 1.1.3. Because of its
In the past various devices have been constructed for analogue orthophoto production of curved
surfaces. Due to their cumbersome operation, analogue orthophoto production achieved little importance.
Before digital orthophoto production has become standard, analytical orthophoto production was widely
applied since 1975. As in the anchor point method, the grid was transformed into the reference image
and the image content inside individual quadrilaterals transferred by means of optical elements (zoom
and dove prisms) onto the orthophoto plane which was a film surface wrapped on a drum. The entire
transfer process was under computer control. The most widely used instrument was the Avioplan OR1
from Wild. This instrument and its principle of operation are described in earlier editions.
practical importance, this topic will also be discussed here, with the aid of Figure 7.3-2.
Consider using the ground model (Figure 7.3-2, left) for orthophoto production. This
model, of course, only describes the ground and not the buildings. In the resulting
orthophoto, for example, a point Q on a building appears at position Q0 and not at
the required position Qo· In an accurate orthophoto ground point Ρ is expected at
position Q 0 . However, if a surface model is used which also contains the buildings
(Figure 7.3-2, right), then an orthophoto is obtained in which point Q on the building
appears at the required location QO using image point Q'. If there is no visibility
analysis then image point Q' appears a second time at position PQ = Q0: there is then
double imaging in the occluded areas.
Figure 7.3-2: Digital orthophoto production without (left) and with (right) modelling
of buildings
Figures 7.3-3 to 7.3-5 illustrate the various orthophoto products. Figure 7.3-3 shows
a conventional orthophoto produced with a ground model. In the tall buildings in the
upper half of the image it is particularly noticeable that their roofs are appreciably
displaced and the vertical fa5ades are also, incorrectly, imaged.
Figure 7.3-4 shows an orthophoto created using a surface model which also includes
the man-made features. However, a visibility analysis was not undertaken and double
imaging occurs. The double imaging can easily be seen in the buildings above the
river; one of the two doubly imaged areas is geometrically correct for an orthophoto,
the other is incorrect, as in Figure 7.3-3.
Figure 7.3-5 shows an orthophoto produced with the same surface model used to pro-
duce the orthophoto in 7.3-4. Here, however, a visibility analysis was undertaken which
eliminates the false image areas in the occlusions and replaces them with dark grey
values. The missing pixels can be transferred from a second photographs taken from a
different location. (The publication from which Figures 7.3-3 to 7.3-5 were taken (see
corresponding footnote) contains such an orthophoto).
Figure 7.3-3: Orthophoto generated using a ground model 7
Figure 7.3-4: Orthophoto generated using a surface model which also contains build-
ings and man-made structures; double imaging occurs here due to the missing visibility
analysis 7
7
Taken from Arahar, F., Jansa, J., Ries, C.: IAPRS 1998.
When producing orthophotos from a block of original images there can be undesirable
step changes in grey level and colour at the joins between the content of different im-
ages. These jumps are mainly the result of different viewing directions due to different
imaging locations. These grey level and colour changes can be eliminated as part of
the mosaicking process.
(Further information: h t t p : / / g i . l e i c a - g e o s y s t e m s . c o m . Kerschner, Μ.:
ISPRS-J 56, pp. 53-64, 2001. Shiren, Y„ Li, L„ Peng, G.: PE&RS 55(1), pp. 49-53,
1989.)
Exercise 7.3-1. Given four points with ξη and XY coordinates and the XY coordi-
nates of a fifth point:
ξ [mm] η [mm] X[m] y[m]

1 26.19 27.14 2325.00 4525.00
2 28.87 26.21 2330.00 4525.00
3 26.08 29.32 2325.00 4530.00
4 28.96 29.01 2330.00 4530.00
5 2327.50 4526.00
Determine the eight coefficients of the bilinear transformation and the ξη coordinates
of point 5. (Solution: ξ5 = 27.528 mm, η5 = 27.175 mm.)
Exercise 7.3-2. Consider the reversal of the transformation and transform point 5 from
the ξη system back into the X Y system. (Because this involves a quadratic equation
there are two solutions; the correct solution lies inside the quadrilateral, the false solu-
tion outside.)
Additional further reading on digital orthophoto production: Baltsavias, E., Käser, Ch.:
IAPRS 32(4), pp. 42-51, 1998. Schickler, W„ Thorpe, Α.: IAPRS 32(4), pp. 527-532,
1998. Weidner, U.: OEEPE Publication 37, pp. 307-314, 1999.
7.3.2 Orthophoto accuracy

The accuracy of differential rectification depends on the following sources of error:
(a) errors in the source material (residual errors of optical distortion, film deforma-
tion, positioning errors of the film scanner, etc.)
(b) errors in the elements of interior and exterior orientation (also control point errors
if the elements of exterior orientation are determined by spatial resection (Sec-
tion 4.2.1))
(c) image displacements of objects not included in the surface model (for example,
roof ridges, tree crowns above the digital terrain model (DTMs) or erosion chan-
nels, embankment edges, if a DTM is used which is based on a square grid)
(d) interpolation errors in the anchor point method, due to linear interpolation be-
tween grid points rather than using the cross ratio
(e) height errors at the grid points used in the DTMs, which can arise during data
acquisition and/or DTM interpolation between arbitrarily distributed points
(f) errors in approximating curved object surfaces by a grid model in which the mesh
elements are defined by hyperbolic paraboloids or tilted triangles (Figure 7.1-1
shows a mesh with facets defined by hyperbolic paraboloids, Figure C 1.1-6,
Volume 2, a mesh with additional triangular elements)
The dominant sources of error cause positional errors in the orthophoto which are close
to zero at the principal point of the photograph (more exactly, the nadir point N°),
increase with distance from this point and are predominantly radial with respect to
the principal point. (Figure 7.3-6 is a practical example from an accuracy study8.) The
largest errors therefore lie at the edges of the orthophoto, a fact which enables the entire
process of production to be checked by comparing neighbouring orthophotos side by
side along their common boundary.
Figure 7.3-6: Planimetrie errors in a differentially rectified orthophoto, derived from

known coordinates of signalized check points
The following detailed comments can be made on some of the sources of error men-
tioned above.
With regard to c) Image displacements of objects not included in the surface
model
Details not contained in the surface model used to create the orthophoto will be incor-
rectly imaged. When using a DTM for orthophoto production, roof ridges, tree crowns,
etc. are displaced in the orthophoto. If the DTM also does not contain break lines, then
8
Otepka, G., Duschanek, E.: Geow. Mitt., vol. 13, pp. 125-150, 1978.
embankment edges, etc. are incorrectly reproduced. Therefore the only correctly po-
sitioned points and lines in the orthophoto are the ones which are also defined in the
corresponding surface model.
Figure 7.3-7 shows the situation for a tree which is not included in the DTM. The
displacement of the tip of the tree in the (tilted) metric image is indicated by Αρ. Fig-
ure 7.3-7 also shows the corresponding displacement Ap' in an imaginary vertical pho-
tograph. The intersection of line Ρ with the (sloping) ground defines a point PQ, given
by the displacement AR in the XY object coordinate system or the displacement Ar in
the orthophoto. The displacements are derived from the following ratios:
AR _ (J_
AZ - AR tan ä c
which can be re-arranged to give:
AR = (7.3-2)
— + tan a
Ρ
If the ground slopes inwards (towards the nadir), as in Figure 7.3-7, angle ά is positive;
if the ground slopes outwards then ä is negative.
The angle ä appearing in Equation (7.3-2) is the slope angle of the ground at point
P° in the direction of line P°N° (Figure 7.3-7). Since it is more convenient to use
the maximum slope angle (denoted by a), the horizontal angle β between the line of
maximum slope and line P°N° must be taken into account in Equation (7.3-2) (lower
diagram, Figure 7.3-7) 9 :
AZ AZ
AR = -g = —r (7.3-3)
— + tan a cos β +tanacos/3
Ρ rmo
It is easy to see that the cosine of angle β is required: for β = 0, i.e. the line of
>
maximum slope is in the direction of line P°N°, the original Equation (7.3-2) is
obtained, for β — 100 gon (90°), i.e. the contour line lies in the direction of ray P ° N ° ,
the tangent term in Equation (7.3-2) is removed by multiplication with cos β = 0. If
angle β is taken over a full circle, then the differentiation between positive and negative
slope angles is unnecessary.
In the second part of the Equation (7.3-3) the ratio cjp' has been replaced by the ratio
H / ( r m o ) . These quantities have the following meaning:
h ...relative flying height above point PQ. It is obtained by subtracting the

ground height, available from the DTM, at location P 0 ° in the orthophoto,
from the Z 0 coordinate of the perspective centre of the metric image,
r . . . planimetric distance of point P 0 ° from nadir point N°, known from coor-
dinates Χο,Υο of the perspective centre of the metric image.
mo . . . scale factor of the orthophoto
'Similar formulae have been given by Blachut, T. and van Wijk, M.: Ph.Eng. 36, pp. 365-374, 1970.
Plan
(in orthophoto) No
Figure 7.3-7: Radial image displacement of details in an orthophoto which were ig-
nored in the data capture
For coarse estimations of the image displacements to be expected in the orthophoto,

the first part of Equation (7.3-3) can be used, with the distance of the corresponding
details (tree, etc) from the principal point of the (near vertical) metric image used as
the value for p'. The displacement Ar in the orthophoto is derived by multiplying AR
(Equation (7.3-3)) by the orthophoto scale 1 : m o ·
Exercise 7.3-3. A vertical photograph (c = 15 cm), taken from an absolute height of
Zo — 2000 m, has been differentially rectified to create an orthophoto at a scale of
1 : 2500. For two 10 m high buildings, estimate the displacement of the top of the
gable end in both the aerial photograph and the orthophoto. Both buildings are 10 cm
away from the principal point in the aerial photograph; one building is at a height of
Ζ = 150 m on an outward-facing slope (a — 10 gon (9°), β — 150 gon (135°)) and the
other at a height of Ζ = 800 m on an inward facing slope {a = 15 gon (13°30'), β =
75 gon (67°30')). (Solution: Ap = 0.54 mm, Ar = 2.9 mm,Ap = 0.83 mm, Ar =
2.5 mm.)
With regard to d) Interpolation errors in the anchor point method

The anchor point method uses linear interpolation along the lines between grid points
and, if there are also triangular facets in the ground model (Section C 1.1.2, Volume 2),
along the sides of the triangles. Along the corresponding lines in image and object,
interpolation must not be linear but must use the cross ratio (Section 2.1.5).
The interpolation error δ in the reference image, due to the anchor point method, is
greatest in the middle of a line on the object, with a length identified by the variable s.
It has the following value (for derivation see the second or third edition of this book):
s2
δ= — t a n a (7.3-4)
4c
tan a . . . slope of this line

s . . . grid spacing or triangle side length of the DTM, given at the scale of
the metric image
Numerical Example. Using the anchor point method, an orthophoto is to be derived
from a wide-angle aerial photograph (c = 150 mm), taken from a height of 3 km above
the ground. The available DTM has a grid spacing of AX = Ay = 50 m. The maxi-
mum slope of the ground is 50%.
The grid spacing in the reference image, indicated by s in Equation (7.3-4), is given
by: s = 50 χ 150/3000 = 2.5 mm.
2 52
(7.3-4)
v : δ = „ ' 0 . 5 = 0.005 mm = 5 μια
' 4 χ 150
In view of the fact that the photographs used to create the orthophotos have a pixel size
between 10 and 30 μπι (Section 3.4.1), this error can be ignored.
Practical note: The numerical example shows that even under extreme conditions the
interpolation error arising in the anchor point method lies significantly within a pixel
and is therefore of little relevance.
Exercise 7.3-4. Repeat the numerical example for a normal-angle camera (c =

300 mm). The flying height is now 6 km instead of 3 km. (Solution: 2.6 μτη. Com-
ment: The interpolation error is smaller in normal-angle images than in wide-angle
images.)
With regard to e) Height errors at the grid points used in the DTMs
These errors primarily have their source in accuracy losses during data acquisition
for the DTM. According to Equation (7.3-3), the AZ errors cause positional errors
in the X Y coordinate plane and orthophoto, which has been derived as indicated in
Figure 7.3-7. The point Ρ shown there can also be viewed as a grid point with height
error. Three times the average error given in Section 6.7.2.3 can be used for AZ.
Exercise 7.3-5. Photographs for data acquisition and orthophoto production at a scale
of 1 : 30000, principal distance c = 150 mm. At the least favourable position in the
corner of the image, the ground has an inward slope of 10%; here the contour lines run
parallel to the edge of the photograph. Orthophoto scale 1 : m 0 — 1 : 10000. Data
acquisition takes place statically in a grid pattern.
Section 6.7.2.3: AZ = 3 χ 0.00015 χ 4500m = 2.0m
Δ Λ = 1 75Π1
(7.3-3): =Ϊ507Ϊ40^ΤΤ07 · °Γ = 1.750/10000 = 0.18 mm
Exercise 7.3-6. The maximum error in the orthophoto must again be determined. The
difference in this case is that normal-angle photographs (c = 300 mm) are used in
place of the wide-angle photographs for the data acquisition and orthophoto produc-
tion. Photo scale is again 1 : 30000. Data acquisition by static raster measurement (see
Section 6.6.2) is now better by a factor of two:
Section 6.7.2.3: AZ = 3 χ 0.00015 χ 9000m = 4.0 m
Orthophoto accuracy, however, changes little:
(73 3): L81:m

' ^ 300/140+^0.1 χ 0.7 = °Γ ^ = = 0.18mm
Result: If data acquisition and orthophoto production are done using the same image
source, then orthophoto accuracy is almost independent of the camera type. The lower
height accuracy provided by normal-angle images is compensated for by the fact that
in normal-angle images the height error has a smaller effect on planimetric error in the
orthophoto than in wide-angle images.
However, if the data acquisition is made with a wide-angle image and the orthophoto
produced from a normal-angle image, then there is an increase in accuracy almost
in proportion to the principal distances of the images used for data acquisition and
orthophoto production, as the following example shows:
Exercise 7.3-7. Data acquisition is defined by the situation described in Exercise 7.3-5;
orthophoto production, however, is done with normal-angle imagery at a scale of
1 : 30000 (c = 300 mm). Planimetrie error in the orthophoto at a scale of 1 : 10000 is

now only:
(7.3-3): AR = 300/140 +°αι χ Q? = 0.90 m or Ar = 900/10000 = 0.09 mm
This large improvement in accuracy suggests that orthophotos could even be produced
at a scale of 1 : 5000 from normal-angle photographs at 1 : 30000. It should, however,
be noted that the enlargement factor from the reference image to the orthophoto would
change from 1 : 3 to 1 : 6 and this would result in a loss of geometric resolution in the
orthophoto. This deterioration can be avoided if the normal-angle images are taken at
a very large scale (e.g. 1 : 15000 for orthophotos at 1 : 5000). The accuracy of these
1 : 5000 orthophotos is indicated in the following numerical example:
Exercise 7.3-8. Again data acquisition is done under the conditions described in Ex-
ercise 7.3-5: m B = 30000, c = 150mm, tana = 0.1, β = 50gon(45°), AZ = 2.0m.
This information is used to derive orthophotos at a scale of 1 : 5000 from imagery at a
scale of 1 : 15000 (c = 300 mm). Orthophoto accuracy is given by:
(7 3 3): Δ Λ
· " = 300/140+^0.1 χ 0.7 = ° - 9 ° m ° Γ Δγ
= 900
/500° = ° 1 8 π ™
It should not be ignored, however, that the accuracy in relation to the orthophoto pixel,
normally chosen to be 50 cm in Exercise 7.3-7 and 25 cm in Exercise 7.3-8, decreases
from 1.8 pixels (= 0.90/0.50) to 3.6 pixels (= 0.90/0.25).
With regard to f) Errors in approximating curved object surfaces by a grid model
The maximum errors will occur where the greatest curvature occurs in the ground. In
order to estimate this, we must know the detail of at least one short profile, along the
line of maximum slope, in the roughest area of the ground. Figure 7.3-8 shows such a
profile.
Figure 7.3-8: Ground profile with minimum radius of curvature

If this profile is digitized with a grid interval AT and intermediate points then interpo-
lated linearly, the maximum height error AZ can be estimated from Figure 7.3-9. The
the maximum separation, AP, between the curve and the chord is:
AT2
AP =
8-Rcos2 a
The following is then obtained for AZ:
AP AT2
AZ= = — 3- (7.3-5)
cos α 8it cos 3 a
Figure 7.3-9: Height error AZ as a function of the radius R and grid interval AT
Numerical Example. In Figure 7.3-8 the radius R = 15 m and the angle a = 35 gon
(31.5°). The maximum height error AZ after digitizing and linear interpolation in a
grid interval AT = 10 m in the orthophoto is then, from Equation (7.3-5):
102
AZ = -———ίτττΓττ3 = 1.34m
8 χ 15 χ 0.853
This error in height produces an error in position AR which can also be calculated
from Equation (7.3-3). If we assume that the image of the roughest part of the ground
is 10 cm from the centre of the wide-angle photograph and that the maximum slope is
directed outwards (β = 200 gon (180°)), then:
1 34
A i ? = =
300/100-0.613 °-56m
This error in position of 0.56 m can be expected at the given position if the orthophoto
is produced using the anchor point method and a grid model with 10 m grid spacing. In
a 1 : 5000 orthophoto this error amounts to 0.11 mm; expressed in orthophoto pixels,
which are normally set to a size of 25 cm in 1 : 5000 orthophotos, the error amounts to
2.2 Pixel ( = 0.56/0.25).
As can be seen from Equation (7.3-3), the error caused by approximating the ground
with a grid model, primarily decreases linearly with increasing principal distance. Fur-
ther, as can be seen from Equation (7.3-5), this error is strongly dependent on grid spac-
ing AT. One way to reduce the error is to interpolate a very dense grid of heights from
the acquired data. If a sophisticated interpolation method is employed, then the spacing
of this grid can be used as the value of AT in estimating the error; however, if surface
interpolation is done using a simple method such as triangular plane between individ-
ual measurement points, then the separation of the original measurements should be
taken as the value for AT. If data acquisition is by digitizing contour lines, then the
value of AT, which is here essentially independent of the interpolation method, should
be the contour line separation.
The problem of the accuracy of digital height models is studied further in Tempfli, K.:
ITC-Journal 1980(3), pp. 478-510, and Frederiksen, P.: IAPRS XXIII(4), pp. 284-
293, 1980. and treated by spectral analysis (already briefly discussed in the context
of filtering in the frequency domain, Section 3.5.2.2). The height errors predicted by
this method serve then as data for Equation (7.3-3). Readers should also consult further
references 10 , in which typical standard deviations for complete orthophotos are derived
from empirical accuracy studies. These figures naturally include all error influences so
that individual influences cannot be isolated. Such information is particularly useful
in practice. The root mean square error of the complete orthophoto should not exceed
±0.2 mm for small-scale orthophotos or ±0.4 mm for large-scale orthophotos. With
respect to ground resolution, i.e. to the orthophoto pixel, the standard error should not
exceed 3 pixels. (Image displacements of objects not incorporated in the ground model
require separate consideration.)
Exercise 7.3-9. Draw a new test profile in which ground heights in comparison with
Figure 7.3-8 are raised by a factor of 2. (In the context of a spectral analysis, amplitudes
increase by a factor of 2.) Repeat the accuracy estimation in this case. (Solution:
Ä » 10m, α « 30gon (27°), AZ = 1.77m, AR = 0.74m.)
Exercise 7.3-10. Draw a new test profile in which ground heights in comparison with
Figure 7.3-8 are rougher by a factor of 2, i.e. the Τ axis is compressed by a factor of 2.
(In the context of a spectral analysis the wavelengths are reduced by a factor of 2.) Re-
peat the accuracy estimation in this case. (Solution: i? « 5 m, α « 31 gon (28°), AZ =
3.6 m , A R = 1.5 m.)
Further reading on orthophotography: Ecker, R., Kalliany, R., Otepka, G.: in Fritsch
and Hobbie (Eds.): Photogrammetric Week '93, pp. 143-156, Wichmann Verlag, 1993.
l0
See the publication referenced by Figure 7.3-6 as well as Blachut, T., van Wijk, M.: IAPRS XIII
(Commission II/4), Helsinki, 1976. Ducher, G.: Test on orthophoto and stereo-orthophoto accuracy.
OEEPE-Publication No. 25, Frankfurt, 1991.
Section 7.4 Analogue, analytical and digital single image analysis 393
7.4 Analogue, analytical and digital single image analysis

Single image analysis is often done on the basis of an orthophoto and this will be
discussed in the following Section 7.4.1. In the subsequent sections methods will be
described which go directly from an original image to a three-dimensional reconstruc-
tion of the photographed object.
7.4.1 Analogue, analytical and digital orthophoto analysis

In many cases an orthophoto is an appropriate medium to employ at the interface be-
tween a photogrammetrist and a partner from a neighbouring discipline. Analogue
analysis requires little in the way of instrumentation: a transparent film is laid on top
of the orthophoto and the feature interpretations are drawn upon it; distances and areas
can also be obtained from the orthophoto.
In analytical analysis the orthophoto is placed on a 2D digitizing table (2D digitizer);
with the aid of a cursor the X Y coordinates of identified object points are measured.
If required, post-processing enables the Ζ coordinates to be obtained from the digital
terrain model (DTM) by interpolation using the X Y positions.
Digital orthophoto analysis takes place in a GIS or CAD environment. The digital
orthophoto is stored on a computer. Using the cursor on the display screen, the X Y
coordinates of identified object points are obtained. The corresponding Ζ coordinates
can be directly taken from the DTM which can also be accessed through the computer.
In place of manual digitizing, the analysis of digital orthophotos also permits the appli-
cation of the automated and semi-automated methods of digital image processing (e.g.
the semi-automatic processing for plan, Section 6.8.7).
Planimetrie analysis of orthophotos, whether analogue, analytical or digital, can never
be more accurate than the orthophotos themselves. Their accuracy has been described
in detail in Section 7.3.2. It will be mentioned again here that object elements, which
do not form part of the digital surface model, can be significantly displaced in the
orthophoto. Typical examples are roof ridges and tree tops (see Equation (7.3-3)).
The third dimension cannot be derived from the orthophotos alone. As already indi-
cated, the third coordinate can be obtained from the DTM. An interesting alternative is
the preparation of stereopartners for the orthophotos (details available in Section C 1.5,
Volume 2). This combination of orthophoto and stereopartner, known as a stereo-
orthophoto, allows the orthophoto to be viewed stereoscopically. Stereoscopic viewing
of the orthophoto makes photo interpretation and analysis more comfortable, and also
more exact than monoscopic viewing (Section C 1.5.3, Volume 2).
7.4.2 Analytical and digital analysis of a tilted image of a flat object

An analytical rectification starts either with measurement of image coordinates by at-
taching the analogue photograph to a 2D digitizer, or by locating points in an original
digital image displayed on a computer screen.
Four control points are required if the photo to be rectified does not have known interior
and exterior orientations. From the image and object coordinates, the eight parameters
of a projective transformation can be determined by means of linear equations. All new
points can then be transformed subsequently into the object coordinate system. This
method has already been presented in Section 2.1.4.
If, for a given interior orientation, the exterior orientation has been determined for
example by a bundle block adjustment (Section 5.3), an analytical rectification can then
be implemented using the collinearity equations (2.1-20). If the flat object is parallel
to the X Y plane, the height of the object plane should be used as the Ζ coordinate.
Generally, this height, and also the heights of the control points, are set to zero.
If a second metric image is available instead of either control points or exterior orien-
tation elements, then the orientation elements necessary for the analytical rectification
of an image can be determined (see Footnote 2 in Section 7.2.2).
7.4.3 Analytical and digital single image analysis of curved object

surfaces
This method of analysis assumes that a digital model of the curved object surface is
available. For simplicity the discussion is limited to a model consisting of Ζ coordi-
nates on a square X Y grid. The grid should be sufficiently dense that linear interpola-
tion is permissible within a grid element. It will also be assumed that the interior and
exterior orientation of the photos to be analysed are known. Figure 7.4-1 shows the
initial configuration.
The collection of identified points in the metric image defines a spatial bundle of rays
(Figure 7.4-1 shows one of these points identified as S'). The intersection points S of
this ray bundle with the digital surface model provides the Χ, Y and Ζ coordinates in
the object coordinate system.
Section 7.4 Analogue, analytical and digital single image analysis 395
The XYZ coordinates of such an intersection point S can be found in the following
steps:
(a) determination of the XYZ coordinates of the corresponding image point S' with
the help of a spatial similarity transformation (Equation (2.1-8)) where the scale
is set to unity:
(7.4-1)
R . . . rotation matrix for the exterior orientation of the metric image
Tip: It helps to visualize Equation (7.4-1) if the unit matrix is (initially) used for
R
The next steps are explained with the help of Figure 7.4-2:
(b) intersection of the line (0)(S') in the X Y plane with the X Y grid lines
(c) calculation of the Ζ coordinates of these intersection points using the DTM along
the (vertical) profile containing the ray OS'
XY plane
I I I I I I I I II I
I I I I τ I I I II I
Y I Μ I : Μ * M l
(0) (S1)
Figure 7.4-2: Vertical plane through ray OS' and the corresponding plan view
(d) intersection of the ray OS' with the vertical profile; the result is the Ζ coordinate
of point S. Its XY coordinates are taken from the ground plan.
Exercise 7.4-1. In the numerical example of Section 2.1.3, image coordinates for both
points Ρ and Q were calculated from their object coordinates. With the aid of Equa-
tion (7.4-1), calculate the XYZ coordinates of both points in the object coordinate
system. (Solution: For Ρ : X = 362530.713 m, Y = 61215.863 m, Ζ = 2005.589 m.)
Single image analytical analysis of curved object surfaces is known as monoplotting 11 .

Monoplotting is used for the analysis of both aerial and terrestrial photographs. The
accuracy of the analysis primarily depends on the intersection angle of the ray OS' with
the surface profile in the vertical plane. When the intersecting rays are very oblique,
large X Y coordinate errors are generated where there are:
• minor height errors in the digital surface model
• small errors in the interior and exterior orientation of the image
• small measurement errors in the image coordinates of point S'
In general, therefore, monoplotting cannot replace stereoanalysis.
Exercise 7.4-2. From the opposite side of a valley, horizontal photographs are taken
with a Wild P32 (Section 3.8.3) for the purpose of monoplotting. The side to be
analysed has a slope of some 50%. How large are the X Y coordinate errors if the
digital ground model has an accuracy of around ±25 cm? (Solution: 1.5 m.)
Exercise 7.4-3. How large are the X Y Z coordinate errors if the measurement accuracy
of image coordinates is ± 10 μηι? The photographing range is 2 km. (Solution: Δ X Y =
1.7 m, ΔΖ = 0.8 m.)
Further reading for Section 7.4.3: Makarovic, B.: ITC-J, pp. 583-600, 1973.
7.5 Photo models

Orthophoto technology from aerial photographs reaches its limits in urban landscapes
or with individual objects which are to be recorded by close range photogrammetry.
The orthophoto is only an orthogonal projection onto a plane, normally the X Y plane
of the object coordinate system. The photographic texture on planes orthogonal to the
X Y plane is therefore missing in an orthophoto; where there are several overlapping
object surfaces only one can be covered with a photo texture. These disadvantages can
be overcome with a photo model. This is taken to mean a three-dimensional geometric
model (frequently a CAD vector model) whose surface is covered by a photographic
texture.
"The analysis of a digital orthophoto with a background DTM, as discussed in Section 7.4.1, may
also be referred to as "monoplotting".
Section 7.5 Photo models 397
Photogrammetric recording and modelling of three-dimensional, geometric models has

already been discussed in Sections 6.6.3 and 6.8.8 (see also Section C 1.7.1, Volume 2).
The following discussion is limited mainly to the transfer of texture from digital im-
ages onto the surface of the three-dimensional object model. Production of a photo
model will be explained using a practical example taken from a university lecture. The
modelled subject is the historic Wartberg church in Lower Austria.
Figure 7.5-1: Camera stations and directions of camera axes
Figure 7.5-1 shows the camera stations and directions of the axes. The images were
taken at different heights and zenith angles. The camera used was a calibrated Kodak
DCS 460c (Section Ε 3.5, Volume 2) with objective lenses / — 28 mm and / = 55 mm.
Pixel size is 9 μπι. 510 object points were determined using a bundle adjustment (Sec-
tion 5.3) to an accuracy of ±2 cm to ± 4 cm), with every object point, on average,
located in 5 metric images. Image point measurement was done manually in the OR-
PHEUS12 multi-image, digital monocomparator. This software tool displays multiple
images, organized in image pyramids, on a computer screen. In this system the polyg-
onal boundaries of the flat surfaces can be defined. Figure 7.5-2 shows the rendered
geometric model with artificial, oblique illumination.
To create the photo model, the polygonal borders of the flat surfaces are transformed
by means of the collinearity equations into those images which are candidates for the
transfer of photo texture. The texture is then taken from the image in which the as-
sociated flat surface is largest. The actual transfer of photo texture is done by indi-
rect transformation (Section 2.2.3), i.e. an image matrix is defined in the appropriate
(tilted) object plane and filled with the photo texture from the corresponding section
of the original image. A bilinear grey level interpolation (Section 2.2.3a) is used for
the required re-sampling. Pixel size on the object surface is chosen to be 5 cm; the
quantity of data for the entire 3D model is 485 kBytes. Figure 7.5-3 shows the 3D
photo model. It was exported in the VRML format (virtual reality modeling language)
to enable distribution over the WWW (world wide web).
12
http://www.ipf.tuwien.ac.at/products/produktinfo/orient/html_hjk/
orpheus_e.html
Figure 7.5-2: Geometric model of the church at Wartberg
Figure 7.5-3: 3D photo model of the church at Wartberg
The photo texture on the roofs of such 3D photo models is frequently taken from very
large-scale aerial photographs. In this example, an amateur camera was used to take
images from a helicopter. The images were calibrated using a bundle block adjustment
(Section 5.7). Details of the analysis can be found in the publications mentioned in the
footnote13.
Comparison of Figures 7.5-2 and 7.5-3 shows the advantages and disadvantages of a
3D photo model compared with a rendered object model using artificial surface texture:
• the main advantage of the 3D photo model is that a rich level of detail without
geometric modelling can be achieved with the aid of photo texture. Particularly
valued details are the windows and structure of the walls. (It should, however, be
13
Zischinsky, T„ Dorffner, L., Rottensteiner, F.: IAPRS XXXIII, pp. 959-965, Amsterdam, 2000.
Section 7.6 Static and dynamic visualizations 399
noted that this level of detail is only provided for visualization and is not present
in the geometric model.)
• because of the texture, which is an optical recording of the real object, 3D photo
models have a very high documentation value.
• When no photo texture is available, an artificial surface texture can be added.
• the texture of a 3D photo model has all the deficiencies associated with a pho-
tograph; complex image processing operations are normally required to create
uniform effects.
7.6 Static and dynamic visualizations

In static visualizations the object and viewing location and direction are fixed; in dy-
namic visualizations the object is fixed (the usual case here) but viewing location and/or
direction vary. A dynamic visualization is put together from a dense sequence of static
visualizations.
Static visualisations have already been used several times in this textbook, with a fo-
cus at infinity and a camera direction perpendicular to the X Y plane. Orthophotos
belong in this category. Oblique viewing directions are much more attractive, as in the
photogrammetric analysis of a village in Figure 6.6-8.
Dynamic visualization is becoming increasingly important, also known as animation.
Digital models of natural and man-made landscapes (e.g. terrain models, city mod-
els, orthophotos) provide the base data for the simulation of virtual flyovers. Digital
models of urban landscapes and historically valuable collections of buildings (e.g. city
models, 3D photo models of historic buildings, Section 7.5) provide the base data for
virtual walkthroughs. Here the flight path or walking route can be pre-defined or de-
fined interactively by the user by means of on-screen controls. A combination of this
visualization (including additional linked information) with a continuous update of the
user's location (for example with GPS) leads to the more comprehensive application of
location based services.
Chapter 8
Laser scanning
In the past ten years, laser scanning has revolutionized both topographic mapping and
close range three-dimensional object recording. Particularly in analysis, there is much
in common between laser scanning and photogrammetry. In this series of books based
on photogrammetric knowledge, it is therefore natural to deal also with laser scanning.
The sections which follow will show the parallels between photogrammetry and laser
scanning, as well as highlight the differences between both these technologies.
For purposes of terminology, and classification of laser scanning, it should be noted
that in the field of remote sensing laser scanning has, for some time, been identified
by the term Lidar (light detection and ranging). The term Laser Radar (radio detection
and ranging) has also been used in remote sensing for the location of objects with the
aid of a laser (light amplification by stimulated emission of radiation). Laser light is
strongly collimated, monochromatic and coherent. Where geometric problems take
precedence in remote sensing, it is common to speak less of remote sensing and more
of photogrammetry (Section 1.1). From that perspective also, laser scanning can be
classified as part of photogrammetry.
In the following sections we will present airborne laser scanning (Section 8.1), terres-
trial laser scanning (Section 8.2) and short range laser scanning (Section 8.3).
8.1 Airborne laser scanning

8.1.1 Principle of operation
Using a laser scanner, points on the ground are sampled (Figure 8.1-1). With the aid of
a narrow laser beam, a pulse of laser light from the scanner is diffusely reflected by a
point on the ground surface. From the elapsed time between transmission and reception
of the pulse, the distance between scanner and ground point can be determined. (The
elapsed time, multiplied by the group velocity of the light pulse which is about 0.03%
less than the velocity of light, gives double the distance value.)
The laser beam in the laser scanner is deflected at right angles to the direction of flight
and this angle of deflection is recorded. The coordinates of the laser scanner's location
and its orientation angles are required in order to convert the polar coordinates of the
measured object point into X Y Z coordinates. These constantly changing values are
determined by means of a dynamic POS (Position and Orientation System), consisting
Section 8.1 Airborne laser scanning 401
Figure 8.1-1: Scanning laser in aircraft
of GPS and an IMU. The same technology which is based on GPS and IMUs to support
aerial photography is equally applicable to airborne laser scanning. This is discussed
in detail in Sections 3.7.3.2 and 3.7.3.3. GPS, IMU and scanner operation are synchro-
nized to microsecond accuracy. Of course, the relative positions, or eccentricities, of
the sensors must be determined before the flight. Gyro-stabilized platforms are rarely
used in laser scanning. (See Section 3.7.3.3.)
Reflections can occur at different locations along the measurement beam. The upper
diagram in Figure 8.1-2 shows the arrival of the first pulse at points on the object sur-
face. The lower diagram indicates the first echoes which return and the arrival of the
pulse at lower lying object points along the beam, if these exist. This occurs if the
surface scattering back the first echo is not fully extended in the beam diameter. The
lower lying objects are then the source of the last echo. If the distance is measured
with the first echo, the recorded laser points are on the crown of the tree (second beam
from left), on the eaves of the roof (third beam from left), on a power line (fifth beam
from left), etc. If the distance is measured with the last echo, then the recorded laser
points lie, using the same order as before, on the tree trunk or ground, on the road, at
the ground below, etc. Almost all laser scanners can measure the distance with both
the first and last echoes, and record both with a small time separation. There are laser
scanners which can additionally utilize any intermediate echoes for range measure-
ment and, in this way determine a whole spectrum of distances for each laser beam.
There are limits to this recording process. Depending on the performance capabilities
of the particular range measuring device, there must be a small minimum separation
between individual records. The separation capability depends on the pulse length and
402 Chapter 8 Laser scanning
Figure 8.1-2: First echo (upper diagram) and last echo (lower diagram) from different
objects, as well as reflection and high absorption
the minimum separation can, for example, amount to 1.5 m:
300000 m χ 10 ns
1.5m = —
2
velocity of light χ shortest measurable time interval
half the measured path length
Generally, for vegetation shorter than 1.5 m high, only the top surface is recorded, even
if there is an occasional echo from the ground (first beam from left in Figure 8.1-2).
It is indicated in Figure 8.1-2 that a number of the transmitted laser pulses does not
return. For example there can be highly specular reflection from very smooth surfaces
such as a car roof. The deflected beam can later strike another surface point which
reflects diffusely and then return an echo to the laser scanner. Ranges measured in this
way via indirect routes (multi path) have gross errors; they are too long (long ranges).
At other surfaces, such as a clear stretch of water (right hand beam, Figure 8.1-2),
absorption can be so strong and reflection so little, that insufficient energy for range
measurement is returned. The absorption also depends on the angle of incidence of
the laser beam: it is especially strong at water surfaces, for example, when the laser
beam meets the surface perpendicularly. For this reason, laser scanner strips over water
surfaces often have fewer points in the middle than at the edges of the strip.
The technical data for the ALS50-II Airborne Laser Scanner from Leica Geosystems,
are summarized below.
Operating Altitude 200 m - 6 0 0 0 m above ground level

Pulse rate max. 150000 measurements per second (150 kHz)
Accuracy (incl. 10 cm GPS error, restricted to max. 40° FOV)
at 1000 m above ground:
11 cm in height, 11 cm in horizontal direction
at 4000 m above ground:
15 cm in height, 44 cm in horizontal direction
Number of returns 4 (first, second, third, last)
Number of intensities 3 (first, second, last)
Intensity digitization 8 bit intensity + 8 bit automatic gain control (AGC)
level
Maximum Field of View (FOV) 75° full angle
Roll stabilization automatic adaptive, range = 75° - current FOV
Laser divergence 0.22 mrad (45") at 1/e 2
Recording media 300 GB Harddisk (17 hours at maximum pulse rate)
Waveform profiling 8 bits at 1 ns interval at 50 kHz (option)
Table 8.1-1: Technical data for the ALS50-II Airborne Laser Scanner
The accuracies of the laser ranging component and the GPS height determination are
primarily responsible for the height accuracy indicated in Table 8.1-1 for the ALS50-
II. Neither effect is strongly dependent on flying height. The dependence of height
accuracy on flying height, which is clearly present, derives from the relatively long
path through the atmosphere, a source of disturbance, the enlargement of the laser
footprint with increasing altitude and the reduction in received energy due to spreading
loss. As the footprint becomes larger, the recorded direction of the polar coordinates
and the object point, to which the distance refers, fit less well together. The planimetric
accuracy is influenced by the accuracy of the GPS location in plan, the accuracy of the
IMU orientation and the accuracy to which the deflection angle of the laser beam is
recorded. The last two mentioned error sources effect a relatively large reduction in
accuracy with increase in flying height.
Further reading on the laser beam and its interaction with surfaces: Wagner, W., Ull-
rich, Α., Ducic, V., Melzer, T., and Studnicka, N.: Gaussian decomposition and calibra-
tion of a novel small-footprint full-waveform digitising airborne laser scanner. ISPRS
Journal of Photogrammetry and Remote Sensing 60(2), pp. 100-112, 2006.
8.1.2 Analysis and processing

8.1.2.1 Georeferencing
The orientation functions Xo{t), Yo(t), Zo(t),u(t), <p(t), κ(ί) can be determined using
the GPS and IMU information and related to corresponding polar coordinates a(t) and
s(t) through the synchronization time t. The resulting XYZ coordinates of individual
laser points are then given in the object coordinate system to which the reference station
for the differential GPS was assigned.
(8.1-1)
The actual object coordinate system is either a global GPS coordinate system (e.g.
WGS84) or the ground coordinate system. In the following text we will take the ground
coordinate system as the object coordinate system. The conversion from global coor-
dinate system to ground coordinate system, which is normally used for presentation of
final results, can be achieved by using datum transformation formulae and the appro-
priate geoid.
In Equation (8.1-1) the zero point of the a scale has been taken as the middle of the
laser scanner's field of view. This zero direction must coincide with the n(t) rotation
axis of the rotation matrix R ^ t ^ t ) ^ ) · A corresponding calibration, which takes into
account the relative positions of GPS, IMU and laser scanner in the aircraft, will not be
discussed here. However, note that companies undertaking laser scanner flights require
reference heights of a number of test surfaces (often only horizontal surfaces) in order
to effect a calibration for each project. These should be given in the ground coordinate
system we have chosen as the object coordinate system 1 .
To achieve rigorous control and improved accuracy, it is recommended to change from
this direct georeferencing to an integrated sensor orientation, i.e. using control and tie
points as is common when georeferencing images from line cameras (Section 5.5).
In place of the type of control point used in photogrammetry, with coordinates given
in the reference coordinate system, special control elements must be made available
which take account of the properties of laser scanner data. As already mentioned,
horizontal planes are suitable as height control elements. The equation of such a control
plane should be determined by a number of points measured on the ground. It has, for
example, the form:
Ziel = αο + d i X + a j Y (8.1-2)
ao, α ι, 02 · · · parameters of the plane determined from a least-squares adjustment

of the terrestrial X Y Z coordinates.
Using the GPS/IMU information and the datum transformation formulae, the direct
georeferenced coordinates X m e a s , ^meas, Zmeas are obtained from Equation (8.1-1). Us-
ing coordinates Xmeas, ^meas in Equation (8.1-2), the height Ztef corresponding to height
Zmeas is obtained. The smaller the tilt of the control plane, the less is the effect of
^meas,^meas errors on the height difference (Z meas - Zref), required for the integrated
sensor orientation.
Due to the relatively large inaccuracy in plan position of the laser points (Table 8.1-1),
and their relatively large separation which, depending on quality requirements (Sec-
tion 8.1.3) can be selected between 0.5 m and 3 m, control points cannot be employed
for the determination of horizontal shifts. Instead, linear or planar control elements
must be used. Such extended control elements could be flat roof surfaces, as shown
in Figure 8.1-3. These should be steeply sloped roofs in different orientations so that
horizontal stability is guaranteed in any direction. As Figure 8.1-3 suggests, three such
roofs intersect in a well-defined point and can therefore be regarded as a "virtual" full
control point. The equations of these control planes should be determined using terres-
trial measurements.
For the integrated sensor orientation, the -X"meas, ^meas, Zmeas coordinates of the laser
points which define the respective control planes must be determined using Equa-
tion (8.1-1) and the datum transformation formulae. The perpendicular offsets of these
points from the (known) control point plane are critical items of information for the
integrated sensor orientation. (The determination of these offsets is explained in Sec-
tion Β 3.5.5.1 in Volume 2.) In the context of a least-squares adjustment, which is still
to be discussed, these offsets have components in all 3 coordinate axes. The example
of a height adjustment plane, discussed above with reference to Equation (8.1-2), is a
special case within this 3D solution.
'Further reading on calibration: Kilian, J., Haala, N„ Englich, Μ.: IAPRS 31(B3), pp. 383-388,
Vienna, 1996. Morin, K., El-Sheimy, N.: Optical 3-D Measurement Techniques V (Eds.: Grün/Kahmen),
pp. 88-96, Wichmann-Verlag, 2001. Burman, H.: IAPRS 34 (Part 3A, Commission III), pp. 67-72, Graz,
2002.
Figure 8.1-3: Three tilted planes defined by roofs, corresponding to a "control point"
The question remains open as to how the laser points belonging to each control plane
can be automatically found. In general, the Xmeas, ^meas, Zmeas point cloud is
searched for planar regions which correspond to the control planes. Once the initial
process of direct georeferencing has been implemented, and is close to the final result,
it is possible to extract areas in the X mca s, Fmcas, -Zmeas point cloud which correspond
to the points measured by ground surveys. These can be examined for planarity, for
example using best-fitting planes, with removal of the outliers such as scanned points
on chimneys or dormers.
This concept of control planes can also be employed when connecting overlapping
laser scanner strips. The difference between a tie plane and a control plane is that the
tie plane parameters are unknown, for example the values αο, α\, az in Equation (8.1-2).
Using the method of least squares, these parameters are introduced as unknowns within
the solution for the integrated sensor orientation. A tie plane has its own set of parame-
ters, regardless of whether two, three or more laser scanner strips are involved. For each
scanner strip in the solution, the perpendicular offsets of the X mea s, imeas, •Zmeas coordi-
nates from the corresponding plane are minimized in the usual least-squares manner.
The derivation of detailed equations for integrated sensor orientation from laser scanner
data lies outside the scope of this book. Only a general outline is therefore given:
• the polar coordinate equations (8.1-1) are used in place of the collinearity equa-
tions, which are central to the integrated sensor orientation of line images (Sec-
tion 5.5). These are solved with respect to the polar coordinates and must, of
course, be linearized, etc. (Further details on the equations of polar coordinates
can be found in Section Β 3.5.3, Volume 2.)
• GPS and IMU models to handle the GPS and IMU information must be con-
structed according to Sections 5.4 and 5.5. Independent models for each scanner
strip are recommended.
• the entire concept of georeferencing using control and tie planes, presented above,
was proposed by Kager. Details of the refinement for height adjustment can be
found in a published paper 2 .
Very good agreement at the edges of the laser scanner strips can be achieved with the
methods presented here. The inclusion of control planes checks the parameters of the
data transformation formulae in those particular interest areas. An improvement in the
parameters is possible as a result. In addition, it is not necessary to take account of the
geoid in the processing, unless the areas are very large. (This problem is handled in
detail in Sections Β 5.3 and Β 5.4, Volume 2.) The IMU information can be included
without modification (Section 5.6). Only the κ heading angle must be corrected by
the convergence of the meridian (5.6-2). In this approach with ground coordinates
defining the object coordinate system, Earth curvature should be taken into account
by modifying the polar coordinates (8.1-1). A simple solution is to apply an Earth
curvature correction to the Cartesian coordinates of the scanned points in a scanner
plane perpendicular to the flight direction, using Equation (4.5-4). Note that A must be
measured from the middle of the strip.
The guidelines of GPS/IMU-supported aerotriangulation (Section 5.4) govern the num-

ber of control planes and their arrangement. As a result of the high quality of GPS and
IMU information which can now be achieved, it is sufficient to have control plane clus-
ters at both ends of the laser scanner strips in a project. Where laser scanning tasks are
designed in blocks, there should be control plane clusters, as a minimum, in the cor-
ners of the blocks. To strengthen formations of blocks, particularly if independent GPS
and IMU models are employed, then laser scanner cross strips are recommended at the
beginning and end of the blocks.
8.1.2.2 Derivation of terrain models
The construction of object models can take place after georeferencing. This section
deals with the derivation of digital terrain models (DTMs) from the digital surface
models (DSMs). Figure 6.6-10 explains the difference between these two models.
Although the scanned laser points from the last echo are used to derive the DTM, there
will still be many scanned points above the surface of the ground, depending on the
density of vegetation and buildings. Figure 8.1-7 (left) represents a DSM. It results
from a resampling (Section 2.2.3) of the irregularly distributed laser scan points on
an orthogonal 2 m grid, for example using a nearest neighbour technique. The shaded
relief of this surface grid model was based on oblique illumination from the top left
during rendering.
The starting point for deriving DTMs from scanned points is shown in Figure 8.1-4
for a single profile. In a first step an adjusting surface to the point cloud is calculated.
The example in Figure 8.1-4 shows a best-fitting curve. Two suitable methods for
achieving this could be linear prediction or bivariate polynomials. With respect to these
best-fitting surfaces, the residuals υ, from the least-squares adjustment are shown. In
2
Kager, H., Kraus, Κ.: Optical 3-D Measurement Techniques V (Eds.: Grün/Kahmen), pp. 103-110,
Wichmann Verlag, 2001.
a second least-squares adjustment, the Ζ coordinates of the scanned points are given
special weights pl which depend on the residuals of the first adjustment.
Such weights, calculated on the basis of a weighting function, are introduced to enable
robust parameter estimation 3 as well as the reliable detection of gross errors. As a
rule, a symmetrical bell curve, centred on the zero point of the residuals v, is used
as the weighting function. However, such a weighting function is not suited to the
robust estimation of a DTM from laser scanned points. It would, in fact, reduce in
equal measure the effect on the fitted surface of points with (large) positive and (large)
negative residuals. The result of such a robust adjustment would be similar to that
shown in Figure 8.1-4.
For the robust estimation of DTMs from laser scanned points, a weighting function
is applied in the second least-squares adjustment. This function takes account of the
following properties of errors in the scanned points:
• points with positive residuals, which are probably ground points, should be given
a higher weight than points with negative residuals, which are probably not
ground points. (In Figure 8.1-4, two points are shown with positive and nega-
tive residuals Vi.)
• the terrain probably lies below the curve or surface fitted during the first least-
squares adjustment (Figure 8.1-4), and a datum shift of the weighting function is
appropriate.
A weighting function, which takes account of these requirements, has the form (Fig-
ure 8.1-5):
for Vi < g Pi = . ΓΓΓ a,b> 0
1 + (a|ui -g\)b (8.1-3)
for Vi > g pi = 1
Vi . . . residuals from the first least-squares adjustment

a, b . . . parameters which control the decreasing effect of the weighting func-
tion from right to left
g ... datum shift which also reduces the effect of points with small positive
residuals vt on the second adjustment
Figure 8.1-6 shows the result of the second least-squares adjustment using the weight-
ing function shown in Figure 8.1-5. It meets our expectations: the DTM is fitted
through the low lying laser scanned points. In this combined interpolation and filtering
process, points above the model have little influence. In practice, several iterations are
needed to reach a final solution.
By selecting an adequate threshold value for the residuals ν in the second least-squares
adjustment, the original scanned points can be classified into points either on the
3
For example, Section Β 7.2.1.5, Volume 2, and Koch, K.R.: Parameter Estimation and Hypothesis
Testing in Linear Models. 2 nd edition, Springer 1999.
Figure 8.1-4: Best-fitting curve through the laser scanned points and residuals of the
first least-squares adjustment
-
1 ^tsffl
/j m m
/
9
! 1 1 "I I II I ! I I1 I! I1 1t !I 1I .1I I I I I I IΜ I II fI Γ
ι I tι \„v
-4.0 -2.0 0.0 2.0 4.0 6.0 8.0 i
Figure 8.1-5: Weighting function for a robust least-squares adjustment which derives
a DTM from laser scanned points
Figure 8.1-6: Results of the second least-squares adjustment using the weighting
function of Figure 8.1-5 and the residuals v^ from the first least-squares adjustment
in Figure 8.1-4
ground or off the ground. Strictly speaking there should be a final least-squares ad-
justment using only ground points without application of variable weights.
For widespread areas of dense forest and extensive building, the process of interpola-
tion and filtering must be expanded into a hierarchical strategy, described in a number
of publications 4 .
Figure 8.1-7 shows the result from part of an international test 5 . Despite the relatively
low point density of 0.25 points/m 2 , i.e. with an average point separation of 2 m along
and across the direction of flight, vegetation and buildings have been successfully elim-
inated, as can be seen in Figure 8.1-7. A small building on the railway embankment
remains: its height is too low.
Figure 8.1-7: DSM (left) and fully automatically derived DTM (right) from a small
part of the OEEPE test at Vaihingen
Figure 8.1-8 shows the DSM from a scanning flight over Vienna. It has a grid spacing
of 0.5 m which is approximately the point separation in the original data. In so far as it
can be defined in urban areas, the DTM can be derived from this model (Figure 8.1-9).
The accuracy of this DTM, which depends on an individual area's land use, has been
determined by the City Council's survey department (Magistratsabteilung 41) with the
help of a large number of ground check points. Table 8.1-2 shows the results. Note that
there is a distinction between standard deviation σ (see also Appendix 4.6-1) and r.m.s.
(root mean square error). The r.m.s. is calculated from the actual errors ε, and these
also contain systematic effects. Laser scanned DTMs, depending on surface roughness
and ground cover, are generally too high.
The standard deviation of ± 1.0 cm on open streets is remarkable at first glance, partic-
ularly in respect of the fact that individual laser points have a height accuracy of only
around ± 6 cm. (This accuracy corresponds to the manufacturer's specifications for the
TopoSys laser scanner which was used for data acquisition.) This large improvement
in accuracy can be simply explained: on surfaces with very limited curvature, the least-
squares adjustment leads to an averaging of the scanned points within a limited area. It
4
E.g. Briese, C., Pfeifer, Ν.: in Optical 3-D Measurement Techniques V (Eds.: Griin/Kahmen),
pp. 80-87, Wichmann Verlag, 2001.
5
Torlegard, K„ Jonas, Ν.: OEEPE-Publication No. 40, 2001.
Figure 8.1-8: Surface model (DSM) Figure 8.1-9: DTM derived using ro-
from laser scanner data (0.5 m grid bust, hierarchical estimation
spacing)
r.m.s. σ
[cm] [cm]
Park, dense vegetation ±14.5 ±11.1
Park, sparse vegetation ±11.4 ±7.8
Park, open ground ±8.6 ±4.5
Streets with parked cars ±9.2 ±3.7
Streets without cars ±2.4 ±1.0
All points ±10.5 ±7.1
Table 8.1-2: Accuracies of the DTM in Figure 8.1-9 for various types of land use
is well known that the arithmetic mean has an accuracy of:
c"m = ~~F= (8.1-4)

\Jn
OE ••• standard deviation of a single observation

σΜ ... standard deviation of the arithmetic mean
η ... number of contributing observations
The accuracy increase from ± 6 cm to ±1 cm is reached when 36 points are included in
the averaging, which is readily achieved at a point separation of 0.5 m, i.e. a 3 m χ 3 m
area on the road.
8.1.2.3 Generation of building models
The automatic generation of building models from laser scanner data is currently a sub-
ject of intensive research. Only the results of two different methods will be presented
here.
The first method starts by forming the difference between a DSM and a DTM. The dif-
ference model contains not only the buildings but also the vegetation and many other
objects lying above the DTM. Building points are extracted from the difference model
using special filter techniques. Within these filtered point clouds, planar areas are sub-
sequently detected and individual planes joined together to form roofs. Figure 8.1-10
shows the results for a building block in Vienna. The data originate from the same
scanning flight which provided the data for the derivation of the DTM in Figure 8.1-9.
The difference model formed at the beginning of this procedure is therefore a subtrac-
tion of the data in Figure 8.1-9 from the data in Figure 8.1-8. In order not to lose the
eaves of the roofs, the extraction of building models uses range measurements based
on the first echo (the third ray from the left in Figure 8.1-2 illustrates this procedure).
Figure 8.1-10: Building model extracted from laser scanner data
The second method was developed under contract to engineering consultancy Schmid,
Vienna, by the "Advanced Computer Vision" technology centre. It has some similar-
ities with the previous method discussed above. However, it makes use of the ground
plans of buildings which are often available in two-dimensional GIS databases. Fig-
ure 8.1-11 shows the result and indicates the laser spots used. This result is a CAD
model.
Further reading: Brenner, C.: DGK 530, 2000. Maas, H.-G.: IAPRS 32 (Commission
3-2 W5), pp. 193-199, Munich, 1999. Vosselman, G., Dijkman, S.: IAPRS 34 (Com-
mission 3/W4), pp. 37^-3, Annapolis, 2001. Peterneil, M., Steiner, Τ.: Computer-
Aided Design 36, pp. 333-342, 2004.
Figure 8.1-11: Building model derived from laser scanner data and a two-dimensional
building ground plan
8.1.3 Comparison of two paradigms and further performance

parameters of laser scanners
The term paradigm implies the pattern of approach and thought which lies at the heart
of a scientific or technical field. In the following section the paradigms behind airborne
laser scanning and (stereo-) photogrammetry will be compared.
The essence of (stereo-) photogrammetry is explained by the simple diagram of Fig-
ure 8.1-12. Here the geometrical reconstruction of an object in three-dimensional
space, from at least two photographic images, is central to the process. One photo-
graph defines a bundle of rays or directions. A point on the object can be reconstructed
in three-dimensional space when it is intersected from at least two directions. The
directions, or rays, generated by the natural light of the sun, are recorded by sensors
known as "passive sensors". Within the context of a paradigm, it is irrelevant if:
• the record is made on film or an electronic imaging device.
• electronic recording is done using a two-dimensional array of detectors or a linear

array in a 3-line camera.
• image points are measured stereoscopically by a human operator or digitally,
using a correlation algorithm.
• a third image, or further images of the same object, or a set of images of an

extended object are available.
• a calibrated metric camera or uncalibrated amateur camera are used.
• the position and orientation of the images are determined from control and tie
points or elements, or from GPS and IMU measurements. (In Figure 8.1-12
"GPS/IMU" is placed in brackets to imply an option.)
The question now presents itself as to which paradigm lies at the heart of laser scanning.
As in (stereo-) photogrammetry, the geometrical reconstruction of objects in three-
dimensional space is foremost, although not from at least two measuring locations but
from only a single location (Figure 8.1-13). In place of a bundle of rays there is a
field of directions and distances, i.e. a vector field. In place of passive sensors there
are active sensors. GPS/IMU information is essential; the connection of two or more
scans in a spatial network can only be achieved using common surface elements and
not points.
(GPSMMU) Scan
GPSMMU
Scan 1 Scan 2
passive sensor passive sensor
Object point Object point
Figure 8.1-12: (Stereo-) photogrammetry Figure 8.1-13: Airborne laser

paradigm scanning paradigm
The remainder of this section will evaluate the contrasts between airborne laser scan-
ning and (stereo-) photogrammetry from an application perspective and then present
performance parameters for airborne laser scanning.
(a) in forested and built-up areas, laser scanning is superior to photogrammetry for
the derivation of terrain models. In photogrammetry, a relevant ground point
must be visible from at least two imaging positions but in laser scanning the
view from a single recording position is sufficient.
(b) active sensing makes laser scanning independent of (natural) sunlight. Laser
scanning flights can, in principle, take place at any time of day or night. Admit-
tedly, clouds are as much of a hindrance to the near infra-red emissions used for
laser scanning as they are to photogrammetric flight operations.
(c) laser scanning requires no texture on the object surface. In contrast to photogram-
metry, laser scanning can be employed for areas of sand devoid of vegetation
(dunes), smooth and dry cultivated fields, bright stretches of concrete, forested
ground in shadow, etc. However, on very flat surfaces which give specular re-
flection, and surfaces with a low reflectance (clear water), there is no laser beam
echo returned to the receiver.
(d) the accuracy of a DTM obtained photogrammetrically is critically dependent on

the flying height and camera type (Section 6.7.2c). The accuracy of a DTM deter-
mined by laser scanning is less dependent on the flying height (Table 8.1-1). In-
stead, the accuracy of a laser scanned DTM is highly dependent on the point den-
sity. The point density during laser scanning is influenced by the flying height,
the flying speed, the scanner's field of view and the scanner's data rate. The
following rule of thumb, from empirically derived accuracies, can be used to
estimate height accuracies:
a H [cm] = ± + 120tano) (8.1-5)
tan α ... ground slope

η . . . point density per square metre
The figures of 6 cm and 120 cm in Equation (8.1-5) remain somewhat uncertain

and should be further verified in the near future.
Numerical Example. To obtain country-wide DTMs, laser scans are currently flown
with point separations along and across the direction of flight of around 2 m, i.e. the
point density is 0.25 points m~ 2 .
(8.1-5) : H a [cm] = ± + 120 tan c ^ = ±(12 + 120 tan a),
i.e. for a ground slope of 10%: o n = ±24 cm
Roughly the same accuracy can be achieved with a wide-angle camera at a flying height
of 1 km ( m B = 6700):
(6.7-5): σΗ [cm] = ±(15 + 100 tan a),

i.e. for a ground slope of 10% : = ±25 cm
In forested areas, a greater photogrammetric standard deviation has to be expected

(compare Section 6.7.2b). Under the assumption that only 25% of scanned forest points
reach the ground (i.e. the average separation of points along and across the direction of
flight is 4 m), a considerable accuracy can still be achieved in a forested area:
(8.1-5): Forest fem] = ± | , 6 + 120 tana ) = ±(24 + 120 tan a )

' V\/0O625 J
Numerical Example. Particularly for the sake of geomorphological fidelity of the
DTMs, which will be examined in more detail in Section (e), laser scanning operations
are flown with an average point separation of 0.5 m along and across the direction of
flight. This leads to the following standard deviation σ of DTMs derived from these
data: ΖΓ
(8.1-5): σΗ [cm] = ± ( - p + 120tana) = ± ( 3 + 120tana)

v4
An area with ground slopes of 5%, for example, has an accuracy of σ η — ± 9 cm.
(e) the geomorphological quality of DTMs derived from laser scanned data is, like
accuracy, crucially dependent on point density. Provided that undulating terrain
is being recorded, the sampling theorem (supported illustratively by Figure 3.3-6)
provides the following connection between the minimum wavelength L mm and
the point separation Δ. (Note that in laser scanning the object is directly "digi-
tized". In terrain scans the point separation Δ is therefore often called the Ground
Sampling Distance, GSD.)
Lmin » 3Δ (8.1-6)
Figure 8.1-14 shows a ground profile with minimum wavelength Lmm and two
different sets of point records with point separation Δ. It is obvious that the curve
can be reconstructed from either configuration of points.
Note: with Lmm = 2Δ, an unfavourable distribution of laser points can deliver 3
identical Ζ values. For this reason, and in particular because of the planimetric
error in the laser points, the factor 2 should be significantly increased. The factor
3 has been chosen here, 2.8 is often used.
Figure 8.1-14: Minimum wavelength of a ground profile and two different arrange-
ments of points to determine it (Δ = point separation)
Laser scanning delivers point clouds in which no structural elements (edges on

the ground, etc.) are emphasized. With photogrammetric data capture, and often
as an extension to automated data capture (Section 6.8.6), considerable value is
placed on the structural elements. In order to ensure the prerequisites for deriv-
ing structural elements from laser point clouds, a high point density is required
during data capture. The spatial relationships for the cross-section of the em-
bankment of a dam are sketched in Figure 8.1-15. In order to determine the
edges (for which there is a publication6 describing this technique), the top and
side surfaces, assumed planar, must be recorded with at least two points. The
relationship between the minimum width Bm\n and the point separation Δ is:
(8.1-7)
Figure 8.1-15 shows two different configurations of points with separation Δ

when the minimum width of the crown of the embankment is £ ? j . In both cases
m n
there is a minimum of two points on the top surface.

6
Briese, C.: IAPRS 35(B3), pp. 1097-1102, 2004.
Ζ Ζ
ΧΥ
Δ Δ Δ Δ Δ Δ Δ Δ Δ Δ Δ
Figure 8.1-15: Minimum width of a dam profile and two different arrangements of
points to determine it (Δ = point separation)
Exercise 8.1-1. An area of interest contains wavelike deformations in the ground with
a minimum wavelength of 4 m and dam walls with a minimum top width of 2 m. What
laser point separation should be chosen for data recording? (Solution: A W a v e = 1.3 m.
Aüam = Im, therefore use a point separation of 1 m.)
(f) compared with photogrammetry, laser scanning has two disadvantages. Firstly,
planimetric resolution is appreciably worse. Laser scanning currently has a reso-
lution in the half metre range. In contrast, photogrammetry can provide resolu-
tion in the decimetre range and occasionally better (Section 3.1.5.2). Secondly,
laser scanning only provides a "range image", in contrast to photogrammetry
with its multispectral information in the visible and infra-red region of the elec-
tromagnetic spectrum which facilitates delineation of land usage and object iden-
tification. It should also be mentioned that currently photogrammetry provides
the better planimetric accuracy and airborne laser scanning the better height ac-
curacy. It is therefore not surprising that, from the separate paradigms of (stereo-)
photogrammetry (Figure 8.1-12) and laser scanning (Figure 8.1-13), a common
paradigm can be created. This common paradigm is illustrated in Figure 8.1-16.
The (straight-line) directions of the imaging rays (Figure 8.1-12) and the
(straight-line) vectors (Figure 8.1-13) have been replaced in the common par-
adigm (Figure 8.1-16) by symbolic (electromagnetic) waves. This serves to em-
phasize the physical line of sight in the common paradigm. The common para-
digm can be characterized as follows:
• for every photogrammetric pixel there is also a "spectrum" of distances

which consists of at least the first and last echoes.
• for every laser scanning pixel there is also a spectrum of electromagnetic
radiation which, depending on sensor, is typified by natural and/or artificial
radiation.
• for every pixel there is, in general, a second pixel with the identical infor-
mation content but recorded from a different direction.
GPSMMU
Scan 1 Scan 2
passive sensor passive sensor
Object point
Figure 8.1-16: Common paradigm based on (stereo-) photogrammetry and laser scan-
ning
A possible restriction to single components of the common paradigm is of con-

siderable importance in practice (object recording using only a single measure-
ment beam, etc).
The realization of the common paradigm can be achieved in different ways. Var-
ious alternatives, starting with simple solutions, are as follows:
• a digital orthophoto is superimposed on the DTM derived from laser data.

• the laser-based DTM is used for digital single image analysis (monoplot-
ting, Section 7.4.3).
• the laser scanner also records the intensity of the reflected laser beam. The
result is a black-and-white infra-red image (Figure 3.2-10), although one
which has the poor resolution of laser scanning.
• a digital photographic camera, either a CCD array camera or a 3-line cam-
era, is built into the aircraft, together with the laser scanner, but without
synchronization between them.
• laser scanner and digital linear array camera are mounted on the same gyro-
stabilized platform (Section 3.7.3.3) and are synchronized with one another.
With this configuration it is particularly important that the separation be-
tween the sensors is kept as small as possible.
Section 8.2 Terrestrial laser scanning 419
Further reading: Baltsavias, E.: ISPRS Journal 54, pp. 138-147, 1999. Jelalian, Α.:
Laser radar systems. Artech House, Boston, London, 1992. Rees, W.: Physical princi-
ples of remote sensing. Cambridge University Press, 2001. Wehr, Α., Lohr, U.: ISPRS
Journal 54, pp. 68-82, 1999.
8.2 Terrestrial laser scanning

8.2.1 Principle of operation
Terrestrial laser scanning differs from airborne laser scanning principally in that, dur-
ing the scanning process, the scanner is not moved 7 . The stationary terrestrial laser
scanner therefore requires deflection mechanisms for pointing in two directions. (In
airborne laser scanners the laser beam must only be deflected in a plane perpendicu-
lar to the direction of flight, Section 8.1.1.) Both deflection mechanisms are shown in
Figure 8.2-1. The pulsed laser beam is sent from the range finder electronics unit (1)
and meets the polygonal mirror element (3) which rotates at relatively high speed. The
laser beam (2) is reflected off the mirror surfaces such that it is scanned through the
vertical angle ζ. After a ζ profile has been recorded, the upper part of the instrument
rotates through a small angle Aa in order to sample the neighbouring ζ profile, and so
on until a full horizontal circle has been covered.
Figure 8.2-1: Principle of the Riegl terrestrial laser scanner
7
However, mobile mapping (Figure 3.8-12), also uses laser scanners whose operation corresponds, in
principle, to the airborne laser scanning concept.
In terrestrial laser scanning, the object is scanned from several measurement stations
such that there is only a small overlap between the point clouds generated at each
station. GPS can well serve the purpose of locating these stations and this will be
discussed in Section 8.2.2. From one station one obtains the polar coordinates α, ζ and
s. The data recorded from one station is, in terrestrial laser scanning, normally called
a scan.
The technical data for the Riegl LMS-Z420i scanner are summarized in the following
table:
Wavelength/Intensity 1550 nm (near infrared), the intensity of

the echo can be recorded
Maximum frequency of vertical line 20 Hz
scan (ζ profile)
Data rate 24 kHz normal scanning, 8 kHz fine
scanning
detectable echos first or last echo
maximum range 200 m for a surface reflectance ρ > 80%,
60 m for ρ > 10% (see Section 3.7.5)
minimum range 2m
FOV, field of view, for a ζ profile 80°
IFOV, instantaneous FOV 0.25 mrad(52"), i.e. the footprint has a di-
ameter of 2.5 cm at a range of 100 m and
1.25 cm at a range of 50 m
Range accuracy ±10mm
minimum increment in both angular 0.07 mrad(14"), i.e. the sampling interval
directions for an object at 100 m range is 7 mm and
at 50 m range it is 3.5 mm
Table 8.2-1: Technical data for the Riegl LMS-Z420i terrestrial laser scanner
8.2.2 Georeferencing
This section will first discuss direct georeferencing. This assumes that the laser scan-
ner's position (Xq, Vq, ZQ), for example with the help of GPS in static mode, and an-
gular orientation (ω, ψ, κ) have been determined. From these parameters, and the polar
coordinates a, C, s, the X Y Z coordinates in a global coordinate system can be derived:
Χ \ ί Χο \ ( s sin ζ cos a \
Y = r0 + I W ssinCsina (8.2-1)
Ζ J \Z0J \ s cos £ /
ζ ... zenith angle with zero direction along the κ axis

a . . . horizontal angle, with zero direction defined by the zero direction for
κ rotation
Comment: It is interesting to make a comparison with the similar relationship (8.1-1)

for airborne laser scanning. It is particularly noticeable that no synchronization of the
transformation parameters is required in terrestrial laser scanning and that the polar
coordinates are derived from two deflection angles.
If the location coordinates and orientation angles are unknown (a technique known also
as free stationing), then these parameters must be determined indirectly from control
points (indirect sensor orientation). In this we initially assume that, in every station
scan, a number of control points can be located and identified. With the Riegl LMS
Z420i (Table 8.2-1) this is an automated measurement procedure provided that control
points have been marked by small retro-reflecting targets. Due to the high contrast be-
tween these targets and their environment, the signals can be automatically identified
in the intensity image (Table 8.2-1). After normal scanning and preliminary location
of the control points, areas around the individual signals are re-scanned at a very high
resolution. In the re-scanned intensity images, the Ca coordinates of the control points
can be determined to sub-pixel accuracy, for example using a weighted centroid calcu-
lation (Section 6.8.5 ff.). Using the 4 nearest scanned points, the s coordinate can be
determined by means of a bilinear transformation (Section 2.2.3a).
The transformation parameters Xo, Yq, Zo and ω, φ, κ for one measurement station can
be determined from at least 3 well distributed control points. A simple solution8,
which corresponds to the absolute orientation of the photogrammetric stereomodel
(Section 4.4), starts with the conversion of polar coordinates into a local Cartesian
coordinate system as follows:
(8.2-2)
The continuation of this "absolute orientation" is described in detail in Section 4.1.1.

The scale factor can be assigned the value of 1.
The economic advantage of indirect sensor orientation can be appreciably increased by
simultaneously connecting all scans in a measurement project. For the full set of scans
only a few more control points need be determined in the object coordinate system.
To connect scans, retro-reflecting targets are again required which can, however, be
automatically found and identified in their overlapping areas. Figure 8.2-2 illustrates
three scanner stations connected through common measured points. After converting
polar coordinates to Cartesian coordinates (Equation (8.2-2)), each scan has its own
local Cartesian coordinate system (Figure 8.2-2 shows this only for the middle station).
In addition, Figure 8.2-2 shows the global X Y Z object coordinate in which a number
of retro-reflecting control point targets must be measured.
The indirect sensor orientation for a complete set of scans, each with local Cartesian co-
ordinates derived from the original polar coordinates, corresponds to the spatial block
adjustment with independent models described in detail in Section 5.2. The scale fac-
tors m in the individual scans can again be assigned the value of unity. Accuracies and
8 A more rigorous solution, which minimizes the least-squares sum of the original observations α, ζ, s,
can be found in Section Β 3.5.3, Volume 2.

Figure 8.2-2: Connection of multiple laser scanning stations (scans)
distributions of control points can also be found in Section 5.2.3. Since terrestrial laser
scanning is often used to record facades on both sides of a street, then the accuracies
and control point configurations of strip triangulation are appropriate for such linear
measurement tasks (Section 5.2.3.4).
In a practical example9 with five scans, the following accuracies were achieved:
±3.5 mm for individual coordinates of a measurement station and ±7.5 mm for the
individual coordinates of a tie point.
8.2.3 Connecting point clouds

Laser point clouds have no sharply defined points which could be identified in several
point clouds and then used to connect them. Retro-reflecting targets, introduced in the
previous section for the simultaneous connection of all scans in a measurement project,
represent a relatively expensive solution.
In Section 8.2.5 we will learn a technique for locating and identifing tie points manually
within the intensity images of the laser scans. To support this, an automatic fine scan
can run in the background, comparable with related methods to determine tie points au-
tomatically in aerotriangulation (Section 6.8.3.5). Laser scanner images are, however,
three-dimensional: the intensity image provides two dimensions, the third dimension
is provided by the distances s. In this context, interest operators (Section 6.8.1.3) can
also provide good service.
Single points have already been rejected for connecting the point clouds of individual
airborne laser-scanned strips. Planes in corresponding regions have been used instead
(Section 8.1.2.1).
9
Pfeifer, N., Kraus, Κ., Schwarz, R., Ullrich, Α.: Ingenieurvermessung 2000 (Eds. Schnädel-
bach/Schilcher), pp. 114-121, Wittwer Publishing, 2000.
The well-known ICP (iterative closest point) algorithm can connect point clouds with-
out using individual tie points. This assumes approximately positioned and oriented
point clouds. In the overlapping region between two point clouds, two subsets of points
are chosen. Subsequently, for every point in one subset a corresponding point, at the
shortest Euclidean distance in the other subset, is sought. In this way, many corre-
sponding point pairs are generated which can be used to transform one point cloud onto
the other, for example using a highly over-determined spatial similarity transform. The
point clouds then move "closer" to one another. In a second iteration, corresponding
point pairs with shortest separations are again determined and through them a second
transformation is carried out. Iterations are continued until the squared sum of point
pair separations is below a given threshold value. Due to the relatively large number
of incorrect matches which can occur, the transformations, which are normally highly
over-determined, should make use of robust parameter estimation. It is conceivable to
extend the method from two point clouds to the simultaneous connection of all point
clouds, comparable with block adjustment using independent models (Section 5.2).
Further reading on ICP: Besl, P., McKay, N.: IEEE Transactions on Pattern Analy-
sis and Machine Intelligence (PAMI) 14(2), pp. 239-256, 1992. Chen, C., Hung, Y„
Cheng, J.: Transactions PAMI 21(11), 1999. Pottmann, H„ Leopoldseder, S., Hofer,
M.: IAPRS34(3A), pp. 265-270, Graz, 2002.
8.2.4 Strategies for object modelling

The result of measurement by laser scanner is a point cloud. In the following explana-
tion it will be assumed that the primary axis of the instrument (Figure 8.2-1) is approx-
imately aligned to the direction of gravity. For one station or scan, the recorded point
cloud is shown projected onto the ζ α coordinate plane in Figure 8.2-3. The various
range values s are the functional values of this "image".
By storing the measurement values as a matrix, neighbouring measurements can be

quickly accessed via the matrix subscripts. The measured points have no "meaning", in
contrast to a manual photogrammetric analysis or a tacheometer (total station) survey.
These last named, conventional methods assign to every measured point an attribute
(house corner, linking data to neighbouring point, etc.). From the laser scanner's point
cloud this information must be determined after data acquisition. As an aside it is
worth noting that laser scanning generates a large amount of data without attributes
and, in contrast, conventional methods generate a small amount of data with manually
assigned attributes.
Before we investigate further the determination of line and point information from
laser scanner point clouds, we will briefly discuss the geometry implicitly contained in
Figure 8.2-3. The columns contain object points which result from the intersection of
a (vertical) plane with the object surface. The particular (vertical) plane contains the
recording location and is defined by horizontal angle a. The line of intersection is a
(vertical) profile. The distances from the measurement station to the profile points are
known; the separations between profile points are defined by the Αζ increments.
α
—*•
ζ|
Figure 8.2-3: Laser points viewed in the ζ a reference plane
A row of the matrix in Figure 8.2-3 contains object points which result from the in-
tersection of a conical surface with the object's surface, where the cone has its apex
at the recording location and opening angle 2ζ (Figure 8.2-1). The distances from the
measurement station to these intersection points are known; the separations between
columns of the matrix in Figure 8.2-3 is defined by the Aa increments. If the object
surface is a plane then the intersection line between the current cone and the object
surface is a conic section, i.e. an ellipse, parabola or hyperbola. In the ζα coordi-
nate plane, these conic sections are drawn as straight lines along the corresponding a
coordinate line. Figure 8.2-4 illustrates the laser points after transformation into the
xyz coordinate system according to Equation (8.2-2). The constant interval between
points has disappeared. The simple neighbourhood connections (topology) between
laser points is, however, unchanged by this transformation. Transformation into the
xyz system has the major advantage that planar regions can be found, for example by
detecting common normal vectors to triangles created from neighbouring points. (In
contrast, planar regions cannot be found in the ( a s coordinate system of Figure 8.2-3;
object planes and lines are bent.) In the example here 4 planar regions can be found
and their laser points are marked with different symbols in Figure 8.2-4.
After segmentation, planes bordering on one another can be intersected. This results
in straight lines and points (Figure 8.2-5). The following strategy for object modelling
therefore lies at the heart of laser scanning: sets of points belonging to individual,
smooth surface elements (in our example these are planes) are extracted from the point
cloud. By intersecting the smooth surfaces, curved edges are obtained (in our example
these are straight lines) and the intersection of the edges provides the required indi-
vidual object points. Conventional methods have a reversed strategy: single points are
'•Α.,
······
Οο ο ο Ο Ο ο ο ο . iV.'.V.V;
Ο ο ο ο Ο ο ο ο ο ο ο> 0 0 0 0 0 0 0 0
Ο Ο Ο Ο Ο Ο Ο Ο
ο ο ο ο ο ο ο ο οο οοοοοοοοο
ο ο ο ο ο ο ο ο οΟ 9οοοοοοοο
ο ο ο ο ο ο ο ο οο οοοοοοοοο ,,···ΜΗ(
ο ο ο ο ο ο ο ο ο ο οο ο ο ο ο ο ο ο
·ο · ο· ο· 'ο ο β' - - - » · · ·
ο • ο*ο· ο ο ο 4ο ο ο ο ο
ο *ο
•·ν.\|
Figure 8.2-4: Laser points in Cartesian coordinate system and the result of segmenta-
tion in 4 planar regions.
Ο Ο ο
ο ο ο •••••«••o.o.ir·. ····:
ο ο ο οοοοοοοοοοοβοοοβ
ο ο ο οοοοοοοοοοβββοββ
ο ο ο Οοοοοοοοοοοβοοοβ •·····
ο ο ο ο οοοοοοοοοοβββοβ »•••··
ο ο ο ο οοοοοοοοοοοοοοο ······
ο ο ο οοοοοοοοοοοοβοοο
ο ο ο ο οοοοοοοοοοββοοβ
······
*·····
Figure 8.2-5: Intersection of surfaces and edges
measured which create lines defining the object's topology and the lines, together with
individual points on smooth surfaces, define the surface elements.
The accuracy of lines and points indirectly derived from laser scans is very high be-
cause the highly redundant set of laser points leads to a significant increase in sur-
face measurement accuracy compared with a single laser point measurement (Sec-
tion 8.1.2.2). In an analogous way to robust least-squares interpolation of surfaces
(Section 8.1.2.2), this reduction, or filtering, of random measurement error should be
undertaken on smooth object surface elements. For the assignment of laser points to
the smooth surface elements it would be advantageous if the random errors were al-
ready filtered out. In practice, therefore, segmentation and filtering are alternately used
in an object reconstruction procedure.
Chapter 8 Laser scanning
Figure 8.2-6 shows the result of a measurement with a Riegl LMS-Z360 laser scanner.
The statue of Marcus Antonius on a carriage being drawn by three lions, was scanned
from several locations. Using a small number of retro-reflecting targets fixed near and
around the statue, all scans were connected together and transformed into a common,
global coordinate system. The surface was modelled using the software package Geo-
magic Studio 10 . Data can be smoothed with this software, i.e. measurement noise is
filtered out, and NURBS elements formed (Non-Uniform Rational B-Splines). Details
are published in: Briese, C., Pfeifer, Ν., Haring, Α.: CIPA conference proceedings,
Antalya, Turkey, 2003.
Figure 8.2-6: Results of a laser scanning measurement of the statue of Marcus

Antonius
8.2.5 Integration of laser data and photographic data

The combination of laser data and photographic data is just as appropriate to terrestrial
laser scanning as it is to airborne laser scanning (Section 8.1.3). For some time terres-
trial laser scanners have been able to record the intensity of the reflected laser impulse,
as well as its range. Figure 8.2-7 contrasts a laser scan made with a Riegl LMS-Z360
and a digital photographic image of the same scene, taken with a Kodak DCS 460c
(Section Ε 3.5, Volume 2).
The resolution of both images is very different. The terrestrial laser scanner, like the
airborne laser scanner (Section 8.1.3ff.), has a worse resolution than the camera by
an order of magnitude. The laser scanner image is deceptive: it is a black-and-white
'"www.geomagic.com/en/products/studio/
Digital Images (Kodak DCS460C)
Figure 8.2-7: Laser scanner intensity image and photographic image
infra-red image (Figure 3.2-10). In this region of the spectrum there are certainly large
contrasts between different types of vegetation but only small contrasts between build-
ing walls (Figure 3.7-18). In addition, the laser scanner image has no three-dimensional
quality: instead of the oblique illumination by sunlight which gives the photographic
image a three-dimensional look, the laser scanner image is generated by an artificial,
near infra-red illumination with the same direction as the recording rays. Compared
with the photographic image, however, the laser scanner image has the significant ad-
vantage that, concealed behind each pixel, is the distance to that particular pixel.
To combine laser scanner and photographic data, the same solutions are possible for
terrestrial laser scanners as are given for airborne laser scanners in Section 8.1.3ff.
However, the connection of different image types into the same georeferenced network
should be briefly discussed. If the different images appear on the same display, as
in Figure 8.2-7, then corresponding points in each can be manually (monoscopically)
identified and, preferably using feature-based correlation (Section 6.8.1.4) located to
sub-pixel resolution. Following this, a common, least-squares adjustment of the terres-
trial laser scanning stations and the photographic image stations can be executed.
Three-dimensional photo models (Section 7.5), arising from both data sources, are par-
ticularly attractive. The geometry of the object surfaces is mainly derived from the laser
data; texture is taken from the digital photographic data and transferred to the surfaces
of the geometric model. A very simple type of photo model can be obtained without
the need for the full effort normally required for photo models. This variant requires
that a digital camera be rigidly mounted on top of the terrestrial laser scanner and its
orientation, with respect to the scanner, known by calibration. The following procedure
is then possible. Directly after recording, the 3D points from a laser scanning station
are transformed (back projected) into the digital photographs where the correspond-
ing colour values are extracted. The colour values are subsequently applied to small
spheres and, in 3D space, placed at the positions of the scan points. This results in a
3D "sphere cloud" which gives an observer the familiar colours of a photograph. This
3D "sphere cloud" can be generated immediately after recording and then visualized
in a very pleasing way.
8.3 Short range laser scanning

At very close scanning ranges, from a few decimetres up to around 2 m, a completely
different scanning principle is adopted. In general terms this principle can be iden-
tified as a light-sectioning technique: a light source projects a plane of light onto an
object surface, thereby generating a profile which is recorded by a digital camera with
a known offset from the light source. In this technique the laser is not used for range
measurement; it is simply used to create a plane of light which "draws" the above
mentioned profile on the object. The various solutions for implementing the light-
sectioning method are described in detail in Section C 2.3.3, Volume 2. In the follow-
ing text, a very specific and efficient solution will be presented which has matured in
recent years to a high level of perfection.
The principle can be very clearly explained on the basis of the horizontal "normal case"
of photogrammetry. Figure 8.3-1 illustrates light sectioning with a laser scanner. This
figure has close similarities with Figure 3.8-1 showing the "normal case" of terres-
Figure 8.3-1: "Normal case" of light sectioning by a laser scanner

Section 8.3 Short range laser scanning 429
trial photogrammetry. It is therefore also possible to describe the method shown in

Figure 8.3-1 as the "normal case" of light sectioning by laser scanning.
The metric camera at recording position Οχ in Figure 8.3-1 is here replaced by a laser
scanner which generates a plane of light. This, in turn, creates a profile of light on the
object. This plane of light can be rotated by an angle αϊ about the ζ axis of the model
coordinate system (= the scanner's internal ζ axis). The laser scanner transmits the
angle αϊ for the different light profiles to the data processing unit of the measurement
system. The digital (CCD) metric camera is positioned at a distance B , which is the
base length in the "normal case", along the χ model coordinate axis and with its opti-
cal axis (viewing direction) parallel to the y model coordinate axis. The CCD camera
makes a photographic recording of the light profile, synchronized with the transmis-
sions of the laser light planes (at varying angles a\). Each profile is separately recorded
on a full digital image. The CCD camera must therefore be capable of a high image
recording frequency.
At a particular time of recording, the digital image of the profile corresponding to the
light plane transmitted at rotation angle αϊ is shown in Figure 8.3-1. The image plane,
shown projecting from the negative position in Figure 2.1-6, has been rotated about the
ξ2 axis for display purposes. (Although these are digital images, the definition of image
coordinate system used in Figure 3.8-1 has been used instead of the normal definition
presented in Figure 2.2-3.)
The coordinates xp, yp, zp, in the model coordinate system, of a point Ρ on the imaged
profile, can be determined from its image coordinates £2, the principal distance c of
the camera, base length Β and angle αϊ of the laser light plane as follows:
— = tan αϊ (8.3-1)
yp
^ ^ = (8.3-2)
yp c
The coordinate yp is then given by Equation (8.3-3a), obtained after elimination of the
XP coordinate in Equations (8.3-1) and (8.3-2). The coordinate xP can subsequently
be derived by substituting yp in Equation (8.3-2) to obtain Equation (8.3-3b). The third
object coordinate, zp, is calculated from the ratios in Equation (8.3-3c):
yp = ^—E- (8.3-3a)
tan αϊ
c
X p = y p b + B (8.3-3b)
c
Zp = yp— (8.3-3C)
c
The processing of the images can be fully automated. The point sequences in images
of individual profiles can be located using simple image processing algorithms. Equa-
tions (8.3-3) are then subsequently evaluated for each located point using its image
coordinates £2 and 7/2 - Although the processing strategy is not very susceptible to er-
ror it should, however, be noted that there is no check with respect to individual object
points. In Equations (8.3-3) there are 3 measurement values ξ2, % a r | d a\ for determin-
ing the 3 coordinates xP, yp, zp. In contrast, in the "normal case" of photogrammetry
there are 4 measurement values ξι, ηι, ξι, r/2 in Equations (3.8-1) for the determina-
tion of the 3 coordinates x, y, z. In photogrammetry there is therefore one redundant
measurement present.
The combination of geometric data with photographic data is readily possible with this
construction principle, illustrated in Figure 8.3-1. After the CCD camera has taken
the sequence of individually imaged profiles, a digital photograph, generally an RGB
image, can then be taken. Three coordinates xP, yP, zP can then be assigned to each
RGB pixel of the digital image. With little effort, it is equally possible to implement
the reverse process in which pixels from the digital image are assigned to the xP, yp,
zP coordinates to create coloured spheres (see end of Section 8.2). The best fusion
results when an object surface is modelled from the laser point cloud and the complete
RGB image draped over this.
A few technical details will be given below for the relatively widely used Minolta
VIVID 9i scanner ( w w w , m i n o l t a 3 d . com):
Scan range: 0.5 - 2.5 m

Laser wavelength: 690 nm
Number of 3D pixels: 640 χ 480
Range accuracy: ±0.008 mm
Colour image array size (RGB): 640 χ 480
Data file size for 3D information and colour values: 3.6 MByte
Recording time for 3D data scan: 2.5 s
Table 8.3-1: Technical data for the Minolta VIVID 9i close-range laser scanner
Figure 8.3-2 shows the instrument. As indicated in Figure 8.3-1, the CCD camera is
mounted in the upper part of the housing. A laser light plane, which can be deflected
about a horizontal axis, is located behind the lower opening.
Figure 8.3-3 shows a three-dimensional object model, which has been recorded by a
Minolta VI-900. The CCD image has been transferred onto the object surface, derived
from the point cloud. Since the doll's head was recorded with only a single scan, the
unmeasured area in shadow at the back of the head can clearly be seen.
Section 8.3 Short range laser scanning 431
Figure 8.3-2: VIVID 9i 3D laser scanner from the Minolta company
Figure 8.3-3: Doll's head measured by the Minolta VI-900 laser scanner
Appendices
Appendix to Section 2.1
2.1-1 Three-dimensional rotation matrix

For the three independent parameters of a rotation matrix in three dimensions we may
use three rotation angles ω, φ and κ about the coordinate axes. Initially the xyz system
is coincident with the fixed (higher order, global) coordinate system XYZ. The xyz
system is then rotated in three steps. In each case an counterclockwise rotation as seen
looking along the axis towards the origin is regarded as positive.
the primary rotation ω of the xyz system about the X axis.

Note: the primary rotation takes place about one of the coordinate axes of the
fixed coordinate system.
Question: What equation is used to transform the χωι/ωζω coordinates of a point
into the fixed system?
It follows from Equations (2.1-1) and

(2.1-2) as well as Figure 2.1-1 that:
X = χω
Y = yω cos ω — ζω sin ω
Ζ = yLÜ sin ω + ζω cos ω
or
/ Ι Ο 0 \ ίχω\
Χ = Ι 0 cos ω - s i n w Ι Ι y u I = Η ω χ ω (2.1-1-1)
\ 0 sin ω cos ω J \ζω )
• secondary rotation φ of the xyz system about the axis.

Note: the secondary rotation takes place about the yu> axis which has been previ-
ously rotated about the primary axis.
Question: What equation is used to transform the χωφνωφζωφ coordinates of a
point through the xuyu,ζω system into the fixed system?
Appendices 433
Note that in the application of Equations (2.1-1)

and (2.1-2) as well as in Figure 2.1-1 a positive
^<0 ^ωφ φ rotation in the χωζω plane is a rotation in the
direction from the ζω axis towards the χω axis.
Positive rotations are to be taken in cyclical or-
der from the χ axis to the y axis, from the y
axis to the 2 and from the ζ axis to the χ axis.
This conforms with the definition that an coun-
terclockwise rotation as seen looking along the
axis towards the origin is positive.
Χω = Χωφ COS ψ + Ζωφ Sin ψ
Vu) = yωφ
ζω = -χωψ sin φ + ζωψ cos ψ
or
cos φ 0 sin ψ
Xti) — 0 1 0 — R'φ^-ωφ (2.1-1-2)
- sin φ 0 cos φ
Substituting from (2.1-1-1) in (2.1-1-2) gives:
(2.1-1-3)
tertiary rotation κ of the xyz system about the ζωφ axis.

Note: the tertiary rotation takes place about that axis which has previously been
rotated about both the primary and the secondary axes.
Question: What equation is used to transform the χ ωφ κ,νωφκ.ζ ωφκ coordinates
(writing χωψκ = χ, νωφκ = y, ζωφκ = ζ) of a point through the χωφνωφζωψ and
through the χ ^ ω ζ ω system into the fixed system?
ζ ζ z
ωφ ωφκ
Using Equations (2.1-1) and (2.1-2) and
Figure 2.1-1 it follows that:
ϋ
ΙΙωφ ωφκ sin κ + y ^ n cos κ
Ζ^ιρ ^ωφκ
or
cos κ - sin κ 0 ωφκ
= | sin κ cos κ 0 'ωφκ = R-rX (2.1-1-4)
0 0 1 ωφκ
434 Appendices
Substituting from (2.1-1-4) in (2.1-1-3) gives:

X = RwRvR,tx (2.1-1-5)
Multiplication of the three matrices R v and R r e to derive the single matrix
R is performed in two steps. First R ^ = R P R K is found and finally Έί ω φ κ =
ΈΙωΈίφκ.
/ CK -S«
CK
\
0'
SK 0
\o 0 11
cv 0 Sψ
0 1 0 SK Ck 0 — RVk
Sψ 0 S^jS K C^
1 0 0
0 cw -So,
0 sw Co,
(2.1-1-6)
If the order of rotations is changed the sequence of multiplications of the matrices
must be changed accordingly. For example, if ψ is taken as the primary rotation,
ω as the secondary rotation and κ as the tertiary rotation, we derive the following
matrix:
(
Cu;SK
c<^cK -f- swsK
cwCk —Sw I (2.1-1-7)

SypCK C(pS^jS^ s^s«
Exercise 2.1-14. Derive the matrix representing the resultant rotation in three dimen-
sions if κ, ψ and ω are the primary, secondary and tertiary rotations respectively, all in
the positive sense.
We sometimes wish to derive the rotation angles from the elements r ^ of the rotation
matrix. From the matrix we wish to find the three angles ω, φ and κ as specified
in Figure 2.1-5. It follows from Equation (2.1-1-6) that:
— Γ23 — ri2
tana;=
, sin<^ = ri3, t a n « = (2.1-1-8)
T33 m
Note: if η 3 > 0 then φ could lie in either the first or the second quadrant while, if
ri3 < 0, ψ could take a value in the third or fourth quadrants. There is, however, no
ambiguity for the angles ω and κ as may be seen when one considers the following
relationships from the matrix (2.1-1-6): 1
ω : T23 — — sin ω cos ψ r33 = cos ω cos ψ
κ: r\2 = — cos yj sin κ r\\ = cos ψ cos κ
'The unambiguous determination of the angles ω and κ is performed in many computers with the
ARCTAN2 function: omega = ARCTAN2 ( — Γ 2 3 , Γ 3 3 ) = [ ARCTAN2 (opposite, adjacent)]
The angle is given in radians with a value between —π and π (excluding —π), counting from the χ axis
in a counterclockwise sense (positive count) or in a clockwise sense (negative count). Example: The
function ARCTAN2 ( - 1 , - 1 ) produces (-2.35619); that is - 1 5 0 g o n or - 1 3 5 ° .
Appendices 435
Numerical Example. From the following rotation matrix which relates to rotations
about the axes in the order specified in Figure 2.1-5
-0.340110 0.999407 0.004822 \

-0.999419 -0.034096 0.000621
0.000784 -0.004798 0.999988 J
find the two sets of rotation angles ω, ψ and κ:
sin</?i = 0.004822 => φχ = 0.3070 gon (16'35") φ2 = 199.6930 gon (179°43'25")

0.000621/costî = - s i n w i = 0.0006210 ω 1 = 399.9605 gon
0.999988/ cos φ χ = cosî = 0.9999996 (359°57'52")
0.000621/ cos ψ2 = — sin ω 2 = -0.0006210 ω 2 = 199.9605 gon
0.999988/ cos ψ 2 = cosw 2 = -0.9999996 (179°57'52")
0.999407/ cos φ χ = - s i n / i j = 0.9994186 κι = 297.8292 gon
-0.034091/costî = cos μ = -0.0340914 (267°18'00")
0.999407/ cos φ2 = - s i n / i 2 = -0.9994186 ^ K2 — 97.8292gon
- 0 . 0 3 4 0 9 1 / c o s ψ2 = cos/i 2 = 0.0340914 (87°18'00")
Exercise 2.1-15. Repeat the above numerical example under the assumption that ψ
is the primary, ω the secondary and κ the tertiary rotation. Hint: The corresponding
rotation matrix is shown above as Equation (2.1-1-7). (Solution: ψ = 0.3070 gon =
16'35" OR φ — 200.3070 gon (180°16'35"); ω = 399.9605 gon (359 0 57'52") ORu; =
200.0395 gon (180°57'52"); κ = 297.8290 gon (268°02'46") OR κ — 97.8290 gon
(88 o 02'46").)
Exercise 2.1-16. The following rotations take place in the order ωφκ: ω = 2 gon
(1°48'), ψ = lOgon (9°), κ = 50gon (45°). In order to achieve the same resul-
tant rotation by means of three sequential rotations in the order φωκ, what would the
individual rotations be? (Solution: ω = 1.9754gon (1°46'40") OR 198.0246gon
(178°13'20"), ψ = 10.0049gon (9°00'16") OR 210.0049gon (189°00Ί6"), κ =
50.3130gon (45°16'54") OR 250.3130gon (225°16'54")·)
The above examples show that numerical problems can occur in the derivation of the
rotation angles from the elements rik of the rotation matrix. These numerical problems
occur when cos φ = 0, that is when φ = 100 gon = 90° or when φ = 300 gon = 270°.
Fortunately such angles do not arise in aerial photogrammetry, although they are quite
possible in close range applications. Similarly, numerical problems arise in the case
of the Ηγ,ωκ rotation matrix (Equation (2.1-1-7)) when cos ω = 0, that is when ω —
100 gon = 90° or when ω = 300 gon = 270°.
The following are some sources of information concerning rotation matrices: ASPRS
Manual of Photogrammetry, 5 th ed., 2004. Shih, T.-Y.: PE&RS 56, pp. 1173-1179,
1990. Sections Β 3.4.1 and Β 3.4.2 in Volume 2.
436 Appendices
2.1-2 Mathematical relationship between image and object

coordinates (collinearity condition)
point Ρ
When the photograph is taken, the object point P , the image point P ' and the cam-
era station Ο lie on a straight line. The above illustrated collinearity condition, as it
is called in photogrammetry, can be expressed in the X'Y'Z' coordinate system in
Equations (2.1-2-1) 2 ; the X'Y'Z' coordinate system is parallel to the ξηζ image coor-
dinate system, both equally tilted with respect to the X Y Z object system. In the image
coordinate system ζ = 0 for all image points and ζ = c for the perspective centre.
ξ-ξο = X'-X'0
c Z'-Z'
(2.1-2-1)
η-ηο = V'-Y0'
c Z'0-Z'
2 Zavoti (Geodetic and Geophysical Research Institute of HAS (GGKI), Sopron, Hungary) provided
the idea for this derivation.

Appendices 437
y — Υ'
<2 1 2 2)
ν - γ - ' -
If the rotation matrix for the image, and therefore also for the X'Y'Z' system, is R
with respect to the X Y Z object coordinate system, it follows from the collinearity of
Ο, P' and Ρ and from Equation (2.1-11) that:
Πι Π2 Π3
T-2ir 2 2 r 2 3 Y'-Y0' I (2.1-2-3)
. Γ31 Γ32 Γ33 Z'
Premultiplying Equation (2.1-2-3) by R _ I = R T and substituting in Equation (2.1-2-2)

from Equation (2.1-2-3) one obtains:
c t rn(X-Xo)+r2i(Y-Yo)+MZ-Zo)
ξ = ςο - c-
rl3(X - Χο) + r23(Y ~ Υο) + r33(Z - Ζ0)
Ζχχ
Ζ
~Ν (2.1-2-4)
ry2{X - Χ0) + r22(Y - Υρ) + r32{Z - Ζ0)
η = ηο - c
rl3{X - Χ0) + r23(Y - Υ0) + r33(Z - Ζ0)
Since the X'Y'Z' system is parallel to the ξηζ image system, the elements rlk are
• the cosines of the angles between the axes of the image and the mapping coordi-
nate systems (see Equation (2.1-10))
• functions of the angles ω, φ and κ through which the camera was tilted with
respect to the object system at the moment of exposure (see Equation (2.1-13) or
Appendix 2.1-1)
For the sake of compact notation in the partial differential coefficients given in Appen-
dix 2.1-3, in Equations (2.1-2-4), above, we write Zx and Zy for the numerators and Ν
for the denominator.
The collinearity equations are sometimes required in another form, a similarity trans-
formation which can be taken directly from Figure 2.1-2-1 (for details see Section
Β 3.5.1 in Volume 2):
'X-Xo
X - X o = | Y-Yo I = mR =mR(£-£0) (2.1-2-5)
Z-Z0
m ... a scale factor, the ratio of the length OP in object space to the length
OP' in image space
438 Appendices
2.1-3 Differential coefficients of the collinearity equations

For a solution of the non-linear equations (2.1-2-4), using Newton's (iterative) Method,
almost invariably together with the method of least squares, one requires the following
partial differential coefficients in which the dependence of the elements r ^ on the
angles ω, ψ and κ (shown in Equation (2.1-13)) is introduced:
δξ ΖΧ δη
de de
Οξ δη z
= ~^{rl3Zx - r u N ) = a2 y - rl2N) = b2
dXo dXo
251 ^
dt
= a3 --^{ri-IZY - r22N) = b3
II
dYo
δξ
= ~^{r33Zx -r3lN) = <14 - rnN) = 64
II
dZ0
- (Y - Y0)r3l + (Z- Z0)r2l) = a5
- (Y - Y0)r32 + (Z- Zo)r22) = b5
δξ c Λ „ • \ZX »τ ι
= — \ZX cos κ — Zy sin κ) — + Ν cos κ I = ae
οψ Ν y Ν
δη c ( Ζ
-— = — ( (Zx cos κ — Zv sin κ) —~ - Ν sin κ1 ] = be
αφ Ν \ Ν
δξ _ __c _ δη _ c_
7
δκ ~ Ν ~ δκ ~ Ν
θζ c θτι C
Οχ = - Jp(Nr" ~ =a« = -7^2(^12 - V13) = h
θ£ c ΘΤΊ C
1 NR21 ZXT2I) = 09 =
ÄR = ( - OY ~N2{NR22 ~ ZYR23) = 69
θ£ C θτι C
=
dz 'N2^Nr31 ~ ΖχΤ33
^ = ai
° dZ =
~Ν2^ΝΤ32 ~ ZyT33
^= 610
Exercise 2.1-17. Consider in what way the derivations of the differential coefficients
will change if the primary and secondary rotations are interchanged. The primary
rotation will therefore become φ and the secondary rotation ω. (Solution: They change
as follows: (Y — YQ) =>· —(X — Xo), r23 =>• —Γ13, r2\ =>· —rn, sin κ => — cos κ, cos κ
=>• sin«;.)
Appendices 439
Exercise 2.1-18. Give the differential coefficients for the case of an exactly verti-
cal, unrotated, image. (Solution (examples of the changes): θξ/dX = Βη/dY =
-c/(Z - Z0), θξ/dY = δη/ΘΧ = 0, Θξ/dZ = c(X - X0)/(Z - Z0)2, δη/θΖ =
c(Y - Y0)/(Z - Z0)2.)
440 Appendices

2.2-1 Derivation of Formula (2.2-5) using homogeneous coordinates
The projective relationship (2.1-24) between the image plane and an object plane can
be given as follows using homogeneous coordinates:
(
αϊ a2 a3 \ /ξ\ _
bi b2 63 U X = A£ (2.2-1-1)
Ci C 2 c3 / \ 1 /
One derives the original coordinates from the homogeneous coordinates as follows:
χ = x/w Y = Y/W (2.2-1-2)
The solution for the image coordinates from equations (2.2-1-1) is:
(2.2-1-3)
Hence
ξ = ί η= 2 (2.2-1-4)
The inverse of matrix A is:
j / 62C3 - &3C2 <23C2 - 0-2C3 a j b j - 0362 \

A - 1 = ——— b^Ci-biCi a i c 3 - a 3 c i 0361-0163 (2.2-1-5)
\b1c2 — biC\ a,2C\ — a\C2 a 162 — «2^1 /
Note: since (X, Y) and (ξ, η) in (2.2-1-2) and (2.2-1-4), respectively, are derived by
fractions which cancel out the scale information in A, scale must be fixed in some way.
By setting C3 = 1 and taking Equations (2.2-1-3), (2.2-1-4) and (2.2-1-5) together one
arrives directly at the relationship (2.2-5).
Note, however, that fixing the scale by setting C3 = 1 is only appropriate as long as
C3 φ 0. This will happen in the case where the line at infinity of the object plane is
mapped to an image line passing through the origin of the image coordinate system.
Appendices 441

4.1-1 Estimation by the method of least squares
We are given a set of linear equations of which the first is:
l\ = anx\ + anx2 Η Va\uxu (4.1-1-1)
aik • · known coefficients

li . . . known absolute terms described below as "observations"
Xfc . . . the unknowns, u in number, of the system of equations
If there are exactly u observations, li, then we have a consistent system of linear equa-
tions which may be solved directly for the u unknowns Xk- If the set of equations is
expressed in matrix form as 1 = A x then the solution is χ = A _ 1 l . In order both
to provide a check on errors in the observations and to increase the accuracy of the
results, it is advisable that we take η observations where η > u. We then have a set
of inconsistent equations which implies an estimation problem; this can be solved, for
example, by introducing the condition that the sum of the squares of the residuals υ of
the observations should be a minimum. We may then write a system of observation
equations:
ν = Ax — 1 (4.1-1-2)
The u unknowns Xk are then found by applying the following minimum condition:
v T v = min. (4Λ
4'2) (Ax - l ) T ( A x - 1) = x T A T A x - 21 T Ax + 1T1
Consequently the solution is 3 :
= 2 x T A T A - 21 T A = 0
dx (4.1-1-3)
χ = (ATA)-'AT1
The matrix ( A T A ) is known as the matrix of the normal equations; the matrix A is the
matrix of the observation equations or, in modern terminology, the design matrix.
The solution shown above is described in the literature as "adjustment by indirect ob-
servations". This algorithm is all that is needed for almost all of the estimation prob-
lems 4 dealt with in this book. Consequently, other approaches to adjustment are not
discussed. Just two extensions to the method of "adjustment by indirect observations"
are required, one to deal with the case when the initial equations (4.1-1-1) are non-
linear and the other when the observations are made with differing accuracies.
3
The observations 1 are stochastic variables (they therefore contain random errors) the calculated
unknowns are simply estimates. The customary notation for an estimated result, with the "hat", χ , is not
uniformly used in this book; it is frequently written just as x.
4
It is also common in surveying and photogrammetry to refer to "adjustment" when errors are itera-
tively "adjusted" to some minimum condition in order to achieve an optimal estimation of the unknowns.
442 Appendices
Let the first equation of a system of non-linear equations be expressed as:
h = f(x h x 2 ,...,Xu) (4.1-1-4)

A set of non-linear equations is commonly solved by Newton's method in which im-
proved approximations to the unknowns are found, in an iterative process, by deriving a
set of linear equations as follows. This lends itself to the method of adjustment by indi-
rect observations which deals only with linear equations. The right hand side of Equa-
tion (4.1-1-4) is expanded in a Taylor's series around initial estimates χ®, x®, • • •, x°u ;
we limit the expansion to linear terms:
Z, = / ( * ? , + + (4.1-1-5)
A comparison with Equation (4.1-1-1) leads to the following parallels:
a-ik — ( j , — ) in which these coefficients are formed by partial

^ Xk ' differentiation and then evaluated numerically
using the current estimates of the unknowns,
". The result is simply a number
which corresponds to the coefficient a ^ of the
linear case. (Why?)
X
li — li ι ' ' ' ' u )
— Ii — I® in which the value of the function

/ ( i p j j , · · · , ^ ) ' computed using the cur-
rent estimates of the unknowns, is denoted by
Z,; the actual observations are denoted by I t .
Xk == dxk in which dxu is a small correction to the current
estimate x(l of the unknown Xk-
Given these correspondences, it may be seen that Equation (4.1 -1 -5) is of the same form
as the linear Equation (4.1-1-1) and may be solved in the same manner by forming the
normal Equations (4.1-1-3).
In the second of the extensions to the method of "adjustment by indirect observations"
we consider the case in which we have differing accuracies in the observations li. The
accuracy is expressed in terms of the standard deviation σι of an observation or by
its variance σ\. In the surveying literature the term "root mean square error" is often
preferred to standard deviation. These different accuracies are taken into account in an
adjustment by indirect observations by means of weights, as they are called, pi, which
are summarized in a weight matrix P«.
/1/c \
\/σ\
P// = (4.1-1-6)
Appendices 443
The minimum condition then becomes: v T P » v = min., which leads to the solution:
χ = (ATP„A)-'ATP,;1 (4.1-1-7)
The (estimated) unknowns x, which can be derived from either of the Equations
(4.1-1-3) or (4.1-1-7), may be substituted in Equation (4.1-1-2) in order arrive at the
estimated residuals ν for all the observations 1.
From all the residuals ν we can derive an estimate for the standard deviation of unit
weight, also called root of the reference variance of an observation which has weight 1
in the weight matrix, (4.1-1-6):
= (4.1-1-8)
V η —u
With this standard deviation of unit weight σ 0 and the weight coefficients qkk the stan-
dard error a Xk of the individual unknowns Xk may be estimated:
Vx k =& 0 y/qkk (4.1-1-9)
The weight coefficients qkk are simply the corresponding elements from the main di-
agonal of the inverse of the normal-equation matrix:
Qxx = ( Α Τ Ρ „ Α ) - ' (4.1-1-10)
The special case in which there is no redundancy, that is when η = u, is also of interest.
In this case the result is the solution of the following linear equation system:
Ax = 1 x = A"'l (4.1-1-11)
The rule of propagation of errors may be applied to the last equation:
Qxx = Α - ^ Α - ' Γ = A"1(AT)"1 = (ATA)-' (4.1-1-12)
Comparison of this equation with Equation (4.1-1-10) means that one gets the accuracy
of the unknowns from a system of linear equations in a roundabout way from a least
squares adjustment. It is necessary to adopt a predetermined a priori accuracy for σο-
Literature: Mikhail, E.: Observations and Least Squares. IEP-A Dun-Donnelly Pub-
lisher, New York, 1976. Koch, K.R.: Parameter Estimation and Hypothesis Testing in
Linear Models. 2 nd ed., Springer 2006.
444 Appendices

4.2-1 Direct Linear Transformation (DLT) with homogeneous
coordinates
The relationship (4.2-3) between an image plane and a three-dimensional object can be
written as follows using homogeneous coordinates:
(4.2-1-1)
Relationship (4.2-3) is then obtained from the homogeneous coordinates ξ, ή and ζ as

follows:
ξ = ξ/ξ η = η/ζ (4.2-1-2)
Note: as has already been mentioned in Section 2.1.3, since the 3 χ 4 matrix of Equation
(4.2-1-1) cannot be inverted, a single picture is insufficient for the reconstruction of a
three-dimensional object with its X Y Z coordinates.
Appendices 445

4.3-1 Differential coefficients for the coplanarity equations
The coplanarity condition for the relative orientation using independent rotations re-
sults in Equation (4.3-11). One obtains the differential coefficients, which are neces-
sary, for example, for Equation (4.3-12), through differentiation of Equation (4.3-11)
with respect to the individual angles. Let us begin with dui^.
dD dp2,z dp2,y ....
= Pi,y ~ö - Pi,ζ-^- (4.3-1-1)
0u>2 σω2 olü2
As an illustration the derivative dp2, z /dw2 will be developed in full. The Equations
(2.1-13) and (4.3-9) give:
' sin u>2 sin K2 — cos u)2 sin ψ2 cos «2

Ρ2,ζ = (ζι V2 ~c) I sin u>2 cos K2 + cos u>2 sin ψ2 cos K2 I (4.3-1-2)
COS W2 COS Ψ2
Differentiation with respect to du>2 leads to:
Q / cos uj2 sin K2 + sin ω2 sin ψ2 cos κ 2

,z
— (^2 V2 —c) I cos l)2 cos K2 — sin u>2 sin ψ2 sin K2 ) (4.3-1-3)
ω2
y — sin UJ2 cos ψ2
Since the column vector of Equation (4.3-1-3) corresponds to the second row of the
rotation matrix (2.1-13), the relationship (4.3-1-3) simplifies to the y component of
the p2 vector (4.3-9), that is, P2,y- The derivative dp2,y/d^2 from Equation (4.3-1-1)
simplifies, after similar considerations, to the negative 2 component of the P2 vector
(4.3-9), that is, - p 2 , z · As a result the complete derivative from (4.3-1-1) is found:
ÖD
= P2,yPl,y+P2,zPl,z
The following derivative is obtained after comparable considerations:

dD
= COS U)2 P2,x Pl,y ~ Sin U)2 P2,x Pi,ζ
θψ2
We obtain the following derivative in a similar way:
dD = (&r2,32 - %i"2,3l)pi,y ~ (6^2,22 ~ mr2,2l)Pl,z
It should first be noted, when it comes to differentiation of D with respect to ψ\ and κι,
that, using ω — 0, the rotation matrix (2.1-13) can be simplified. The two remaining
derivatives are:
dD
dD . . . ,
-Q^- = {ξΐη,22 - mri,2l)P2,z - (ξΐΠ,32 ~ 7*1,31 )P2,y
446 Appendices
Exercise 4.3-3. If both rotation matrices, R i and R2, are replaced by the unit matrix,
these derivatives are considerably simplified. Equation (4.3-12), which applies to the
relative orientation of highly tilted photographs, then becomes Equation (4.3-6), the
equation for near-vertical photographs. Verify this conversion.
Appendices 447

4.6-1 The empirical determination of standard deviations and
tolerances
The standard deviation σ is soundly based in statistics. Sixty-eight percent of the actual
errors lie within the range ±σ. In the related disciplines of civil engineering, mechan-
ical engineering and so on, tolerances are preferred. The definition of tolerance is,
however, relatively arbitrary. A widely used tolerance t is the 2σ tolerance:
ί 9 5 = ±1.97σ « ±2σ (4.6-1-1)
As one can deduce from the Gaussian error distribution (normal distribution), about
95% of the true errors lie within this 2σ range. A somewhat more conservative toler-
ance is:
t 99 = ±2.58σ « ±2.5σ (4.6-1-2)
But there are always real errors even outside this 2.5σ tolerance. We can tell from the
Gaussian error distribution that there are about 1 % more. Nevertheless, the tolerance
f 9 9 is used in such a way that all real errors are presumed to lie within these bounds.
Not infrequently, when photogrammetric services are called for, an accuracy is re-
quested without defining the measure of accuracy. Associated professions prefer toler-
ances. If tolerance is adopted as a measure of quality in calling for services it must be
stated whether it is a question of i95 (4.6-1-1) or i 9 9 (4.6-1-2) or some other tolerance.
Assessment of the extent to which the specified tolerance tref has been met is carried
out with the help of check measurements of superior accuracy. To start with, an r.m.s.
value (root mean square error) (4.6-1-3) can be calculated from the differences which
arise:
r.m.s. ^ (4.6-1-3)
τι
η . . . number of check measurements
Frequently the r.m.s. is referred to as the absolute accuracy since it reflects all error
influences. If, from the disparities ε, one removes a possible systematic contribution
by averaging, the differences are diminished; from the differences reduced in this way,
which we denote by v, one obtains the standard deviation as follows:
^ (4.6-1-4)
71 — 1
The standard deviation a.m is also frequently called the relative accuracy. Continuing
the analysis, the tolerance tm can be calculated with Equation (4.6-1-1) or (4.6-1-2)
and compared with the specified tolerance i re f. One can also check directly whether
95% of the differences v, found after removal of a possible systematic part from the
differences ε, lie within the specified tolerance i 9 5 jre f·
448 Appendices
Check measurements of a superior accuracy are very costly. In many cases one is satis-
fied with a second set of measurements of the same accuracy as the first measurements.
The agreed tolerance iref can be checked from the differences d which arise. In this case
one first computes the standard deviation ad, m of the differences (in the same way as
explained above, one can also separate out the systematic part);
(4.6-1-5)
η . . . number of repeated measurements
Since two measurements of approximately equal accuracy contribute to the differences

d, the standard error of a single measurement, the actual standard error, comes to:
(4.6-1-6)
The tolerance tm can then be found from Equation (4.6-1-1) or (4.6-1-2) and compared
with the stipulated tolerance i re f.
In conclusion, it should be emphasized that a great deal of attention should be paid to
systematic errors when evaluating standard errors and tolerances empirically.
Numerical Example. When calling for photogrammetric point determination a toler-

ance of i95,ref = ± 1 0 c m is demanded. A value of a d > m = ±8.5 cm arises from the
differences d of repeated measurements (no systematic component contributed to the
differences d). Is the specified tolerance i95)ref = ± 1 0 cm met?
(4.6-1-6): a m
m = 8 . 5 c m / \ / 2 = 6.0cm
(4.6-1-1): £95,m = ± 1 2 cm
The specified tolerance i95jref is therefore not met. The photogrammetric point deter-
mination has to be repeated, possibly with a larger photo scale.
Completion of the references
ASPRS: American Society for Photogrammetry and Remote Sensing.

AYN: Allgemeine Vermessungs-Nachrichten, Wichmann Verlag, Karlsruhe, Germany.
BuL: Bildmessung und Luftbildwesen, Wichmann Verlag, Karlsruhe, Germany.
Can. Surv.: The Canadian Surveyor, The Canadian Institute of Surveying, Ottawa.
DGK: Deutsche Geodätische Kommission bei der Bayerischen Akademie der Wis-
senschaften, Verlag Bayerische Akademie der Wissenschaften, Munich, Ger-
many.
Geow. Mitt, der TU Wien: Geowissenschaftliche Mitteilungen der Studienrichtung
Vermessung und Geoinformation der TU Wien, Austria.
GIS: Geo-Informations-Systeme, Wichmann Verlag, Karlsruhe, Germany.
IAPRS or IAPR or IAP: International Archives of Photogrammetry (and Remote
Sensing) of ISPRS (ISP).
ISPRS: International Society for Photogrammetry and Remote Sensing.
ISPRS-J: ISPRS Journal of Photogrammetry and Remote Sensing, Elsevier Science
Publishers, Amsterdam, The Netherlands.
ITC-J: ITC Journal, Enschede, The Netherlands.
JEK: Jordan, Eggert, Kneissl, Handbuch der Vermessungskunde, 10th ed., J.B. Metz-
lersche Verlagsbuchhandlung, Stuttgart, Germany.
Mitt, der TU Graz: Mitteilungen der Geodätischen Institute der TU Graz, Austria.
Na.Ka.Verm.: Nachrichten aus dem Karten- und Vermessungswesen, Institut für
Angewandte Geodäsie, Frankfurt/Main, Germany.
OEEPE Publ.: Publications of the Organisation Europeenne d'Etudes Photogramme-
triques Experimentales (European Organisation for Experimental Photogram-
metric Research), now EuroSDR (European Spatial Data Research), h t t p :
//www.eurosdr.org.
OZfV or OZ: Österreichische Zeitschrift für Vermessungswesen und Photogramme-
trie, Vienna, Austria.
PE&RS or Photogr.Eng. or Ph.Eng.: Photogrammetric Engineering and Remote
Sensing, Journal of the American Society of Photogrammetry.
450 Completion of the references
PFG: Photogrammetrie, Fernerkundung und Geoinformation. Ε. Schweizerbart'sche

Verlagsbuchhandlung (Nägele und Obermiller), Stuttgart, Germany.
Ph.Rec. or The Photogrammetrie Record: The Photogrammetrie Society, London,
United Kingdom.
Phia: Photogrammetria, Elsevier Scientific Publishing Company, Amsterdam, The
Netherlands.
UNISURV: The school of surveying. The University of New South Wales, Sydney,
Australia.
VPK: Vermessung, Photogrammetrie, Kulturtechnik; Schweizerischer Verein für Ver-
messungswesen und Kulturtechnik, Schweizerische Gesellschaft für Photogram-
metrie etc., Zürich, Switzerland.
VT: Vermessungstechnik, Zeitschrift für Geodäsie, Photogrammetrie und Kartogra-

phie der (ehem.) DDR, VEB Verlag für Bauwesen, Berlin, (former) GDR.
VGI: Österreichische Zeitschrift für Vermessung und Geoinformation, Vienna, Aus-

tria.
Wiss. Arb. Geod. u. Photogr. TH Hannover: Wissenschaftliche Arbeiten der Fach-
richtung Vermessungswesen der Universität Hannover, Germany.
ZfV: Zeitschrift für Vermessungswesen, Konrad Wittwer Verlag, Stuttgart, Germany.
zfv: Zeitschrift für Geodäsie, Geoinformation und Landmanagement, Wißner-Verlag,
Augsburg, Germany.
ZPF: Zeitschrift für Photogrammetrie und Fernerkundung, Herbert Wichmann Verlag,
Karlsruhe, Germany.
Index
0,1,... recording of buildings, 322
12-parameter transformation, 228 relative, 244
3-line cameras, 55, 95, 146 relative orientation, 214
georeferencing, 277, 279 stereometric cameras, 166
gyro-stabilized, 154 strip triangulation, 267
intersection, 183 terrestrial laser scanning, 425
usage, 364 Active stereoviewing, 292
3D digitizer, 296 Adjustment by indirect observations, 441
3D digitizing, 316 Adjustment by the method of least squares,
441
A ADS from Leica Geosystems, 146
Abbe comparator, 303 Aerial photo interpretation, 93
Aberration, 65 Aerial photographs
Absolute orientation, 191, 219 cameras, 137
approximate values, 228 constraints, 77
error theory, 226 films, 91
terrestrial laser scanning stations, 422 flight planning, 131
Absorption, 403 overlap control, 138
Absorption filters, 78, 80 survey aircraft, 161
Accommodation, 286 Aerotriangulation, 246
Accuracy automated, 339
absolute orientation, 226 GPS- and IMU-assisted, 276
automated relative orientation, 337 Affine rectification, 378
block adjustment with independent mod- Affine transformation
els, 258 three-dimensional, 20
bundle block adjustment, 274 two-dimensional, 13
bundles in close range photogramme- Air photo-interpretation, 292
try, 284 Aliasing, 101
contours obtained indirectly from a All-round imaging, 177
DTM, 321 Amateur cameras, 170
distance, 244 Amateur pictures, 28, 187, 201
DTM model, 411, 415 Amplitude spectrum, 126
GPS-positioning, 152 Anaglyph system, 290
IMU-orientation, 152 Analogue photogrammetry, 297, 376, 381, 393
inertial navigation system, 152 Analytical photogrammetry, 300, 381, 394
least square matching (LSM), 326 Anchor point method (orthophoto), 380
location of the fiducial marks, 334 Animation, 399
monocomparator, 304 Aperture (critical), 66
of directly drawn contours, 320 Aperture stop, 59
orthophoto, 384 Automated recording of surfaces, 350
photogrammetry (in general), 34 Automatic fine measurement, 360
point determination in a stereopair, 238, critical situation, 362
317 Auxiliary image, 140
processing of lines, 318 Aviograph AG1, 299
452 Index
Β CCD metric camera, 429

Base, 31 CCIR/PAL, 172
Base/distance ratio, 33, 295 Central projection
Baselength (terrestrial), 174 in three-dimensional space, 21
Basic lens equation, 58 of a plane, 24
Beam splitter, 104 of the straight line, 29
Best-fitting surface to a point cloud, 407 Central shutter, 139
Bilinear interpolation Centre of perspective, 21
grey values, 42 Chain code, 37
Bilinear transformation Chained plane similarity transformation, 250
anchor point method, 381 Chained spatial similarity transformation, 258
fiducial marks, 74 Characteristic curve (films), 82
in a reseau, 169 Circle of confusion, 58
Binary images, 38 Circular graduated filter, 71
Binomial filter, 123 Close range photogrammetry, 163, 173, 197,
Black and white 282,315
digital images, 38 CMOS technology, 104
infra-red image, 88 Collinearity condition, 21
photography, 79 derivation, 436
Block adjustment differential, 438
by independent models, 248 normalized image, 344
terrestrial laser scanning stations, 421 similarity transformation, 437
Blooming, 104 Colour films, 87
Blunder detection Colour imaging
bundle block adjustment, 264
CCD camera, 104
Body coordinate axes, 151
Colour infra-red films, 87
Breaklines, 311
Colour infra-red photography, 84
Brightness enhancement
Colour photography, 84
in digital images, 111
Colour reversal process, 85
Building model, 316
Colour values (film scanning), 107
Bundle block adjustment, 247, 269
Complementary colour, 78
in close range photogrammetry, 282
Computer vision
with additional parameters, 274
with self-calibration, 274 disparity map, 343
with unknown interior orientation, 274 processing for plan, 353
relative orientation, 201
c Connectivity relationships (in pixel images),
37
CAD model, 314, 367, 396
Calibration Contour interval, 320
certificate, 54, 142 Contour lines
of a film scanner, 110 flat land, 311
of photogrammetric cameras, 50 forested areas, 310
on-the-job, 56, 284 Contrast
with a test field, 55 definition, 68
Camera configurations (terrestrial), 176 Contrast transfer function, 68, 90, 98
Camera constant, see Principal distance Control points, 23
Camera shutter, 139 automated location, 339
Cardan axes, 15 block adjustment, 259
CCD area array sensor, 96 groups of, 264
CCD cameras, 95, 145, 172 in high mountains, 228, 267
CCD linear arrays, 95 o f a G I S , 341
Index 453
retro-reflecting targets, 421 Digital topographic models, 316

Convergence of the meridian, 281, 407 Digitizing analogue images, 106
Convolution operator, 123 Direct Linear Transformation (DLT), 187,
Coplanarity condition, 192, 198, 202 444
differential, 445 Disparity map, 343
Copying with contrast control, 91 DMC from Intergraph, 144
Correlation Double images
3-line camera, 365 aerial photograph, 155
among model coordinates, 219 orthophoto, 382, 383
area based matching, 332 Drift correction, 140
feature based matching, 331, 337 Dynamic range
in the subpixel region, 326 CCD cameras, 105
of more than two images, 332 film, 81
radiometric information, 352 filmscanner, 109
two-dimensional, 329 image processing, 111
Correlation algorithms, 323
Correlation coefficient, 324 Ε
Corresponding points Earth curvature correction
identifying, 349 airborne laser scanning, 407
Critical cylinder, 211,213 dependent on the type of objective, 237
Cross-ratio, 29 for near-vertical photographs, 235
Cycle slips, 148 horizontal photographs, 233
image coordinates, 236, 280
D with tangential coordinate system, 281
Datum transformation, 404 Edge extraction, 125, 356
Deformation Emulsion carrier, 72
metric image, 367 Energy function (snakes), 354
plane objects, 370 Entrance pupil, 49
Deformation of the photogrammetric model, Epipolar geometry, 337, 341
215 in normalized images, 343
Density (Film), 81 in three images, 349
Depth of field, 58, 174 in tilted metric photographs, 345
Design matrix, 441 using projective geometry, 348
Detector spacing, 100 Epipolar plane, 342
Determination of heights Epipole, 342
automatic, 350 Estimation, see Adjustment
computer assisted, 310 Euclidean metric, 37
direct, 310 Exit pupil, 49
indirect, 311 Expansion
Differential GPS (DGPS), 148 grey values, 118
Diffraction blurring, 62, 63, 66 Exposure measurement, 82
Diffraction disc, 63 Exterior orientation, 23
Digital camera, see CCD camera
Digital image, 36 F
colour image, 38 Förstner operator, 331
Digital image processing, 35, 110, 323 horizontal shifts, 359
Digital object models, 316 Fa9ade evaluation, 315
Digital semi-metric cameras, 172 False colour film, 84
Digital terrain model (DTM), 311, 316 Far limit, 58
from airborne laser scanning, 407 Fiducial centre, 21
454 Index
Fiducial marks, 21, 140 Geometrical resolution (laser scannig and

automated location, 334 photogrammetry), 417
number and arrangement, 74 Geomorphological quality
Fiducial-mark cameras, 171 DTMs derived from laser scanned data,
Field completion, 309 416
Field of view, 135, 174 Georeferencing, 192
Film camera, 54 airborne laser scanning, 404
Film deformations, 72 direct, 181, 276
Film developing machines, 93 direct (3-line camera), 364
Film scanner, 107 indirect, 181
Film sensitivity, 82, 106 terrestrial laser scanning, 420
Films for aerial photography, 91 Gimbal axes, 15
Filter, 77 Glass plates, 72
"cut-off" or edge, 79 Global radiation, 159
Filtering GLONASS, 147
elimination of noise, 124, 128, 408 GPS (Global Positioning System), 147
frequency domain, 125 GPS time, 149
high-pass, 128 GPS- and IMU-assisted aerotriangulation, 276
low-pass, 128 Gradation, 81
spatial domain, 122 Granularity, 89, 106
Five parameter transformation, 76 Grey values
Fixed base stereometric cameras, 174 bilinear interpolation, 42
Fixed-focus cameras, 61 Grey values (film scanning), 107
Flat-bed scanners, 109 Grey wedge, 81, 142
Flight planning, 131 Grid model (errors in approximating), 385
digital line cameras, 136 Gruber points, 209
Flying height Gyro-stabilized platforms, 153
flight planning, 134 Gyros (gyroscopes), 150
influence on accuracy, 240, 275, 319
FMC (forward motion compensation), 157 Η
Focussing, 58 Haze light, 160
aerial metric cameras, 135 Height accuracy, see Accuracy
terrestrial, 174 Height control elements (airborne laser scan-
Forward overlap, 134 ning), 405
Fourier series, 97 Height control point, 191, 220, see Control
Frame cameras, 171 point
Frame grabber, 103, 172 chains of, 265, 277
Free net adjustment, 225 Height exaggeration, 295
terrestrial photogrammetry, 283 Hierarchy of axes, 15, 278, 432
Full control point, 191, 220 High-pass filter, 128
Fundamental matrix, 203 Histogram equalization, 114
epipolar rays, 348 Histogram normalization, 114
with additional contrast enhancement,
G 120
GALILEO, 147 Homogeneous coordinates, 44
Gaussian filter, 123 between two planes, 440
Geo-information system, 316 Direct Linear Transformation (DLT), 444
orthophoto, 366 Horizontal (or plan) control point, 191, 220
Geometric models, 315, see Topographic Horizontal parallax, 287
models Human vision, 296
Index 455
Hyperbolic paraboloid, 42 Κ
Hyperfocal distance, 60 Knots (nautical miles per hour), 162
I
L
Laplace operator, 124, 356, 358
ICP (iterative closest point), 423
Laser data and photographic data, 426
Illuminance, 81
Laser Radar, 400
Illumination (aerial photogrammetry), 159 Laser scanner
Image compression, 129
on gyro-stabilized platforms, 153
based on wavelets, 130 Laser scanning, 400
Image contrast, 69 airborne, 400
Image coordinate refinement, 230 first pulse, 401
Image coordinate system, 21, 39, 53 last pulse, 401
Image data management, 306 short range, 428
Image distance, 48, 58 terrestrial, 419
Image edge (definition), 356 Least square matching (LSM), 326
Image line (definition), 356 Least squares estimation method, 326
Image motion, 155 Lens stereoscope, 289
Image motion compensation Level of Detail (recording of buildings), 314
aerial film cameras, 157 Lidar, 400
CCD area arrays, 158 Light fall-off from centre to edge of image,
line cameras, 159 70
Image pyramids, 128 correction, 119
Image repetition frequency, 306 Light sectioning, 429
Image scale, 27, 34, 134 Line camera, 95
influence on accuracy, 274 gyro-stabilized, 154
orthophoto, 375, 389 Line extraction, 356
relationship between map scale and photo Line jitter, 103
scale, 319 Line maps, 366
IMU (inertial measurement unit), 148 Line photogrammetry, 338
IMU information Location based services, 399
tangential coordinate system, 281 LoG operator (Laplacian of Gaussian), 357
Independent metric cameras, 166 extraction of lines, 358
Inertial navigation system shift in position, 359
Look-up table (LUT), 111
accuracy, 152
Low-pass filter, 128
Information on object form, 283
INS (inertial navigation system), 149
Intensity image Μ
Machine vision, 296
terrestrial laser scanner, 427
Map and image scale, 319
Interest operators, 330
Map sheet preparation, 309
Interference filter, 105
Medium format cameras, 170
Interior orientation, 23, 39, 47, 50
Metric
automated, 334
of raster data, 37
calibration certificate, 54
Metric aerial cameras, 137
CCD cameras, 102
resolution, 66
semi-metric camera, 169
Metric cameras, 47, 170
Intersection of rays in three dimensions, 181
digital, 38
Metric film cameras
J aerial, 137
JPEG compression, 130 terrestrial, 165, 169
456 Index
Metric image, 367 Objective lens distortion, 52

analogue, 71 correction, 56
digital, 38 Objective lenses (for aerial photogrammetry),
Mirror stereoscope, 289 135
Misalignments, 153 Observation equations, 441
Mixed pixels, 324 Opacity, 81
Mobile mapping system, 178, 284 Optical axis, 48
Model coordinate, 21 Optical gyroscope, 151
Model deformations, 217 Optical principal points, 47
Model formation (after relative orientation), Orientation functions, 277
199 Orientation procedures, 180
Modulation transfer function, 70, 98 single-stage, 188
filtering, 126 two-step, 189
Moment of exposure (synchronization), 140 with known exterior orientation, 181
Monochromatic spectral colours, 77 with unknown exterior orientation, 184
Monocomparator, 302 Orthochromatic black-and-white-film, 88
Monoplotting, 396 Orthogonal matrix, 11
laser scanning, 418 Orthogonality conditions
Moravec operator, 331 three-dimensional, 15
Mosaicking of orthophotos, 384 two-dimensional, 11
Moving average, 122 Orthophoto, 366
Multi-image correlation, 333 Orthophoto analysis
Multipath effect, 148 analogue and analytical, 393
Orthophoto production
analogue, 381
Ν
analytical, 381
Nadir point, 372, 376
curved objects, 380
Navigation plan, 136
tilted plane, 378
Near limit, 58
Orthoscopic effect, 292
Near vertical images, 133, 180 Overlap control, 138
Negative position of a metric image, 21
Neighbour relationships (in pixel images), 37 Ρ
Nodal points (optical), 47 Panchromatic black-and-white-film, 88
Normal atmosphere, 231 Parallaxes, 32
Normal case accuracy, 287
light sectioning, 429 Passive stereoviewing, 292
of aerial photogrammetry, 31 Pattern recognition, 323
of terrestrial photogrammetry, 164 Perspective centre, 50
Normal equations, 255, 273, 441 Perspective centres
Normalized images, 342, 344 in a spatial block adjustment, 256
derivation, 346 Photo models, 367, 396, 427
Photo scale, see Image scale
Ο Photo texture, 397
Object colours (reflected light), 160 Photo-electric effect, 93
Object contrast, 68 Photogrammetric processing
Object distance, 58 analogue, 308
Object model, 316 automatic, 323
Object modelling computer assisted, 309
terrestrial laser scanning, 423 computer vision based, 323
Objective, 48 Photogrammetry
with variable principal distance, 61 point measurement, 246
Index 457
Photographic emulsion, 71 Raypath, 160

resolution, 89 Real-time photogrammetry, 172
Photographic infra-red, 77 Recording in plan
Photomaps, 366 computer assisted, 308
Pixel, 36 semi-automated, 353
Pixel geometry, 36 Recording of buildings
Planar accuracy, see Accuracy computer assisted, 312
Planar control elements, 405 from laser scanner data, 411
Planar regions, 406 semi-automatic, 360
terrestrial laser scanning, 424 topology assisted, 312
Planarity (film) Reference matrix (correlation), 323
departures from, 76 Reflection
Planimetrie adjustment of a block, 248
laser beam, 403
Planning
Refraction correction
terrestrial photogrammetry, 173
for horizontal photographs, 233
Pneumatic flattening, 138
Point clouds, 422 for near-vertical photographs, 230
Polar coordinates Relative orientation, 193
airborne laser scanning, 404 y-parallaxes, 205
laser scanner strip, 406 alternative formulation, 201
terrestrial laser scanning, 421 automated, 335
Polarization, 291 critical surfaces, 210
Primary colours error theory, 213
additive, 78 flat ground, 209
subtractive, 78 Gauss-Helmert, 200
Principal distance, 21, 142 highly tilted photographs, 197
Principal planes, 47 line-based, 338
Principal point, 21 mountainous country, 206
of autocollimation, 49, 54 of near-vertical photographs, 193
of best symmetry, 52, 54 using rotations only, 195
offset, 167 Remote sensing, 400
Projective photogrammetry, 28, 29, 40, 201 Rendering, 314
three-dimensional, 187 laser scanner data, 407
Projective rectification, 40, 376, 378 Resampling, 40
Proper orthogonal matrix, 11 correction of distortion, 57
Pseudo ranges, 147 correlation, 329
Pseudoscopic effect, 292
two normalized images, 344
virtual correction image, 237
Q Resection in three dimensions, 185
Quality management Resolution (CCD camera)
absolute orientation, 228 geometric, 106
relative orientation, 215 radiometric, 106
R Resolving power, 63
Reseau cameras, 76, 169, 171 actual, 99
Reseau scanning, 110 optical, 64,67, 89
Radial displacement, 374 photographic emulsions, 89
orthophoto, 376 theoretical, 99
Range image, 417 total, 89
Range measurement, 401 RGB image, 38, 292
Raster measurement, 311, 351 Roaming, 307
458 Index
Robust parameter estimation numerical problems, 435

of a DTM, 408 two sets of rotation angles, 435
Root mean square error, 442 Spatially related information system, 316
Rotation matrix, see Spatial rotation matrix Spectral reflectance, 160
inversion, 12 Spectral sensitivity
terrestrial photogrammetry, 164 CCD cameras, 103
three-dimensional, 15 films, 87
two-dimensional, 10 Spline function, 353
Rotations Standard deviation
successive, 17, 432 estimation, 442
Rule of propagation of errors, 443 Stereo-orthophoto, 393
Run-length encoding, 130 Stereocomparators, 300
Stereographic images
s height exaggeration, 295
Sampling theorem, 100, 107 Stereometer, 293
Search image, 324 Stereometric cameras, 165
Self-calibration, 56 Stereophotogrammetry
Semi-metric cameras, 169, 170 3-line camera, 136
Sensor flight planning, 131
chemical, 47 Stereoprocessing, 180, 286
electronic, 47 analogue, 297
opto-electronic, 93 analytical, 300
Sensor orientation, 280 digital, 290, 306
direct, 276 topology assisted, 312
indirect, 421 universal analytical, 304
integrated, 276 Stereoscopic acuity, 287
integrated (3-line camera), 365 Stereoscopic images, 288
integrated (laser scanning), 405 analogue, 289
Serial metric cameras, 138 digital, 290
Short range laser scanner, 428 Stereoscopic measurement, 297
Side overlap, 134 Stereoscopic model, 180
Signal-to-noise ratio, 105 deformation, 215
Similarity transformation Stereoscopic observation systems, 286
chained plane, 250 Stereoscopy, 287
chained spatial, 258 significance, 292
three-dimensional, 20 Still video cameras, 173
two-dimensional, 14 Storage space
Single image photogrammetry for a metric image, 107
digital line camera, 136 image pyramids, 128
flight planning, 134 Strapdown inertial system, 151, 152
Single image processing Street model, 316
of a flat object, 393
Superimposed stereoscopic images, 290
of curved object surfaces, 394
Superposition system, 312
Skylight, 159
Surface model, 316
Snakes, 353
Survey aircraft, 161
Spatial block adjustment, 256 Systematic errors, 448
Spatial bundle of rays, 269 compensation, 274
Spatial frequency, 69
Spatial rotation matrix, 14, 432 Τ
in the body coordinate system, 278 TDI (time delay integration), 158
linearized, 220 Temporal frequency, 69
Index 459
Terrestrial laser scanner

technical data, 420
Three-dimensional image, 288
Three-line camera, see 3-line camera
Tie points, 249
automated determination, 335
number, 264
Tolerance, 244
Topographic information system, 316
Topographic models, 453
Topology
buildings, 312
roof surfaces, 314
Transform function
linear, 111
logarithmic, 111
Transmission of filters, 80
Transparency, 81
Types of film, 88
u
Uncertainties of definition
related to the image, 242
related to the object, 243
Undersampling, 101
V
Vertical parallaxes, 288
Videocameras, 103
Virtual correction image, 237
Virtual flyovers, 399
Virtual spatial image, 288, 293
Virtual walkthroughs, 399
Visibility analysis
orthophoto production, 382
Visible light, 77
Visualization
digital orthophotos, 366
dynamic, 399
photo models, 396
static, 399
VLL (vertical line locus), 351
w
Weber-Fechner law, 81
Weighted centroid method, 349
Weighting function for filtering of airborne
laser scanner data, 408
WGS84 (World Geodetic System 1984), 147
Ζ
Zero crossings, 357

(De Gruyter Textbook) Karl Kraus, Ian A. Harley, Stephen Kyle - Photogrammetry - Geometry From Images and Lase PDF

Uploaded by

Copyright:

Available Formats

(De Gruyter Textbook) Karl Kraus, Ian A. Harley, Stephen Kyle - Photogrammetry - Geometry From Images and Lase PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(De Gruyter Textbook) Karl Kraus, Ian A. Harley, Stephen Kyle - Photogrammetry - Geometry From Images and Lase PDF

Uploaded by

Copyright:

Available Formats

de Gruyter Textbook

G Berlin · New York

First English edition:

© Printed on acid-free paper which falls within the guidelines

Bibliographic information published by the Deutsche Nationalbibliothek

Volume 1 was additionally translated into several languages, including Serbocroatian,

In the discussion of stereo-plotting instruments most attention is given to digital soft-

Norbert Pfeifer Vienna, Summer 2007

Kraus, Karl: Photogrammetry, Volume 2, Advanced Methods and Appli-

Volume 2 is a completely separate textbook and is currently out of print. It covers

2 Preparatory remarks on mathematics and digital image processing . . . 10

3 Photogrammetric recording systems and their application 47

3.1.6 Light fall-off from centre to edge of image 70

3.6.2 Image compression 129

4 Orientation procedures and some methods of stereoprocessing 180

4.3.2.1 Gauss-Helmert model of relative orientation 200

5 Photogrammetric triangulation 246

5.6 Accounting for Earth curvature and distortions due to cartographic

6 Plotting instruments and stereoprocessing procedures 286

6.8.3.1 Near-vertical photographs with 60% forward over-

7 Orthophotos and single image analysis 366

7.5 Photo models 396

8 Laser scanning 400

Completion of the references 449

• numbers—coordinates of separate points in a three-dimensional coordinate sys-

That branch of photogrammetry which starts with conventional photographs and in

'See PE&RS 58, Copy 1, pp. 49-115, 1992.

If the photographs are taken with specialized cameras, photogrammetric processing is

1.3 Some remarks on historical development3

Meydenbauer had, however, been preceded in 1849 by the Frenchman Laussedat, a

In considering the history of photogrammetry the work of Scheimpflug in Austria

Efforts were therefore directed towards developing a means of continuous measure-

including both Pulfrich, one of the greatest contributors to photogrammetric instrumen-

Analytical photogrammetric triangulation is a method, using numerical data, of point

Subsequently, stereoplotters were equipped with devices to record model coordinates

Photogrammetry has progressed as supporting sciences and technologies have supplied

The introduction of digital cameras into a photogrammetric system allows automation,

Although development continues apace, photogrammetry is a mature technology with

Further reading. Adams, L.P.: Fourcade: The centenary of a stereoscopic method

Preparatory remarks on mathematics

2.1 Preparatory mathematical remarks

2.1.1 Rotation in a plane, similarity and affine transformations

Representing matrices and vectors by means of bold symbols this becomes:

Figure 2.1-1: Plane rotation Figure 2.1-2: Introduction of the unit

Properties of the rotation matrix R

iTi = cos2 α + sin2 α = 1 = r 2 , + r\x

Exercise 2.1-1. Consider a rectangle in an xy system, the vertices being transformed

Inverting the rotation matrix R

Rewriting this result:

A transformation with a non-orthogonal matrix (see the first numerical example) is

The affine transformation is of the form:

of which the solution is

One obtains a similarity transformation by replacing the nonorthogonal matrix A of

2.1.2 Rotation, affine and similarity transformations in

f X \ / c o s (ZxX) cos (ZyX) cos {ZzX) \ f x\

Figure 2.1-3: Rotation in three-dimensional space

That is, a rotation in three dimensions is prescribed by three independent parameters.

Figure 2.1-4: Rotations about the axes in gimbals

primary secondary tertiary