See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220611764

Object detection using image reconstruction with PCA

Article in Image and Vision Computing · January 2009

DOI: 10.1016/j.imavis.2007.03.004 · Source: DBLP

CITATIONS READS

82 1,749

2 authors, including:

Olac Luis Fuentes

University of Texas at El Paso
159 PUBLICATIONS 1,888 CITATIONS

SEE PROFILE

All content following this page was uploaded by Olac Luis Fuentes on 11 October 2019.

The user has requested enhancement of the downloaded file.

 ARTICLE IN PRESS

Image and Vision Computing xxx (2007) xxx–xxx

www.elsevier.com/locate/imavis

Object detection using image reconstruction with PCA

Luis Malagón-Borja a, Olac Fuentes b,*

a
Computer Science Department, I.N.A.O.E., Tonantzintla, Puebla 72840, Mexico
b
Computer Science Department, University of Texas at El Paso, El Paso, TX 79968, USA

Received 25 January 2006; received in revised form 8 November 2006; accepted 5 March 2007

Abstract

In this paper, we present an object detection system and its application to pedestrian detection in still images, without assuming any a
priori knowledge about the image. The system works as follows: in a first stage a classifier examines each location in the image at different
scales. Then in a second stage the system tries to eliminate false detections based on heuristics. The classifier is based on the idea that
Principal Component Analysis (PCA) can compress optimally only the kind of images that were used to compute the principal compo-
nents (PCs), and that any other kind of images will not be compressed well using a few components. Thus the classifier performs sep-
arately the PCA from the positive examples and from the negative examples; when it needs to classify a new pattern it projects it into
both sets of PCs and compares the reconstructions, assigning the example to the class with the smallest reconstruction error. The system
is able to detect frontal and rear views of pedestrians, and usually can also detect side views of pedestrians despite not being trained for
this task. Comparisons with other pedestrian detection systems show that our system has better performance in positive detection and in
false detection rate. Additionally, we show that the performance of the system can be further improved by combining the classifier based
on PCA reconstruction with a conventional classifier using a Support Vector Machine.
 2007 Published by Elsevier B.V.

Keywords: Object detection; Pedestrian detection; Principal Component Analysis; Support Vector Machines

1. Introduction Many object detection systems that have been developed

focus on face detection. An early very successful system
The object detection problem can be seen as a classifica- was presented by Rowley et al. [20], which consists of an
tion problem, where we need to distinguish between the ensemble of neural networks and a module to reduce false
object of interest and any other object. In this paper, we detections. Similar example-based face detection systems
focus on a single case of the object detection problem, have been developed by Sung and Poggio [22], Osuna
detecting pedestrians in images. et al. [17], and Yang et al. [28].
Pedestrian detection is more difficult than detecting Most pedestrian detection systems use motion informa-
many other objects due to the fact that people can show tion, stereo vision, a static camera or focus on tracking;
widely varying appearances when the limbs are in different important works include [5,8,10,29]. Papageorgiou has
positions. In addition, people can dress in clothes with reported a system to detect pedestrians in images, without
many different colors and types. For the characteristics of restrictions in the image, and without using any additional
the pedestrian class we need a robust method that can learn information [15,16,9]. It uses the wavelet template to repre-
the high variability in the class. sent the image and a Support Vector Machine (SVM) to
classify. The system has been improved in [12,4], detecting
pedestrians through the detection of four components of
* the human body: the head, legs, left arm and right arm.
Corresponding author.
E-mail addresses: jmb@inaoep.mx (L. Malagón-Borja), ofuentes@ Viola, Jones and Snow developed a system to detect pedes-
utep.edu (O. Fuentes). trians from image sequences. This system uses a large set of

0262-8856/$ - see front matter  2007 Published by Elsevier B.V.

doi:10.1016/j.imavis.2007.03.004

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

2 L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx

simple filters as features, and then applies the Adaboost at least twice times and eliminating the detections that
algorithm to generate a cascade of classifiers [26]. overlap.
We present an object detection system to detect pedestri- Fig. 1 shows the complete process to detect pedestrians
ans in gray level images, without assuming any a priori in an image, starting with the gray level image and finishing
knowledge about the image. The system works as follows: with the image with the detected pedestrians.
in a first stage a classifier based on Principal Component
Analysis (PCA) examines and classifies each location in 2.2. Stage 1: a classifier based on image reconstruction with
the image at different scales. Then, in a second stage, the PCA
system tries to eliminate false detections based on two
heuristics. In this stage, we present a classifier that decides if an
The system uses PCA as a classification tool; the main image of size 105 · 45 belongs or does not belong to the
idea is that PCA can compress optimally only the kind of pedestrian class. This classifier is based on doing image
images that were used to do the PCA, and that any other reconstruction using PCA and comparing the recon-
kind of image will not be compressed well in a few attri- structed with the original images. First, the reasons to work
butes, so we do PCA separately for positive and negative with both the gray level image and the edge image are
examples; when a new pattern needs to be classified we explained, later we explain how the reconstruction of an
compare the reconstruction made by both sets of principal image is performed using PCA, and finally, we present
components (PCs). In order to improve the performance of the way in which a classifier can use these reconstructions
the classifier we have used the edge image as additional to decide if an image belongs or does not belong to the
information for it. Additionally, we show that the perfor- pedestrian class.
mance of the system can be further improved by combining
the classifier based on PCA reconstruction with a conven- 2.2.1. Edge images
tional classifier using a Support Vector Machine. Because pedestrians appear in many colors and different
The organization of the reminder of this paper is as fol- textures, it is not advisable to use characteristics based on
lows: Section 2 presents a detailed description of the sys- color or texture to do pedestrian detection. For this reason,
tem. In Section 3 the performance of our system, and a we have chosen to use the edge image with the idea of
comparison with similar systems are presented. Section 4 obtaining the typical silhouette of a pedestrian and to elim-
reports conclusions and possible directions for future work. inate useless information for the classifier.
The edge images were computed using x and y Sobel fil-
ters, this edge image serves as complementary information
2. The detection system to the gray level image and it allows the classifiers to obtain
more data to decide if an image is a pedestrian or not.
2.1. Overview of system architecture In Fig. 2 we can see examples of the corresponding edge
images of some pedestrian gray level images. In these
The system works scanning the whole image by means images we can observe that although the gray level images
of a detection window of size 105 · 45 pixels; the window are very different in color and background, the edge images
is shifting with two pixel jumps to accelerate the process present fewer changes from one image to another. This is
without losing much information from one window to the reason why the edge images are very important to aid
another. We need a classifier that decides for each window in the classifier’s task.
if it contains a pedestrian or not. The construction of the
classifier is the most complicated stage, we have created a
2.2.2. Image reconstruction with PCA
classifier based on image reconstruction with PCA, this
Principal Component Analysis is a popular technique
classifier uses the edge image in addition to the gray level
for data compression and has been successfully used as
image.
an initial step in many computer vision tasks, including
The scanning of the whole image is part of an iterative
face recognition [2,23] and object recognition [14]. The for-
process where the image is resized several times to achieve
mulation of standard PCA is as follows. Consider a set of
multi-scale detection. For our experiments, the image has
m images, each of size r · c. Each image Ii is represented by
been scaled from 0.26 up to 1.35 times its original size, with
a column vector vi of length rc. The mean object of the set
increases of approximately 17% in every cycle, thus the
is defined by
image is processed at the following 12 different scales:
0.26, 0.3, 0.35, 0.4, 0.47, 0.55, 0.64, 0.74, 0.86, 1, 1.17 1 Xm

and 1.35, this implies that pedestrians of sizes between l¼ vi

m i¼1
78 · 33 and 404 · 173 pixels will be detected by the system.
When the system has finished examining the image in all C, the covariance matrix, is given by
scales, a second process eliminates some detections that are X
m
T
believed to be false detections. The form in which this pro- C¼ ðvi  lÞðvi  lÞ
cess works is eliminating the detections that do not repeat i¼1

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx 3

Fig. 1. Architecture of the system for pedestrian detection in images.

Fig. 2. Edge images. The edge images eliminate information about color and texture, therefore they present less variation among pedestrians.

The principal components are then the eigenvectors of (see [3]). In our system we compute these eigenvectors using
C. These eigenvectors can be computed in several ways. Per- the implementation provided by Matlab, which is based on
haps the easiest one is to solve the generalized eigenvector the QZ algorithm [13].
problem using the QZ algorithm or its variants [13]. It is If we sort the eigenvectors by decreasing order of their
also common to formulate the problem as that of finding corresponding eigenvalues, a projection onto the space
the basis vectors that minimize the reconstruction error defined by the first k eigenvectors (1 6 k 6 rc) is optimal
and then solve it using standard least-squares techniques with respect to information loss. Let P be the matrix whose

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

4 L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx

columns are the first k eigenvectors of C. The projection of From this fact we can create a classifier based on image
an image u into this eigenspace is given by reconstruction with PCA, which decides if an image
p ¼ P ðu  lÞ belongs or does not belong to the pedestrian class. The
algorithm to do this classification is the following:
When we speak of reconstructing an image with PCA, what Before doing any classification:
we understand is to project the image into the PCs, and
from this projection, try to recover the original image. 1. Perform PCA on the set of pedestrian gray level
Thus the reconstructed image u 0 is images to obtain the projection matrix Pgp and the mean
u0 ¼ P T p þ l ¼ P T P ðu  lÞ þ l lgp.
2. Perform PCA on the set of pedestrian edge images to
Let P be the matrix whose columns are the first k eigenvectors obtain the projection matrix Pep and the mean lep.
of C. The projection of an image into this eigenspace is given by 3. Perform PCA on the set of non-pedestrian gray level images
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X ffi
2 to obtain the projection matrix Pgn and the mean lgn.
d ¼ ju  u0 j ¼ ðui  u0i Þ
4. Perform PCA on the set of non-pedestrian edge images
In general, the more PCs we use to obtain the projection, the to obtain the projection matrix Pen and the mean len.
less information loss we will have, thus the reconstruction of
the image will be more accurate. Also, the more similar u is When we want to classify a new gray level image g:
to the images used to generate P, the better the reconstruc-
tion will be for a fixed number of eigenvectors. 1. Obtain the edge image e from g.
2. Do four reconstructions:
2.2.3. Classification using reconstruction (a) rgp ¼ P Tgp P gp ðg  lgp Þ þ lgp
By definition, PCA looks for the set of PCs that best (b) rep ¼ P Tep P ep ðe  lep Þ þ lep
describe the distribution of the data that are being analyzed. (c) rgn ¼ P Tgn P gn ðg  lgn Þ þ lgn
Therefore, these PCs are going to preserve better the infor- (d) ren ¼ P Ten P en ðg  len Þ þ len
mation of the images from which PCA was performed, or 3. Obtain reconstruction errors:
of those that are similar. Thus, if we have a set of PCs that (a) dgp = jrgp  gj
were obtained from a set of pedestrian images only, these (b) dep = jrep  ej
must reconstruct better the images of other pedestrians than (c) dgn = jrgn  gj
any other type of images, and viceversa, if we have a set of (d) den = jren  ej
PCs obtained from images of anything except pedestrians, 4. Let total error be given by dt = dgn + den  dgp  dep
the reconstruction of the pedestrian images will not be as 5. Classify the image according to the following criterion

good. We can observe this fact in Fig. 3, both for gray level Pedestrian; dt P 0
images and for edge images. classðgÞ ¼
Non-pedestrian; d t < 0

Fig. 3. Image Reconstruction with different sets of PCs. Row (a) shows the original images; row (b) shows the images reconstructed using 100 PCs
obtained from pedestrian images, and row (c) shows the images reconstructed using 100 PCs obtained from non-pedestrian images. We can see that, for
both the gray level images and the edge images, the pedestrian images are better reconstructed with the PCs obtained from pedestrian images (row b) than
with the PCs obtained from non-pedestrian images (row c). This does not happen with the non-pedestrian images, which are better reconstructed with the
PCs obtained from non-pedestrian images (row c).

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx 5

2.2.4. Adding a Support Vector Machine classifier The kernel function used for mapping the input space
The main feature of the first stage in the algorithm is its was a polynomial of exponent one. We used the implemen-
simplicity. The preprocessing phase requires to find the tation of SVM included in the WEKA environment [27].
edges in every image in the training set and then to com- The optimal separating hyperplane found by an SVM
pute four sets of principal components, while the classifica- algorithm for a particular training set is given by the vector
tion phase just requires to project the subimage to the four w and the scalar b. Thus a test example x is classified as
eigenspaces and back to the original spaces and apply a positive iff w Æ x  b P 0. Let wp and bp be the parameters
threshold. We will show that this simple method yields of the optimal separating hyperplane obtained using the
accurate results in the pedestrian detection task. training set of pedestrian and non-pedestrian images. Then
If higher training times are acceptable, we could use a the classification rule that combines our original classifier
potentially more accurate learning algorithm. We experi- and the SVM is:
mented using a Support Vector Machine as base classifier, 
Pedestrian; a1 d t þ ad ðwp  u  bp Þ P 0
using the gray levels of the original as well as the edge classðgÞ ¼
images as attributes. We will show in the experimental Non-pedestrian; otherwise
results section that, although the SVM-based classifier does
not perform as well as our classifier based on reconstruc- where a1 and a2 are positive weights that control the rel-
tion errors, a weighted combination of the outputs of both ative influence of both classifiers. Experimentally, we found
classifiers yields better results than either of them individu- that classification accuracy was not very sensitive to the
ally, albeit at a significantly higher computational cost. choice of weights as long as a1 was greater than a2, since
SVM is a learning algorithm developed by Vapnik [24], the first classifier is more accurate than the second. For
that is based on the method of structural risk minimization, the experiments we used a1 = 3 and a2 = 1.
which minimizes a bound on the generalization error. The
main idea is to construct a hyperplane as the decision sur- 2.3. Stage 2: reduction of false detections by means of
face in such a way that the margin of separation between heuristics
positive and negative examples is maximized. Instead of
constructing this hyperplane in the original input space, The output after classifying all the windows of the image
SVM uses a nonlinear kernel to project the original vari- in multiple scales still contains a significant number of false
ables into a higher-dimensional feature space, which yields detections, in this stage we present two heuristics that allow
a nonlinear decision surface in input space. This is a very to reduce the number of false detections by means of two
powerful feature, because it allows SVM to overcome the processes, namely, eliminating single detections and elimi-
limitations of linear boundaries. For more information nating nearby detections.
about this algorithm we refer the reader to [21,24].
Given that Support Vector Machines have proven to 2.3.1. Eliminating single detections
perform well over high dimensionality data, they have been As we can see in Fig. 4a, most of the pedestrians are
successfully used in many vision-related applications, such detected at multiple nearby positions and scales, while false
as face detection [17], 3D object recognition [19], and track- detections usually appear at a single position. This observa-
ing [1]. tion allows us to eliminate some false detections, eliminat-
In our work, the optimization algorithm used for train- ing detections that appear only once.
ing the support vector classifier is an implementation of Each detection found can be grouped with those detec-
Platt’s sequential minimal optimization algorithm [18]. tions whose centroid is inside the same neighborhood,

Fig. 4. Process to eliminate single false detections. In figure (a) we can see the original detections found by the classifier. In figure (b) each detection is
grouped with those detections whose centroid is in the same neighborhood. Finally in figure (c) we have the grouped detections composed by two or more
original detections.

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

6 L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx

2
obtaining a new set of detections which we will call WeightðheightÞ ¼ ðheight  50Þ
grouped detections, composed by one or more of the origi-
nal detections. Once we have the set of grouped detections There are very few cases where this heuristic does not
we will ignore the original detections, and we will eliminate work, and thus it allows to eliminate many false detections
the grouped detections composed by only one original when the classifier confuses the arms, the legs, or some
detection. We can see this process in Fig. 4. other object with a pedestrian. Fig. 5 shows an example
where several false detections are eliminated applying this
heuristic to the output of the classifier.
2.3.2. Eliminating clustered detections
If a window is identified correctly as a pedestrian, then it
is very likely that there are no pedestrians either above or 3. Experimental results
below it, and if there are pedestrians beside it, they cannot
be too overlapped. This heuristic allows us to eliminate As we explained in the previous section, we need a set of
nearby detections. With this purpose, we define a region pedestrian images and a set of non-pedestrian images to
around a detection which we are going to use to eliminate obtain the four sets of PCs from which we are going to per-
any detection whose centroid is inside this region. The size form the four reconstructions, and to train the SVM
of the region was defined empirically as 1.4 times the detec- classifier.
tion height upwards and downwards from the centroid and The pedestrian images were obtained from the MIT
between 0.5 and 0.75 times the detection width towards pedestrian database, which contains pedestrians in frontal
each side of the centroid. or rear views under different scene conditions. We con-
We know that when we have multiple detections we must verted these color images to gray level and we cropped part
choose only one to keep, but how do we make this decision? of the background to reduce the variation that exists
A reasonable way to choose is to maintain the grouped detec- among pedestrian images, finally obtaining a set with 500
tion composed by the most original detections; nevertheless, gray level images.
we observe that usually the biggest detections were the cor- For the negative images we obtained two different sets,
rect ones, due to the fact that arms, legs and head are often the first set of negative images had 2315 images that were
confused with pedestrians, so when we need to decide among obtained randomly from a set of 90 images of scenery that
a set of detections that are in the same region, we must con- did not contain any pedestrian and was used in the classi-
sider the number of times that they have been detected orig- fier based on image reconstruction with PCA. The second
inally as well as the size of the detected regions. set had 2248 images obtained in a bootstrap manner [22]
To achieve this, the detections that compose a grouped and it was used to train the SVM classifier.
detection are weighted by their height, then the grouped For the classifier based on image reconstruction we used
detection with the greatest Preference, according to the fol- 200 PCs in each set, that contain between 75% and 85% of
lowing formula is chosen. the variance, to do the reconstructions. It was observed
Preference ¼ Detections  WeightðheightÞ that this number of PCs allowed a good classification.
The SVM classifier used as kernel a Radial Basis Func-
where Detections is the number of original detections tion (RBF) with r = 0.00015 and C = 10,000. The SVM
that compose the grouped detection that we are evaluating was trained with the projection of the 2748 pedestrian
and Weight is a function that determines the value that and non-pedestrian images onto the first 200 PCs of six dif-
each detection has, according to the height of the grouped ferent PCs sets, each one obtained from a different set of
detection, and is given by the formula: images; these sets are the following:

Fig. 5. Process to eliminate nearby false detections. Figure (a) shows the detections found by the classifier. Figure (b) shows the grouped detections. In
figure (c) the grouped detections with the greatest Preference have been preserved and the nearby grouped detections have been eliminated.

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx 7

1. The 500 pedestrian gray level images. The system was tested with a database containing 204
2. The 2248 non-pedestrian gray level images. pedestrian images in frontal or rear view to determine the
3. The 2748 pedestrian and non-pedestrian gray level pedestrian detection rate; these images were not used
images. before. The false detection rate was obtained by running
4. The 500 pedestrian edge images. the system over a database with 17 images that did not con-
5. The 2248 non-pedestrian edge images. tain any pedestrian; by running the system over these 17
6. The 2748 pedestrian and non-pedestrian edge images. images 4,850,103 windows were classified.
In general, the performance of any object detection sys-
So each image of size 105 · 45 pixels was described by tem shows a tradeoff between the positive detection rate
1200 attributes (200 for each projection). and the false detection rate. We ran the system over the test
images at several different thresholds. The results were
plotted as a Receiver Operating Characteristic (ROC)
curve, given in Fig. 6. We can see that the best individual
classifier is the classifier based on image reconstruction, this
is better than the best reported in the literature of pedes-
trian detection in systems than do not assume any a priori
scene structure or use any motion information. The SVM
shows the second best performance among individual clas-
sifiers, and the best overall performance is obtained by the
ensemble that outputs the weighted combination of the
classifier based on reconstruction error and the SVM.
The curve indicates that the system can achieve a detection
rate of 99.02% with one false positive every 53,890 win-
dows examined, or if we want a more conservative system,
it can achieve a detection rate of 90.69% with one false
detection every 808,351 windows examined. Fig. 7 shows
the result of applying the system to sample images in clut-
tered scenes under different conditions.

4. Conclusions and future work

In this paper, we have presented an object detection sys-

Fig. 6. ROC curves comparing the performance of our classifiers versus tem for static images, without assuming any a priori knowl-
the best reported in the literature. The detection rate is plotted against the edge, applied to the specific problem of locating pedestrians
false detection rate measure on logarithmic scale. in cluttered gray level images.

Fig. 7. These images demonstrate the capability of the system for detecting people in still images with cluttered backgrounds.

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
 ARTICLE IN PRESS

8 L. Malagón-Borja, O. Fuentes / Image and Vision Computing xxx (2007) xxx–xxx

Our system is able to detect frontal and rear views of [5] Sumer Jabri, Zoran Duric, Harry Wechsler, Azriel Rosenfeld,
pedestrians, and usually it can also detect side views of Detection and location of people in video images using adaptive
fusion of color and edge information, in: Procedings of International
pedestrians despite not being trained for this task. Conference on Pattern Recognition, vol. 4, 2000, pp. 4627–4631.
The success of PCA for pedestrian detection comes from [6] Fernando De la Torre, Michael Black, Robust principal component
its capability to capture most of the information about the analysis for computer vision, in: Proceedings of IEEE Conference on
objects of interest by using both intensity and edge images. Computer Vision and Pattern Recognition, 2001, pp. 362–369.
This allows to distinguish between a pedestrian image and [7] Anuj Mohan, Object detection in images by components, A.I. Memo
1664, Center for Biological and Computational Learning, MIT,
any other image in the huge universe of non-pedestrian Cambridge, MA, 1999.
images. The success of the ensemble of classifiers is due [8] Anuj Mohan, Constantine Papageorgiou, Tomaso Poggio, Example-
to the low correlation in errors between both classifiers based object detection in images by components, IEEE Transactions
and the capability of each classifier to learn the pedestrian on Pattern Analysis and Machine Intelligence 23 (4) (2001) 349–361.
class accurately. [9] C.B. Moler, G.W. Stewart, An algorithm for generalized matrix
eigenvalue problems, SIAM Journal on Numerical Analysis 10 (2)
An interesting possibility for future work is to use alternate (1973) 241–256.
projection spaces to derive the attributes used by the SVM. In [10] Shree K. Nayar, Sameer A. Nene, Hiroshi Murase, Real-time 100
particular, Fisher’s linear discriminant (FLD), which has been object recognition system, in: IEEE International Conference on
used successfully in the face recognition domain [2], could be Robotics and Automation (ICRA), vol. 3, Minneapolis, MN, 1996,
used to derive those features. This has the potential to provide pp. 2321–2325.
[11] Michael Oren, Constantine Papageorgiou, Pawan Sinha, Edgar
improved results because, in contrast to principal compo- Osuna, Tomaso Poggio, Pedestrian detection using wavelet templates,
nents, FLD uses class information to find a projection that in: Procedings of IEEE Conference Computer Vision and Pattern
separates examples of different classes. Recognition, 1997, pp. 193–199.
The current system does not work as well for side views [12] Michael Oren, Constantine Papageorgiou, Pawan Sinha, Edgar
of pedestrians as for pedestrians in frontal or rear views. Osuna, Tomaso Poggio, A trainable system for people detection,
in: Procedings of Image Understanding Workshop, 1997, pp. 207–
To solve this, we can add side views of pedestrians to the 214.
training set, or we can create an additional part of the sys- [13] Edgar Osuna, Robert Freund, Federico Girosi, Training support
tem that could be specialized for these views. vector machines: an application to face detection, in: IEEE Confer-
Another way to improve the system’s performance is to ence on Computer Vision and Pattern Recognition, 1997, pp. 130–
obtain more positive and negative examples for training. 136.
[14] Constantine Papageorgiou, Object and pattern detection in video
We only use 500 positive examples and 2315 negative sequences, Master’s thesis, M.I.T., Cambridge, MA, 1997.
examples, while other works in object detection use around [15] John C. Platt, Sequential minimal optimization: a fast algorithm for
2000 positive examples and 10,000 negative examples. training support vector machines, in: Bernhard Scholkopf, Christo-
The framework described here is applicable to other pher J.C. Burges, Alex J. Smola (Eds.), Advances in Kernel Methods:
domains besides pedestrian detection; it can be generalized Support Vector Learning, MIT Press, Cambridge, MA, USA, 1999,
pp. 185–208.
to the detection of several different types of objects, such as [16] Massimiliano Pontil, Alessandro Verri, Support vector machines for
faces, vehicles, and others. A promising direction for future 3D object recognition, IEEE Transactions on Pattern Analysis and
work is to apply the method presented in this paper in a Machine Intelligence 20 (6) (1998) 637–646.
component-based approach. This approach has shown bet- [17] Henry A. Rowley, Shumeet Baluja, Takeo Kanade, Neural network-
ter performance in pedestrian detection (see [12,4]) than a based face detection, IEEE Transactions on Pattern Analysis and
Machine Intelligence 20 (1) (1998) 23–38.
similar full-body pedestrian detector [15,16,9]. Also, we [18] Bernhard Scholkopf, Alexander J. Smola, Learning with Kernels:
intend to investigate if the classification approach based Support Vector Machines, Regularization, Optimization, and
on reconstruction error with PCA can be applied to other Beyond, MIT Press, Cambridge, MA, USA, 2001.
problems in computer vision and other areas. [19] Kah-Kay Sung, Tomaso Poggio, Example-based learning for view-
based human face detection, IEEE Transactions Pattern Analysis and
Machine Intelligence 20 (1) (1998) 39–51.
References [20] Matthew Turk, Alex Pentland, Eigenfaces for recognition, Journal of
Cognitive Neuroscience 3 (1) (1991) 71–86.
[1] Shai Avidan, Subset selection for efficient SVM tracking, in: [21] Vladimir N. Vapnik, The Nature of Statistical Learning Theory,
Proceedings of IEEE Conference on Computer Vision and Pattern Springer, New York, USA, 1995.
Recognition, 2003. [22] Paul Viola, Michael Jones, Daniel Snow, Detecting pedestrians using
[2] Peter N. Belhumeur, Joao P. Hespanha, David Kriegman, Eigenfaces patterns of motion and appearance, International Journal of Com-
vs. fisherfaces: recognition using class specific linear projection with a puter Vision 63 (2) (2005) 153–161.
view-based representation, IEEE Transactions on Pattern Analysis [23] Ian Witten, Eibe Frank, Data Mining: Practical Machine Learning
and Machine Intelligence 19 (7) (1997) 711–720. Tools and Techniques with Java Implementations, second ed.,
[3] Ronan Fablet, Michael J. Black, Automatic detection and tracking of Morgan Kaufmann, 2005.
human motion with a view-based representation, in: Proceedings of [24] Ming-Hsuan Yang, Dan Roth, Narendra Ahuja, A SNoW-based face
European Conference on Computer Vision, 2002, pp. 476–491. detector, in: Advances in Neural Information Processing Systems,
[4] Ismail Haritaoglu, David Harwood, Larry S. Davis, W4: Who, when, vol. 12, 2000, pp. 855–861.
where, what: a real time system for detecting and tracking people, in: [25] Liang Zhao, Charles E. Thorpe, Stereo and neural network-based
Procedings of Third Face and Gesture Recognition Conference, 1998, pedestrian detection, IEEE Transactions on Intelligent Transporta-
pp. 222–227. tion Systems 1 (3) (2000) 148–154.

Please cite this article in press as: L. Malagón-Borja, O. Fuentes, Object detection using image reconstruction with PCA, Image Vis.
Comput. (2007), doi:10.1016/j.imavis.2007.03.004
View publication stats