Kourosh Khoshelham

Automated building extraction from multi-source data has attracted great attention in recent years. This paper presents an approach to automated building extraction by fusion of image data, height data and 2D ground plans. In this approach buildings are detected using ground plans and height data. A split-and-merge process is applied to fuse image and height data and derive the parametric forms of roof planes. Vegetation regions are identified and discarded using the image information in red and infrared channels. Walls are reconstructed as vertical planes upon the ground plan. The model planar faces are finally intersected and resulting plane patches are assembled together to form a generic polyhedral model. Results of the experimental testing indicate the promising performance of the proposed approach in automatic detection and reconstruction of buildings.

x, 160 leaves : ill. (some col.) ; 30 cm. PolyU Library Call No.: [THS] LG51 .H577P LSGI 2004 Khoshelham Digital building models are used in various applications and are traditionally acquired by manual digitization of aerial images in stereo view using photogrammetric stereoplotters. Manual digitization is a tedious and time-consuming task, which requires skilled operators and expensive equipments. Therefore, methods that can assist the operator in performing the entire or part of the task automatically are of great significance. Automated methods for extraction of buildings face a number of problems. Semi-automated approaches involve a considerable amount of interaction. Fully automated approaches that work solely based on a single source of data, suffer from the lack of robustness due to complexities involved in data as well as in buildings. In this thesis, a new approach is presented for semi-automated extraction of buildings from a single aerial image with minimum interaction. ...

Automated detection and 3D modelling of objects in laser range data is of great importance in many applications. Existing approaches to object detection in range data are limited to either 2.5D data (e.g. range images) or simple objects with a parametric form (e.g. spheres). This paper describes a new approach to the detection of 3D objects with arbitrary shapes in a point cloud. We present an extension of the generalized Hough transform to 3D data, which can be used to detect instances of an object model in laser range data, independent of the scale and orientation of the object. We also discuss the computational complexity of the method and provide cost-reduction strategies that can be employed to improve the efficiency of the method.

Over the past few years, there is an emerging growth in mobile mapping systems which can effectively capture the geospatial data in an efficient way. A typical terrestrial mobile mapping system consists of camera, laser scanners, GNSS and INS (Inertial Navigation System). Imagery data is captured by camera and the point clouds are acquired by the laser scanners. GNSS and INS are used for measuring the positional and orientation information of the mapping sensors respectively to achieve direct geo-referencing. GNSS/INS system is very expensive which makes the overall mapping system very expensive. An alternative to make this system is to use Structure from Motion Approach (SfM), to minimize the relative cost of this system. SfM generates 3D point clouds of scene and estimate the orientation parameter for mapping sensor by using imagery data. The major issue with SfM is that it generates point clouds in arbitrary coordinate system with arbitrary scale. Therefore, this research work was focused on feasibility of mobile mapping by integrating SfM approach (for the estimation of camera orientation parameters) with GNSS (for exposure station position) for comparatively low cost terrestrial mapping system and to make whole system (captured 3D scene) direct geo-referenced with proper scale without using ground control points (GCPs). In the first step, sequences of images were captured by measuring the exposure station positions. Then feature extractions and matching were done on the sequence of overlapping images. Bundle adjustment was then applied on it to generate the 3D scene (point clouds) of rigid body and for estimation of camera orientation parameters. The generated point clouds were in arbitrary coordinate system, so tie points were selected to transform into mapping coordinate system. Space photo intersection was applied on tie points with the use of exposure orientation parameters and matched feature points in sequence of overlapping images to transform into world coordinate system. Further, point-based similarity transformation was used to generate transformation parameters from tie points. These transformation parameters were applied on whole generated point cloud to transform it into world coordinate system with proper scale. Then the accuracy assessments on point cloud were carried out using internal and external accuracy assessment. Sometime the epipolar lines do not exactly cross at a fixed point in different overlapping images due to which a distorted 3D scene (point clouds) was created. There was an error in the estimation of orientation parameters due to no ground measurements were used in bundle adjustment. Thus it was observed that the shape of the point cloud was concaved near start and end edges of the scene. RMSE were 32.21 cm, 20.50 cm and 23.56 cm in easting, northing (depth) and height respectively.

This paper discusses a method to overcome problems in image and height data in order to move one step closer toward a fully automated building reconstruction system. In the following these problems with image and height data as the main data sources are further illustrated

The roughness of rock surfaces is traditionally measured by using manual tools such as carpenter's comp and compass and disc clinometers. The manual measurements are limited to small samples at accessible parts of the rock. Terrestrial laser scanning is an attractive alternative measurement technique, which offers large coverage, high resolution, and the ability to reach inaccessible high rock faces. The application of laser scanning to the study of rock surface roughness faces a major challenge: the inherent range imprecision hinders the quantification of roughness parameters. In practice, when roughness is in millimeter scale it is often lost in the range measurement noise. The parameters derived from the data, therefore, reflect noise rather than the actual roughness of the surface. In this paper, we investigate the influence of laser scanner range measurement noise on the quantification of rock surfaces roughness. We show that measurement noise leads to the overestimation of...

The application of terrestrial laser scanning to the study of rock surface roughness faces a major challenge: the inherent range imprecision makes the extraction of roughness parameters difficult. In practice, when roughness is in millimeter scale it is often lost in the range measurement noise. The parameters extracted from the data, therefore, reflect noise rather than the actual roughness of the surface. In this paper we investigate the role of wavelet de-noising methods in the reliable characterization of roughness using laser range data. The application of several wavelet decomposition and thresholding methods are demonstrated, and the performances of these methods in estimating roughness parameters are compared. As the main measure of roughness fractal dimension is derived from 1D profiles in different directions using the roughness length method. It is shown that wavelet de-noising in general leads to an improved estimation of the fractal dimension for the roughness profiles....

ABSTRACT 3D models of indoor environments are important in many applications, but they usually exist only for newly constructed buildings. Automated approaches to modelling indoor environments from imagery and/or point clouds can make the process easier, faster and cheaper. We present an approach to 3D indoor modelling based on a shape grammar. We demonstrate that interior spaces can be modelled by iteratively placing, connecting and merging cuboid shapes. We also show that the parameters and sequence of grammar rules can be learned automatically from a point cloud. Experiments with simulated and real point clouds show promising results, and indicate the potential of the method in 3D modelling of large indoor environments.

ABSTRACT Building indoors reconstruction is an active research topic due to the importance of the wide range of applications to which they can be subjected, from architecture and furniture design, to movies and video games editing, or even crime scene investigation. Among the constructive elements defining the inside of a building, doors are important entities in applications like routing and navigation, and their automated recognition is advantageous e.g. in case of large multi-storey buildings with many office rooms. The inherent complexity of the automation of the recognition process is increased by the presence of clutter and occlusions, difficult to avoid in indoor scenes. In this work, we present a pipeline of techniques used for the reconstruction and interpretation of building interiors using information acquired in the form of point clouds and images. The methodology goes in depth with door detection and labelling as either opened, closed or furniture (false positive).

ABSTRACT Airborne Laser Scanning (ALS) is widely used in many applications for its high measurement accuracy, fast acquisition capability, and large spatial coverage. Accuracy assessment of the ALS data usually relies on comparing corresponding tie elements, often points or lines, in the overlapping strips. This paper proposes a new approach to strip adjustment and quality assessment of ALS data by using planar features. In the proposed approach a transformation is estimated between two overlapping strips by minimizing the distance between points in one strip and their corresponding plane in the other. The planes and the corresponding points are extracted in a segmentation process. The point-to plane distances are used as observables in the estimation model, where the parameters of a transformation between two strips and their associated quality measures are estimated. We demonstrate the performance of the method on the AHN2 dataset over Zeeland province of The Netherlands. The dataset consists of 13 overlapping strips, from which a total of 522 gable roof and dike slope planes are extracted. The results show planimetric offsets between the strips that range from 3.13 cm to 55.32 cm. These offsets are in agreement with previously reported results using linear features. In addition, we estimated vertical offsets in the order of a few centimeters, which were not estimated in previous studies. The rotation parameters between the strips were also estimated; however, these did not show a significant difference in the orientation of the strips.

ABSTRACT Rock joint roughness characterization is often an important aspect of rock engineering projects. Various methods have been developed to describe the topography of the joint surface, for example Joint Roughness Coefficient (JRC) correlation charts or disc-clinometer measurements. The goal of this research is to evaluate the accuracy, precision and limits of Terrestrial Laser Scanning (TLS) for making remote measurements of large-scale rock joints. In order to find the most appropriate roughness parameterization method for TLS data and to analyse the capability of TLS for roughness estimation, experiments were made with a 20 × 30 cm joint sample. The sample was scanned with TLS and compared to reference measurements made with the Advanced TOpometric Sensor (ATOS) system. Analysis of two roughness parameterization methods, virtual compass and disc-clinometer, and angular threshold method, showed that the latter is less sensitive to noise. Comparative studies of ATOS and TLS roughness parameters indicate that the TLS can adequately quantify surface irregularities with a wavelength greater than 5 mm from a distance of 10 m.

Automated reconstruction of buildings from different data sources has been one of the most challenging problems in photogrammetry and computer vision. Systems for automated building reconstruction fail in many cases due to complexities involved in the data including image noise, ...

1. Introduction Automated extraction of building objects from different data sources has been a topic of intensive research in past years. There is an increasing demand for digital building models, which can be used in numerous applications including urban planning, ...

The roughness of rock surfaces is traditionally measured by using manual tools such as carpenter's comp and compass and disc clinometers. The manual measurements are limited to small samples at accessible parts of the rock. Terrestrial laser scanning is an attractive alternative ...

The application of terrestrial laser scanning to the study of rock surface roughness faces a major challenge: the inherent range imprecision makes the extraction of roughness parameters difficult. In practice, when roughness is in millimeter scale it is often lost in the range ...

Abstract: This paper presents a learning Dempster-Shafer model for the detection of buildings in aerial image and range data. The process of evidence assignment in the Dempster-Shafer method is implemented through membership functions in an adaptive-network-based ...