Kubik Kurt 
STEREO IMAGE MATCHING USING ROBUST ESTIMATION AND IMAGE ANALYSIS 
TECHNIQUES FOR DEM GENERATION 
Yihui Lu ^ and KurtKubik ‘ 
1 
Department of Geographical Sciences and Planning 
University of Queensland, QLD 4072 Australia 
2 
Department of Computer Science and Electrical Engineering 
University of Queensland, QLD 4072 Australia 
KEY WORDS: Image matching, computer vision, image understanding, DEM 
ABSTRACT 
Digital Elevation Models (DEM) produced by digital photogrammetry workstations are often used as a component in 
complex Geographic Information Systems (GIS) modeling. Since the accuracy of GIS databases must be within a 
specified range for appropriate analysis of the information and subsequent decision making, an accurate DEM is needed. 
Conventional image matching techniques may be classified as either area-based or feature-based methods. These image 
matching techniques could not overcome the disparity discontinuities problem and only supply a Digital Surface Model 
(DSM). This means that matching may not occur on the terrain surface, but on the top of man-made objects such as houses, 
or on the top of the vegetation. In order to get more accurate DEM from overlapping digital aerial images and satellite 
images, a 3D terrain reconstruction method using compound techniques is proposed. The area-based image matching 
method is used to supply dense disparities. Image edge detection and texture analysis techniques are used to find houses 
and tree areas. Both these parts are robustified in order to avoid outlyers. The final DEM comes from the two parts of image 
matching and image analysis and hence overcomes errors in the DEM caused by matching on tops of trees or man-made 
objects. 
1 INTRODUCTION 
A major research area in computer vision and digital photogrammetry is image matching for the reconstruction of a Digital 
Elevation Model (DEM). This process, which is a fundamental problem in stereo vision, involves the determination of 
corresponding points in a stereo image pair. From the image coordinates of these corresponding points, their 3D positions 
can be computed by triangulation, from the known camera geometry, and additional points on the terrain surface can be 
obtained by interpolation. However, 3D terrain reconstruction from aerial or satellite images will be subject to errors in 
built-up and treed areas [Baltsavias et al 1995,Henricsson et al 1997 & Tonjes 1996]. In order to obtain a more accurate 3D 
terrain model, it is necessary to develop better methods to overcome these problems. In this paper, procedures are described 
that combine image analysis and image matching methods in an attempt to ensure that the elevation points are measured 
only on the natural terrain surface, and not on the top of vegetation or man made features such as houses. Section 2 
introduces the proposed system. Section 3 and Section 4 describe the stereo image processing procedure and the single 
image processing procedure respectively. Section 5 gives experimental results, and conclusions are drawn in Section 6. 
2 GENERAL DESCRIPTION OF THE TERRAIN RECONSTRUCTION SYSTEM 
Figure 1 illustrates the architecture of the proposed 3D reconstruction system. The goal of this technique is to achieve 
more accurate reconstruction of elevations from overlapping aerial or satellite images over a wide variety of terrain types 
and ground cover. The key functions of data acquisition and pre-processing, are to acquire the images in digital form and 
improve the output for the subsequent processes by the production of epipolar images from the original left and right 
images. 
  
520 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.