4 
na 
ed. 
age 
del 
es, 
lite 
ing 
ses 
age 
ade 
eve 
pes 
and 
ight 
  
Kubik Kurt 
  
The system developed for DEM determination consists of three main parts. Part 1 performs the matching of the stereo 
image pair, derives a disparity map, and produces a digital surface model (DSM). An analysis of the disparity map then 
reveals possible house and tree areas. Part 2 applies standard image segmentation and texture analysis techniques to the 
  
Part 1 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
ug : Image segmentation 
Image acquisition à Stereo Image matching... (of disparity map : 
i Disparity 
map : 
Pre-processing  |__, ; Part 2 ; > Fusion — > 
; Image segmentation Texture analysis : po» DEM | 
| à | of a single image & classification 2, Part 3 
  
  
  
  
  
  
  
  
  
Figure 1 Architecture of the proposed reconstruction system 
left image to recognize houses and to locate trees. Based on a combination of the 3D information extracted from the 
disparity map and the 2D image segmentation, the elevations derived in regions which do not appear to represent the terrain 
surface can be removed from the DSM in Part 3, thus leading to a more accurate DEM. In the case of the houses, the 
elevations can then be interpolated from the surrounding terrain. Where trees exist, the DSM heights can be reduced by the 
tree heights. 
3 PROCESSING OF STEREO IMAGE PAIR 
3.1 Derivation of the Disparity Map 
The first step in the recovery of 3D terrain information from overlapping aerial or satellite images is based on the matching 
of corresponding pixels in the stereo images. From the matched points, the 3D coordinates of a point can be obtained by 
triangulation using information of the image capturing geometry. Many computational algorithms have been used to solve 
the stereo matching problem. Conventional image matching techniques may be classified as either feature-based or area- 
based. Each of these approaches has advantages and disadvantages. Feature-based matching generally produces good 
results, is less expensive and is more tolerant of illumination differences and geometric distortions. However, only a few points 
may be matched in some regions due to the scarcity of the features, which leads to large areas being subjected to inaccurate 
interpolations. ^ Area-based matching algorithms can provide denser disparity maps. However, they are intolerant to 
geometric distortions caused by steep terrain slopes or imaging geometry. 
In order to produce a dense, reliable matching result, the hierarchical area-based stereo image matching using robust 
estimation was been employed. Since this paper concentrates on the process of recognizing houses and trees in images, and 
correcting for their effects on derived elevations from image matching, the disparity values obtained from matching have 
been directly used in the subsequent stages of the system in Figure 1. For these developments, a dense sample of points in the 
disparity map is required in order to avoid some of structures being missed. Hence, a matching grid interval of 5 pixels in 
column and row directions has been used. The derived disparity map is then interpolated to the same size as the original 
image for further processing. 
3.2 Edge Detection Applied to the Disparity Map 
Figure 2 illustrates the stereo image processing procedure of Part 1 in detail. 
  
  
  
  
Edge More accurate 
Stereo image isparity Sobel edge map| Threshold ) Laplacian of Gaussian » map 
pair » Matching map detection ——3 edge map | [ + Zero crossings 
  
  
  
  
  
  
  
  
  
  
  
  
Figure 2 Stereo image processing procedure of Part 1 in Figure 1 
Although automatic area-based matching algorithms are not able to distinguish between the terrain surface and objects on and 
above this surface, the output of stereo image matching can supply significant information to identify man-made structures 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 521