wanderings in other areas than occlusions seldom happen. The former comes 
from the requirement for the procedure to find out. occlusions from 
their surrounding x-parallaxes. Therefore this algorithm is useful for 
matching of small or middle scale aerial images which are usually 
plentiful in textures and do not contain many occlusions. And the latter 
comes from the requirement for the median-filter to work well. 
Note 2: If the maximum width of occlusions is beyond 8 pixels on the 
reduced patches I» in the first step, one can easily reform the algorithm 
to consist of more steps. 
5. Performance Test . 
Fig.5 shows the stereo images of 400m x 400m used for tests, the original 
scale of which is 1:30,000. They are digitized with 20 pm spacing to 256 
levels from 0 to 3D. 20 ym spacing corresponds to 60 cm on the ground 
and the parallax difference of 20 pm corresponds to 1.2 m in height 
difference. Further by rearranging pixel arrays along epipolar lines with 
the bi-linear method, y-parallaxes are removed from the images. The 
grid spacing of 8 pixels corresponds to about 5m on the ground. There 
exist the occlusions A and C, which are enclosed with dotted lines, along 
y-axis on the both sides of a central flat region. And the area B 
contains terrain discontinuity in y-direction. 
For comparison a contour map is plotted by a skillful operator with 5m 
interval as shown in Fig.6.(One might notice the line pointed by an arrow 
is very different in shape from corresponding ones in the contour maps 
obtained by experiments below. In fact it is plotted by following the 
edges of a field, not by following equal terrain height.) Moreover an 
image display is used for checking the matching precision to the order 
of a pixel. 
The contour map plotted when using moving-averaging of x-parallaxes with 
cross-shaped windows of 5 points, filtering of image densities with the 
ideal low-pass filters and correlation using cross-covariances, is shown 
in Fig.7. In this case, 2 grid points on the margin are discarded and 
20x20 grid points are used for plotting for every patch pair. One can 
see that the neighbours of occluding areas are reproduced smoothly 
(within the enclosed areas with thick lines). In the 3rd step some 
positional conversions of conjugates occur, which are pointed by a 
notation x. A lot of small errors of 1-2 pixels are seen over the entire 
  
Fig. 5 Stereo images used for tests 
- 324 -