step. And if zero-crossings do not exist at some grid points, matching 
of those points is thought of to be impossible and the x-parallaxes are 
interpolated from those of surrounding grid points. But actually such 
cases seldom occur since we employ as wide as 4-octave band width of cor- 
relation window, while the band width of the LOG filter is 1.6 octaves. 
4. How to find occlusions 
Matching wanderings occur in occluding areas. We discuss here a way of 
evading wanderings due to occlusions by the help of median-filtering of 
x-parallaxes. Associated with the discussion we should first argue about 
removing vibrating errors in x-parallaxes. 
Even though matching wanderings due to occlusions were not to occur, 
vibrating errors are usually contained in measured x-parallaxes. They 
make a bad effect on rearranging of pixel arrays to eliminate perspective 
distortions. Such vibrating errors are effectively removed by some kind 
of a low-pass filter, e.g. a moving-average. But if gross errors or 
wanderings are contained, low-pass filters do a harm than a good on 
account of the averaging effect. Furthermore since occlusions are 
always accompanied by abrupt changes in terrain height (or in x- 
parallaxes), if we employ a low-pass filter, the terrain is forcedly 
reproduced in a smooth form across occlusions. Then the following 
algorithm is developed to solve such difficulties. 
We attend to the fact that occluding areas, widths of which are smaller 
than the shortest wavelength of included signals never appear. Since 
the grid spacing is set to 8 pixels and the correlation window width is 
set to 15 pixels on the reduced patches, occlusions smaller than 4 
pixels in width can be neglected in correlation. If the widths of 
occlusions are between 4 and 8 pixels, wanderings may occur. But they 
can be expected not to occur at successive 2 grid points in x-direction. 
Such wanderings are regarded as impulse noises, which can be removed by 
median-filtering completely. The median filter is a filter that replaces 
every signal by a median in the window taken around the signal. Its most 
significant characteristic is a nature that it removes impulse noises 
with edges reserved. Further it hold an ability of removing vibrating 
noises as well, though a little weaker than that of the moving-average 
with the same-sized window/3/. We adopt the window of 3x3 grid points, 
which is determined through some preliminary experiments. * 
Thus the processes (4)-(8) in the algorithm stated in section 2 can be 
given a body and substance as follows. 
The 1st step: 
1)The maximum width of occlusions involved in the reduced patches I, is 
assumed smaller than 4 pixels. 
2)Correlation is performed for all the grid points to search for 
conjugate points. ; 
3)The measured x-parallaxes are median-filtered. 
4)It is checked if the measured x-parallaxes satisfy the consistency 
condition on the positions of conjugate points; 
The order of the conjugate points standing in a row in x-direction 
must be the same as that of the correponding grid points. 
The grid points that do not satisfy the condition are regarded as 
occluding. 
The 2nd step: 
5) Initial values of the x-parallaxes of the grid points on the left 
patch I5 are interpolated from those of the grid points on the patch I,. 
The grid points found to be occluding in the 1st step are processed ina 
particular way. Since in our algorithm, grid points are placed on the 
left patch, we should consider the case that the conjugates on the right 
patch vanish as shown in Fig.4. Let assume in Fig.4 that the grid points 
P490; P12:; P44; P259: P22 »,;P34 be on I1, and P410: P414; P12; P43-++++>P34 
- 322 -