Error Distribution Map
B B B B B B B B e C B
A : Flat
B : Loose Slope
C : Steep Slope ( 235°)
B B B B e B B A c c B
Total Error Frequency
error(pixel) A B e
8 à € € A à AC. BB, C8 9 71 115 61
B B 3 C À à A C C B C 8 1 0 0 0
BB ho CACCADSUADLCOA 8 C 2 0 3
2. B...C3..C. Ar +A Bail FCB Co» 3 9 0 2
Fig. ll Matching errors at every 5 grid points
Terrain slopes are classified to A, B
and C. Affixed figures mean matching
errors in pixel (in the case of no errors
zero's are omitted)
region. Besides large wanderings of 4-8 pixels occur in the enclosed
areas with break lines. They are thought of due to unstableness of
matching.
The contour map plotted when using median-filtering of x-parallaxes
instead of low-pass-filtering is shown in Fig.8. 12x16 points are used
for plotting for every patch pair. Small errors of 1-2 pixels are often
seen near boundaries of patches. Though the occluding areas are rather
well reproduced, wanderings of about 4 pixels still occur in a square
area drawn in the figure, because the stability is still insufficient.
The contour map plotted when using median-filtering of x-parallaxes and
LOG-filtering of patches is shown in Fig.9. A series of the filtered
image pairs with the LOG filters are shown in Fig.10. This procedure
produces no wanderings and the terrain is reproduced precisely enough.
For checking the precision on a display, every 5 grid points are sampled
diagonally. The result is shown in Fig.11. No errors occur at almost
all grid points. Even at the most steep slopes errors are at most 3
pixels.
6. Conclusion
The coarse-to-fine correlation algorithm for stereo matching of aerial
images is developed and its performance is tested. The key point of the
algorithm is narrow-band-pass-filtering of images for stable matching
and median-filtering of x-parallaxes for finding occlusions. The
algorithm is forecasted to be especially available for middle or small
scale aerial images which are plenty of textures and contain relatively
- 327 -
