„matched points from the automatic orientation to dispar-
ity values. These disparity values are then interpolated to
generate a disparity map to determine the location of the
search window. After the matching process is completed in
one level of the image pyramid, the disparity map is up-
dated to ensure that better approximations in the next level
are available. This is quite important, particularly in urban
areas where the disparity values may abruptly change. It
should be noted that the disparity map always corresponds
to the resolution of current level in the image pyramid.
3.2. Surface Analysis in High Level Matching
The use of the disparity map provides not only good approx-
imations for correlation matching but also a closer surface
approximation after each level of matching. In higher levels
of the image pyramid such a surface can even provide a lot of
three-dimensional object information. This information can
then be used for real object surface analysis, as discussed in
[Wang et. al. 1992]. On the other hand, the information
gained from the 3-dimensional analysis can be fed back to
guide the matching.
One of the most difficult matching cases are urban aerial im-
ages in which there exist man-made features with extreme
height, such as tall buildings or chimneys. The deformations
and disparities of such features in stereo images can be very
large causing the matching to be either incomplete or unsuc-
cessful. The solution here is to analyze the disparity map.
As one application of 3D feature analysis discussed in [Wng
et. al. 1992], a contour map can be generated after segment-
ing a disparity map. The following rules are implemented to
superimpose knowledge to the existing disparity map and
used to guide the program to find potential high features.
® A cluster of close-centered contours indicates a poten-
tial hump
e If the inner disparity values are much larger than the
outer ones, a potential high hump is indicated
e For a potential high hump, information is fed back to
guide a future matching
® The boundary of a potential high hump is the second
closed outer contour since the first one may indicate
boundary of the environment
The potential disparity values inside the selected
boundary are the average disparity values of the
matched points inside the boundary
The obtained disparity values of the potential hump
are appended to the matched data and a new disparity
map is interpolated
3.3. Figural Continuity Constraint
The figural continuity criterion implies that the disparity
values along zero-crossings must be continuous. We imple-
mented the figural continuity constraint by performing a
Hough transformation of all the matched points belonging
‘to one segment of a zero-crossing contour. Continuous dis-
parity values show up as clusters in the Hough space. If fewer
than 15 points fall into the cluster a flag is set to indicate
146
that there is no corresponding zero-crossing segment. Fi-
nally, the location of the corresponding segment in the right
image is determined by the Hough transformation and the
correlation threshold.
4. EXPERIMENTS
The matching algorithm was tested with several pairs of
aerial photographs. In this paper, we present the results
from stereo-images (193, 195) taken over the campus of the
Ohio State University. This model represents a very typi-
cal urban area of all the different models tested, it was the
most difficult one. The photo scale here is approximately
1 : 4000. The diapositives were scanned to a resolution of 304
pixel size by Intergraph Corporation using the PhotoScan.
However, we only used a resolution of 604 which yielded a
4096 x 4096 pixel image. The ground coverage of a pixel is
approximately 25 x 25cm.
Fig. 2 and 3 show the original aerial images at the coarsest
resolution of 512 x 512. Zero-crossings were first detected
with the LoG operator (w = 5), and then matched with a
single average disparity approximation. The range of the
search window was set to 10 pixels in order to avoid wrong
matching. The matched zero-crossings are shown in Fig. 4
and 5. À disparity map was interpolated by using Modular
function on Intergraph workstation. The result is shown in
Fig. 12. which outlines the surface of the whole overlapping
area of the model. Some humps are clearly visible.
The interpolated disparity map was then converted into an
image of 512 x 512 resolution and the disparity values were
treated as graylevels. Fig. 6 and 7 show matched zero-
crossings of a 512 x 512 image patch selected from stereo
images of 1K x 1K resolution. Fig. 13 shows the interpolated
3D disparity map. The humps are now more prominent. Fig.
8 and 9 depict matched zero-crossings of a 512 x 512 image
patch from 2K x 2K resolution images. The interpolated
disparity values are shown in Fig. 14. The surface is fairly
well approximated at this level.
The procedure is repeated at the finest resolution, again with
an image patch size of 512 x 512 pixels. In this example
the disparity values range from 0 to 118. The segmented
disparity image resulting from matching is shown in Fig.
16 where the hump is clearly indicated. Fig. 10 and 11
show the matching results superimposed to the resampled
images, while Fig. 15 and 17 show the interpolation of the
final matching results in the disparity map and the three-
dimensional object space, respectively.
5. CONCLUSION
The presented matching scheme combines the merits of both
area-based and feature-based matching methods and proved
succussful in the aerial image matching. The use of a hi-
erarchical approach and surface approximation makes this
approach particularly suited for urban area image matching.
It is found that the precise detection of prominent features
is helpful for recovering the object surface. The reliability of
the correlation matching is improved by the employing the
figural continuity constraint. Finally, this matching scheme
shows a great potential for object surface analysis and re-
construction.
Cho,
ital I
Li, J.
on Z
Note.
Scier
Marr
Hum
Schei
proac
tified
28(5/
Scher
copy
ginee
pp. 1
Wang
ysis 1
Zong,
Inter
ceed
