ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001 
250 
Since chained edge elements are not direct representation of 
object’s shape, they must be approximated as line segments 
(line segmentation), which are the fundamental information in 
the following processes such as edge matching. The algorithm 
of line segmentation is as followings. 
Consider edge element with m nodes {n u ...n m ). 
(i) Approximate line L, with node n, and n,. At initial i = 1 , j = 
m. 
(ii) Calculate distance d k (k = i+1,...J-1) between each n k and 
4 
(in) If any d k is greater than a pre-defined threshold dt, 
substitute j-1 for j and go to step (i). 
(iv) Register L, as line segmentation and substitute j for /. 
(v) If i equals to j, process is finished, otherwise go to (i). 
Line segments with length below a predefined threshold are 
regarded as of low reliability and discarded. 
4. EDGE MATCHING 
The line segments obtained in the above process differ from 
each other in stereo image pairs in many ways. To make line 
segment usable for stereo matching process, the following 
algorithms are proposed to select and modify reliable line 
segment pairs. 
We apply geometrical constraint and matching criterion of image 
texture for detecting conjugate edge segments. Candidates of 
conjugate edges are those of detected by processes of section 
3. Fig.4 shows the flow of proposed process. 
Fig.4 The Process Flow of Edge Matching 
(1) Filtering with Geometric Constraint 
In DPW system, the orientation parameters of an aerial image is 
known, and can be used as geometric constraints for candidates 
of conjugate points or line segments. In this step, it is assumed 
that all edge segments that have the same epipolar line in the 
other image are the candidates of conjugate pair. Combinations 
of conjugate candidates are filtered with the following geometric 
constraints. 
(a) Edge candidates in the other image are supposed to exist 
in a certain search width, which means that there is an upper 
limit to parallax between conjugate edges. 
(b) Conjugate edges have certain common section length 
across epipolar lines. In other words, the distance in real 
world between sharing section must exceed a specified 
threshold. 
(c) Altitude calculated with edges must be within a pre-defined 
range. Anything out of this range will be rejected. 
(d) Angle of elevation must be less than a specified threshold. 
This condition gives priority to detection of edges that form 
the top of buildings. 
(2) Filtering with Image Similarity 
In this phase, candidates of conjugate edges are further filtered 
with evaluation of image similarity, which is based on the 
correlation coefficient calculated with the templates shown in 
Fig.5. The templates are chosen to be both sides of edge 
segment between the overlapped range and the shape of 
template is parallelogram, which is parallel to edge segment and 
epipolar lines. The higher correlation coefficient of two templates 
is adopted for evaluation criterion and the combination with 
highest one is selected for candidate of conjugate edges. 
(3) Expansion and Contraction 
Normally, the conjugate lines are not of the same length. The 
shorter edge of supposed conjugate edges is first extended to 
that of the longer one. If the image similarity described above is 
improved by this operation, expansion of edge is considered to 
be valid. This operation is executed for all candidates of 
conjugate edges. The next step is equalizing length of conjugate 
edges (contraction process). In this stage specific edges in other 
image have a possibility of referencing by multiple edges as a 
candidate of conjugate edges, so in such a case, the above 
processing is performed after copying those edges. 
Edge Segment 
Fig.5 The Layout of Template for Correlation Calculation 
(4) Conjugate Edge Selection 
As the last phase, conjugate edges are selected by the following 
criterion 
(a) An edge pair is adopted unconditionally when in the 
assumed conjugate edge candidates refer to each other 
as the highest edge and not referenced by other edges. 
(b) If an edge A in a image is referenced by multiple edges B, 
in another image, and edge A has connectable edges in 
its neighborhood with a specified distance threshold, and 
the same condition applies to a B of B„ the pair {A,B} is 
selected. 
(c) For cases similar to case (b), except that there are no 
connectable edges, the edge before being copied in step 
(3) is used, and the pair that gives the highest similarity 
is selected. 
5. CEM ENHANCED WITH EDGE CONSTRAINT 
5.1 Coincidence Enhancement Method 
Coincidence Enhancement is an extension of Hebb’s rule for 
self-organized process in neural network modeling. This process 
can be modeled by using the principle of competition and 
consensus. CE model can realize smooth mapping between 
input signals and output pattern. When take the left stereo 
image as input and the right as output, this concept is applicable 
to stereo matching. 
The concept of stereo matching with CEM is illustrated in Fig.6. 
In the competition phase, each pixel or local area in image A 
tries to find an optimal poison in the search area. A shift vector 
is formed by connecting the original position in A and the optimal 
position in B. In consensus phase each shift vector is modified 
with surrounding shift vectors in consensus area. These