ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001 
249 
Matching 
ting information in 
e to urban area is 
e-based matching 
I. This approach is 
peration [4]. Edge 
process and thus 
/ere executed with 
with edge support 
ja bytes at high 
ddle range cost, 
I of the processes 
nate goal for the 
issible automatic 
juce the cost of 
information. 
at improving the 
iss by introducing 
3 information. The 
)f the proposed 
etail the improved 
nation. The third 
with the reliable 
le and constraints 
en conducted and 
is are capable of 
stereo matching 
30RITHMS 
and precision of 
ion. The reliability 
iformation that is 
on is improved by 
constraints. The 
e as follows. 
ig. In this study, 
is for initialization 
equal to general 
r constraining the 
ion will be given 
. The third one is 
lapping between 
ional approaches 
> to only detect of 
highly reliable conjugate edges for general registration. Since in 
the process of mapping between stereo image pairs do not 
necessarily require conjugate edges, more edges that are not 
strictly matched can be used for maintaining uniform parallax. 
CE is one kind of neural network model and gives very unique 
mapping process, which realizes nonlinear mapping from inputs 
to outputs. In this study, this technique is applied to stereo 
matching process, with the input and the output being the stereo 
image pairs. Since CEM produces matching results less 
sensitive to erroneous local minimum, it can reduce the 
possibility of mismatching caused by occlusions or noises. In 
order to reduce the effect of discontinuous shift vectors or 
parallax changes, we use the edge information of high reliability 
as constraints. 
2.2 Process Flow 
Fig. 1 shows the general process flow of our proposed 
algorithms. 
Fig. 1 The General Process Flow 
In the first stage, edges are extracted using rectified stereo 
image pairs, which are across epipolar lines. The processes 
consist of several steps, such as edge enhancement, thinning, 
tracing of enhanced edge image, segmentation, filtering by 
geometrical constraint and template matching with the 
consideration of edge boundaries [2]. We apply two different 
types of edge operators for edge image enhancement and 
compare the result to choose the more effective one. The 
operators are SUSAN (Smallest Univalue Segment Assimilating 
Nucleus) operator [3] and Haar’s wavelet transformation 
respectively. 
In the second stage, the edges extracted from stereo image 
pairs are matched with each other to form conjugate edge pairs. 
Geometric constraints from the orientation parameters, and 
image similarity calculated from edge dependent template are 
introduced to guarantee the high reliability of matching result. 
The matched edges are further trimmed to the same shapes 
according to their common section. 
In the last stage, nonlinear matching process is performed. This 
process is based on CEM. CEM realizes nonlinear normalization 
at local area in the target system and acquiring of global 
agreements by iterative computation. In this approach we 
conceptually identify neuron's activity with matching or shifting 
vector. This network model has two phases in optimizing 
procedure of shifting vector. The first one is competition phase 
in which shifting vectors are individually searched for optimal 
displacement with certain constraints. The second one is 
consensus phase in which shifting vectors are operated to have 
local continuity of shifting vector or parallax. These two phases 
are repeated to achieve local normalization with respect to the 
edge matching results. We tried to extend CEM with the result of 
edge matching. 
3. EDGE DETECTION FROM RECTIFICATION IMAGE 
To simplify the task of edge detection and refinement, stereo 
image pairs are translated to rectified images using adequate 
perspective projection, as long as proper tie points are specified. 
This step includes photogrammetric operations such as 
estimation of orientation parameters by relative orientation. 
Rectification images provide a stereo model that has no y- 
parallax. As a result, original 2 dimensional matching becomes 
the equivalence of 1 dimensional matching (in x-direction, 
corresponding to epipolar line), and only extraction of edges 
across epipolar lines need to be considered. 
The process flow of edge detection after rectification is shown in 
Fig.2 [2]. In this study, two types of edge detection operators are 
applied. One is the SUSAN operator proposed by Smith et al. 
and the other is wavelet transformation. SUSAN operator is a 
unique interest operator that pays attention to the variation of 
regional brightness distribution in a local window. The brief 
descriptions of each process are as follows. 
Fig.2 The Process Flow of Edge Detection 
(1) Pre-processing 
As pre-processing, stereo image pairs are smoothed by 3 x 3 
size median filter for noise reduction, and brightness or contrast 
adjustment is applied to have almost the same brightness. 
(2) Edge Enhancement 
Two types of operators (SUSAN and wavelet) are used to 
enhance edge component to get the candidates of building’s 
structural edge. Haar’s wavelet is used here and operation is 
applied in x-direction. The result of level-1 process is selected 
for enhanced image. Binary edge images are also produced in 
this stage. Eq.1 shows the basic formula of wavelet 
transformation and Fig.3 shows Haar’s mother wavelet. 
(WJ){b,a) = Eq. 1 
where f(x) is input signal, <t>(x) is mother wavelet function, a is 
scale parameter and b is translation parameter. 
1 
1 
-1 
u 
Fig.3 Haar’s Mother Function 
(3) Edge Thinning 
Applying mathematical morphology operation (dilation, erosion, 
hit and miss etc.), binary edge components are locally 
connected and thinned to trace edge segments. 
(4) Edge Tracing 
Edge tracing is performed by 8-direction neighbor connection 
process. In this stage chained edge elements are extracted from 
raster edge images. 
(5) Edge Line Segmentation