Hee Ju Park 
  
  
  
  
ü Source Images | 
t 
| Epipolar Images | 
Y Y 
  
| Interest Points | | Edges dl 
Y 
  
  
  
| Point matching | 
| Blunder suppression | 
  
v 
* Corresponding points position 
* sdv(dG), E(dG) 
  
  
  
  
  
  
  
  
  
  
* sdv(dx) 
L Line Matching | 
L Results A Image 1 Image? 
Figure 1. Basic flow of proposed matching method Figure 2. Neighbouring points in point matching 
threshold of Interest Value w which is related with the contrast, we use a positive value of Wi» for 
example W. =1.0. We don’t consider the value q which is related with shape. The reason of this is the line matching of 
the next step. In literature many methods for matching between Interest Points can be found [Foerstner et al, 
1987;Zhang, 1994]. One simple method is the correlation coefficients method. To increase the reliability of matched 
points, we perform an additional check for the case when the template window and search window is reversed. If the 
result for each point pair matched is the same for both - normal case and reversed case-, it is accepted. This method is 
called back-matching [Hannah, 1989]. To improve the reliability we accept the case when both of the matched points 
are Interest Points. Till this stage the matching is performed with the images’ gray values because of better 
performance. Through the above process we get possible matched points sets. Then we check the correlation 
coefficients between each matched point pair for each RGB colour channel. If any of correlation coefficients for each 
colour channel is less than 0.5, that point pair is rejected. Till this process we check the local similarity of areas near the 
points. 
3.2 Blunder suppression to possible matched point set 
Based only on the local similarity comparison, avoiding blunders is difficult. To solve this problem we check the global 
similarity between a possible corresponding point pair and its neighbouring possible corresponding point pairs. 
Let two points of a possible corresponding pair be point i, and point j. We assume that there is a number M of 
neighbouring points near the point i, and point j. Suppose a point m, and its possible corresponding point n are near the 
point i, and point j as shown in Figure 2. The coordinates of point i, point j, point m, point n are (ji), (j,j,): 
(m,m ),(n, n,). We define a measure of Strength of Matching SM, for the pair of point i, point j as follows : 
  
Sc exp(—abs(dx)/2sdv(dx)) 
C, + 
SM.- (mn) 1+[(d(i,m)+ d( j,n)]/ 2 a 
  
  
j 1 
1+ 
x 1 [(d(i, m)  d( ], n)]/2 
(m,n) 
C, : correlation coefficient between point j and point j 
C,,,: correlation coefficient between point m and point n 
d (i, m) : distance between point i and it's a neighbouring point m 
d ( j, 1) : distance between point j and it's a neighbouring point n 
dx x (i, —m )—(j —-n) 
  
700 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
This 
relat 
Poin 
33 
If W 
poss 
as de 
like 
COIT 
simi 
Bec: 
the € 
poin 
follo 
If bo 
cons 
In tl 
com] 
One 
COTTE 
epip 
This 
simp 
The | 
left f