ırban 
1odel 
s are 
made 
rains 
gs 18 
irban 
d the 
| and 
ional 
matic 
m of 
1 full 
matic 
s and 
)n of 
ng of 
f the 
| 3D 
hang, 
gular 
' ofa 
| 
g 
1995; 
5, 1S a 
/ever, 
r low 
ted in 
icture 
n the 
  
  
Zuxun Zhang 
First, by using the correlation coefficient ©, select n candidate points of homologous image in neighbor image: j,, /,, .. 
.5 
J,» Compute the initial probability of each candidate point: 
right 
p'G,j)- pa. D/ Y. pa.) 
Then, considering the relation between image point i and its 
neighbor point k, for instance, calculate the compatible coefficient 
C( i, j, k, | ) between them. As demonstrated in figure 1, the 
horizontal axis represents the left image while the vertical the right. 
Based on the principle of local maximization of correlation 
coefficient, compute for two image points (i and k) in the left 
image their candidate points of homologous images in the right 
image (j,, jo) (ln I» Ij. If (i, jj) and (k, 1j) are the right two of 
homologous point pairs, the similarity of image segments between 
the left and the right image is in accordance with the compatible 
coefficient: C (ij; KI) 9 p. Cik, j;l;) . Tt may render an efficient Figure 1. Bridge mode 
calculation of geometrical distortion of the homologous points, 
which is the essence of the bridge-mode method; that is, it 
connects all the matching points between i and its neighbour points 
(set k ), which is a set of neighbour points. The probability p'(i , j ) 
is modified with the method of probability relaxation in an iterative 
procedure. 
  
qi Do Cd i s 
p'G, j)2 p^ (i. ))-(4+B-4(, j)) 
p' Gs p' GJ (p 6.3) 
The algorithm has a full consideration of global relation of image 
matching and enhances significantly the reliability of the result. 
Figure 2 shows DSM of the urban area generated by VirtuoZo, in 
which the houses are appearing as lumps. 
  
Figure 2. Urban DSM 
2.2 Detection of Local Prominence Area 
The above analysis tells that houses are shown at ‘lumps’ in DSM. And the so-called detection of local prominence area 
(LPA) is a process of separating the lumps from DSM and extracting the boundaries. The paper adopts a heuristic 
searching method based on the principles of maximalization of gradient differences and minimalization of height 
difference (Qiu, 1997) , which reguires that the boundaries of local prominence area should be placed where gradient 
difference of DSM is largest and where the height difference (or parallax) of two neighbouring points in the the 
boundaries should be minimal. 
If s(x,y) is the slope value of the point(x,y), then the module G(s(x,y)) of the slope difference rate is: 
G(s(x, y)) =[(0s/0x)" + (0s/0y)*]? 
And the parallax between point I and point j is : 
v=|n-v/ 
The cost function from point ; to / is acquired: 
C, j)- y [GGOG;.y;)) * GGG; y ;)]/(Ap * 1) 
According to the criterion: 
C(i, j) max 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 1017