| Table 4: Number of matched end points after mismatch 
of $ detection. 
Cost Benefit 
Function function 
No. of matched points from 96 103 
relational matching 
No. of matched points after 42 44 
; > radiometric mismatch detection 
# e Y a No. of matched points after 41 44 
geometric mismatch detection 
f E 
shows the results of automatic relative orientation from the 
proposed relational matching scheme with two evaluation 
functions and an experienced human operator. As shown, the 
results of automatic relative orientation were less accurate than 
  
    
   
  
  
    
  
    
  
  
  
  
  
  
  
  
   
  
   
  
  
  
  
   
  
  
  
  
  
  
  
  
  
  
   
  
  
   
  
  
   
  
   
   
  
   
   
   
  
   
   
   
  
  
  
   
   
  
   
    
  
  
  
  
Figure 9: Matched line primitives with benefit function: 
(a) left (unit) image and (b) right (label) image. 
118 
developed and implemented. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
that of a human operator. However, the results of automatic 
A relative orientation are less than one pixel accuracy, which is 
- accurate enough for relational matching to be used as a means 
N of providing the approximation for a highly accurate matching 
| " J | method such as area based matching. 
Y SR Table 5: Results of automatic relative orientation. 
"cip a 
UE y "| Cost Benefit Analytical 
= \ V La 1a function function plotter 
A en enl Ae e | Pixel resolution 480um 480um 
e cdl e ui | Number of points 40 44 9 
On Xo (mm) 90.681 90.618 90.586 
uis n/ ee, Yo (mm) -2.439 -2.450 -2.393 
er mb M Zo (mm) 150.182 | 150175 | 150341 
L ol Bar EL \ | o (deg) 0.644 0.647 0.604 
AP x \ vit | $ (deg) -1.183 -1.210 -1.327 
uie ota fv S^ A IN P A | K (deg) 359.151 359.168 359.064 
i ^ 4 ei. 2 By a | Maximum residual 758.7 722.6 10.3 
| Standard deviation 329.1 304.5 4.5 
(a) 
5. CONCLUSIONS 
M it In this study, relational matching with two evaluation functions 
SN f was implemented to solve the image matching problem, 
si specifically automatic relative orientation. As shown, the 
3 proposed relational matching scheme provided reliable results 
T : SN of automatic relative orientation for urban area images 
wi r dn \ containing breaklines, occlusions, and repetitive pattern by 
M ung M using three different binary relationships between two straight 
= ifl line primitives. 
| AM Lh m e 9 In the proposed relational matching scheme, locally consistent 
: y 103 1, \ és m " relations were extended to a globally consistent solution. From 
| Tio T. a practical point of view, this is a way of obtaining a solution 
| uS for a large set of primitives. Furthermore, the 2-D tree 
Ey x/ Ma technique was developed to speed up the tree search and 
rol ms wa s\ \ manipulate the local binary relations efficiently. Implementing 
“a iem ah heuristics such as unit ordering and the modified forward 
a ee my ) checking also helped the relational matching reach a solution 
: hs a ^ RM pe without expanding unnecessary subtrees. 
a a, pe aye TA e The investigation of two evaluation functions in relational 
"a sad E m 5 s x matching showed that the benefit function performs little better 
than the cost function, particularly for the domain of the image 
(b) matching problem. One way of obtaining an initial 
approximation between two images with little rotation was 
It showed that the distinct and 
prominent rt 
solving the 
experiments 
matching su 
images. 
This rese 
problems ii 
expenses ar 
amount of ti 
relational m 
computing 
poor manne 
features. Fi 
well constr 
solution. A 
reaching ar 
From the 
identified fc 
e A feat 
well w 
shown 
when t 
e  Thehi 
and fa 
way o 
descri] 
e Some 
orient: 
match: 
cases, 
invari: 
e [n thi: 
The pi 
to ma 
that i! 
proble 
esoscdhec 
form 
existir 
maxin 
e Thep 
this r« 
maps, 
invest 
future 
Ballard, 
Computer 
Boyer, K. 
Vision. /E 
Intelligenc 
Burns, J. 
Extracting 
Analysis a 
455. 
Cho. We 
Orientatio 
Columbus