ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002 
  
Qa. 
1500 E 
1500 
  
ü D 
3.a 3.b 
Figure 3.a: The results of the parameter space clustering with respect to the scale and rotation angle. Figure 3.b: shows the results of 
the parameter space clustering with respect to the translations along the x and y-axes. The emerged peaks in both figures point to the 
locus of the expect solution that can align the two images. The grids of the two plots are based on the matrix indices. 
Figure 4: The automatically matched points overlaid over their original subimages. The automatically matched points are used as a 
basis for precise registration using least squares solution. 
  
  
Figure 4: shows the automatically matched points overlaid over their original subimages. The automatically matched points are used 
as a basis for precise registration using least squares solution. 
  
  
Point description Number of points Parameter Value Std Dev. 
  
  
Detected points in image 1962 X-translation 35.33 pixels + 0.0917 Pixel 
(1987) 
Detected points in image 1932 Y-translation 330.5 pixels t 0.0917 Pixel 
(1991) Scale 0.9768 +10“ 
Matched Points 328 EC 
Rotation -0.0023deg t1.74x10 deg. 
  
Table 1: The number of the detected and matched points. Table 2: The registration para eters and their standard deviations. 
  
4. DISCUSSION 
The developed approach, detailed in this paper, successfully 
registers the two images, as shown in Fig. (5) . The correct 
matches define a peak in the parameter space, (see Fig. (3)). 
Incorrect matches define non-peaked clusters. It is evident 
from table (1) that this approach is highly robust , since the 
percentage of the matched point compared to the number of 
the detected points in each image is very small (<16%). In 
other words, this approach is able to handle more than 84% 
of incorrect match (outliers). The results of the least squares 
solution, presented in table (2), give important information 
about the final accuracy of the registration, which is about 
A- 322 
1/10" of the pixel size in the x and y directions. It is 
interesting to note that the accuracy of feature extraction is 
around +1 pixel. This excellent subpixel registration 
accuracy, in the final localization, is obtained because all of 
the points that have been identified as a corresponding pairs 
(328 points) are used in the final adjustment. 
S. CONCLUSIONS 
In this paper we have developed a formulation and 
methodology to handle the problem of image registration in 
an autonomous, robust and accurate manner. In this method 
the problem of image registration is characterized, not by the