JT 
Ying Chen 
  
The constraints derived from feature chain is expressed as: 
n-l 
cu = 20095 a, -avg,) -( b, avg,)|)> max (12) 
0 «| davg |« T, 
n-l 
C, Seo 2 d iy -avg,) -( b -avg,) | — max (13) 
(T, — 4) <| davg |< T, +4 
in which: avga, avgb and davg are the directions of contour chain a,b (i.e. the average of corrected chain code [Ying, 
1998.4]) and their rotation angle, T, is threshold of rotation angle, generally not over 22.5° The cos is to make 
correlation coefficient Cy, less than 1. Formula (12) shows that correct matching is available if C, approach to 
maximum while the direction of contour start point is same as that of the end point in real-time scene and reference 
image. While davg is near to 4, it maybe exists a rotation of approximately 180° between two contour ‚or start 
point is right-about to end point. Then using formula (13) could avoid false matching. While they are homologous 
features , right matching would be attainable from (13), though start point is right-about to end point. 
When above conditions are satisfied, a reliable matching with one pixel precision can be obtained. According to the 
result, an enough accurate initial value is provided for least square image matching to realize sub-pixel matching. 
After finishing the above initial matching, a series of centroid coordinates(x;y;) and(x;.y;) could be attainable. 
Therefore, corresponding between two different sensor images would be got. Meantime, rotation angles between 
two images could be determined. : 
k = avg, — avg, (14) 
According to this rotation relationship, we can preprocess real-time scene , and make its coordinate system 
homologous to reference image. Finally, least square matching between two images could be done as follows 
[Ying, 1998.4 ]. 
ver TS 
abuso eue 
mu M 
x-|a^ amv af an] 
  
À = 
À = (A" PA) ‘(AT PL) ds 
3 | EXPERIMENT RESULTS AND PRELIMINARY CONCLUSIONS 
In this paper, matching experiments have been made by several SPOT digital images and low altitude CCD images. 
All of these have gained favorable effect. Fig.3 is one example of them. 
In experiments, three pyramid images were used to do matching and edge detection. In order to improve efficiency 
and reliability of edge detection, gradient map must be preprocessed before detection . For this goal ,a histogram 
filter was used so that a great lot of non-characteristic information was removed [Ying,1998.4]. The contour 
feature with and without filter are shown in fig.3. 
During feature matching, several candidate positions that satisfy formula (10) were obtained in top pyramid. 
Sequentially,image least square matching was done in ground pyramid, so that sub-pixel result was obtained. The 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 181