The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
648 
point to the tail point of the chain, as shown in figure 4, is 
above the threshold value N ; 
(3) The vertical distances are calculated from every points in 
the chain to AB . The maximum value M of the vertical 
distances and corresponding point D are recorded. 
(4) This chain represents a straight line segment if M < A. 
The straight line is fitted using all pixel points in the chain. 
Two endpoints of the straight line segment are determined 
by point A and B. The endpoints should not exceed the 
scope area of A and B .Then the process turns to the next 
chain. 
(5) This chain does not represent a straight line if M > A . 
The chain is divided into two chains AD and DB. The 
process turns to (1). 
With the algorithm above, all straight line segments in edge 
chains are extracted. In practice only prominent features are 
used in registration process. So it is necessary to pick prominent 
straight lines. A convenient rule is to hold those straight line 
segments, which have the maximum length. Figure 5 shows the 
result of straight line segment extraction from airborne image, 
where N = 2Q pixels, A = 2.0 pixels. Top left comer of image 
is shown in figure 5. 
Figure 5. Result of straight line segment extraction 
4. AUTOMATIC MATCHING OF STRAIGHT LINE 
SEGMENTS 
This is the critical step for the whole process of automatic 
registration. In this article we suppose that the images have 
preliminary registered. There are some approaches to get 
preliminary registered airborne image sequences. If video scene 
changes slowly, just shortening sampling interval can obtain 
preliminary registered image sequences. If video scene changes 
quickly or sampling interval is long, we can do rough geometric 
rectification for all image sequences, using the parameters of 
platform position and sensor attitude. Thus all images are taken 
to one reference coordinate frame. But because of the error of 
parameters, in this way we can not get precise rectified images. 
So image registration process is necessary yet. 
In this article, three constraints of corresponding straight line 
segments in the preliminary registered images are proposed as 
follows. 
(1) The distance of the mid-points of corresponding straight 
line segments should be below a threshold value M , 
which is determined by the residual shift error of the 
preliminary registered image. 
(2) The included angles between every pairs of corresponding 
straight line segments should equal approximately. This 
angle value is determined by the residual rotation error of 
the preliminary registered image. 
(3) All pairs of corresponding straight line segments should be 
one-to-one correspondence. Every pairs of corresponding 
straight line segments should hold the same set model 
parameters of image registration. 
According to these three constraints, the algorithm of automatic 
straight line segments matching is designed as follows: 
(1) For every straight line segments i in source image, 
calculate distances ‘S from it’s midpoint to the midpoints 
of all straight line segments j in reference image. Every 
straight line segments j is treated as a candidate matching 
'CMj of i if ‘Sj < M ■ 
(2) For every pairs of candidate matching of straight line 
segments, calculate included angles e [0,^/2] between 
them. The included angle Q with the maximum 
probability is found. From point of calculation, this value 
is a space -8,0 p +8] • Any candidate matching 'CM 
is abandoned if 'q. g -S,6 p +8] ■ 
(3) Now, each straight line segments in source image may 
have just one corresponding straight line segment in 
reference image, may several features, or may no feature. 
To eliminate all wrong correspondences, we establish a 
simple registration model with those one-to-one 
correspondences, and give up all correspondences which 
don’t fit this model. 
The correspondences of straight line segments established by 
the algorithm above are incomplete. That is to say, the 
algorithm doesn’t establish all correspondences of extracted 
straight line segments. But the established correspondences are 
reliable, which satisfies the requirement of image registration. 
5. EXPERIMENTAL RESULT 
Figure 6 shows results of straight line segments extraction, 
matching and image registration of aerial image sequences. The 
reference image and source image are sampled from airborne 
aerial video, with a sampling interval of 0.1 second. In both 
reference and source image 50 prominent straight line segments 
are extracted, respectively. The top left comers of images are 
shown in figure 6. Although the source image is preliminarily 
aligned to the reference image, there is offset in y direction. The 
source image is registered with a second-order polynomial 
model. In registered destination image, the offset is corrected.