The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
610 
Feature extraction and matching are prepared for image 
registration. The image registration that implements frame-to- 
frame registration of the video image sequence is the key point 
of motion compensation. The result of image registration could 
be used in two directions, image stabilization and image 
mosaicking. Former can restrain the moving background and 
facilitate the detecting of moving target, and latter can update 
the local image (always express with the ortho-image) and help 
to form the trajectory of tracked object. 
RANSAC—random sample consensus algorithm (Fischler et al., 
1981) is a nonlinear algorithm. Fitting data model with 
RANSAC maximally restrains the impact of outliers, and 
reduces the computation to a certain extent. The fine matching 
is fitting the fundamental matrix through iteration computing 
and identifying most of the outliers. Figure 2 presents the 
results of matching after eliminating wrong correspondences 
from the candidate matches which got from the coarse matching. 
It can be seen that though bulk of mismatches have been 
removed, there still a few incorrect correspondences remain. 
2.1 Feature Extraction and Matching 
In feature extraction, choosing a right kind of feature should be 
considered for one thing. The feature could be point, line or 
surface. It has been proven that comer feature is robust and easy 
to operate. Harris operator (Harris et al., 1988) is a typical 
comer detector, and its principle is that recognising the features 
by judging the difference of gray-level’s change while moving 
the search window. Detecting results of two series frames 
shown in figure 1, and there is good coherence between the two, 
so it should be thought that the operator has a stable 
performance and the results could be taken as the input of 
matching. 
Figure 1. Detecting results using Harris comer operator 
After extracting the features, a coarse matching would be made 
to get approximate matching results, and this course is realized 
by measuring the similarity of corresponding features. Because 
there are many mismatches in the approximate results and they 
cannot meet the requirements of registration, so it has to 
implement a fine matching to remove the mismatches. 
Figure 2. Overlay of two successive frames after 
eliminating wrong correspondences with RANSAC 
A suitable way to keep inliers is combining of epipolar 
geometry and RANSAC algorithm. Epipolar geometry offers a 
model—fundamental matrix to the matching, cause the two 
views should satisfy the epipolar restriction in stereo vision. 
2.2 Image Stabilization 
Image stabilization is compensation of unwanted motion in 
image sequences. The matter of image stabilization is image 
registration. The transformation model of image registration is 
not complicate. A usual choice is affine transformation or 
projective transformation. 
Figure 3. The comparation of difference results before and 
after image registration 
The normal mode for registration is calculating the parameters 
of the model using corresponding points. Whether the precision 
of image registration is good or not depends on the results of 
matching. So image stabilization could be done by computing 
the registration parameters with the outputs of fine matching 
and rectifying the prepared frame to reference frame. In order to 
optimize the result of registration, repeating the course until the 
accuracy of registration good enough. Figure 3 shows the 
comparation of difference results before and after image 
registration. The left one is the difference result previous 
registration. Except some regions with same textures, most of 
the background image can not be subtracted, especially some 
obvious objects and linear features. The right one is the 
difference result after image registration. Though there are 
objects edges still distinct, majority of background image got 
better elimination. 
2.3 Image Mosaicking 
Mosaicking of video image sequence is rectifying all frames to 
the reference frame and piecing them together as a panoramic 
image. The reference frame may be the first frame or a chosen 
one. A key step for the generation of panorama is image 
registration. 
It is unavoidable accumulate registration errors during aligning 
the image sequences. The accumulation of errors could induce 
misalignment of adjoining frames. To resolve the problem, 
there are many methods have been tried, such as refining 
registration and introducing reference data. An UAV video 
image mosaicking is illustrated in figure 4, and there are some 
piecing seams for registration errors.