Image Matching Based On Interest Points And Gray-value Differential Invariants

Interest points are useful low level features where the image gray-value changes two dimensionally. The detection of interest points is the basis of kinds of computer vision applications, such as: camera calibration, 3D reconstruction, image matching, video retrieval, motion estimation, etc. In this paper, three impersonal criteria:delocalization, false-detection rate, and repeatability are presented to evaluate the performance of an interest points detection algorithm. Based on these three criteria, an improved interest points detection algorithm is presented after we theoretically analyze the characteristics around the interest points and compare different detectors according to the criteria specified above. Good experiment results show that our algorithm is robust enough to solve the delocalization and false detection problem. And at the same time, our algorithm also gives high repeatability. Once the interest points were detected, the image matching process in an image sequence is performed using local gray-value differential invariants. The local gray-value differential invariants are robust to image occlusions, clippings, and rotations, since they are locally calculated and rotationally invariant. But they are very sensitive to image noise and points delocalization error. In order to use the invariants, the local geometric constraint must be added. Experiments have shown a quite high correct matching rate using the method specified above. Finally, the only available geometric constraint, namely, the epipolar constraint, is exploited robustly using the above initial set of matches. More accurate matches are eventually found, as in stereo matching, by using the recovered epipolar geometry. The good performance of the proposed algorithm was justified by extensive experiments on real images. The results show a very fast matching speed and a very high correct matching rate.