Effective corner matching based on Delaunay triangulation

Matching corners between images is an important and difficult problem in stereo vision and many other vision applications, and no effective method has been developed to cope with general cases. In this paper, we present an improved SUSAN (smallest univalue segment assimilating nucleus) corner detection algorithm and an effective algorithm to establish corner correspondence between two images based on Delaunay triangulation. First we construct Delaunay triangulations among corners of each image and compute interior angles of the triangles. The corner correspondence is established based on an observation that these angles completely and uniquely characterize the corners and their values are not affected by scale change, less affected by rotation and translation to some extent. At the matching stage, we first obtain the most similar triangle pairs, and then extend their edges circularly until all matching corners are triangulated and mismatching corners are discarded. Experimental results are provided which illustrate the good performance of the algorithm.