Three-frame corner matching and moving object extraction in a sequence of images

This paper presents a three-frame matching method for finding the correspondences of corner points. After a two-stage corner detector is applied to each frame to extract a set of corner points as the matching primitives, candidate transition paths, which are formed by three corner points among three consecutive corner sets, are found by utilizing the smoothness constraint of motion due to inertia. Initially, each transition path is assigned an initial probability of being correct transition based on the similarity of curvatures of the three corner points. These probabilities are iteratively modified by a relaxation process according to the consistency properties of both acceleration and velocity. After several iterations, the paths with sufficiently high probabilities are taken as the correct transition paths. A new segmentation process which integrates both velocity and contrast information is presented to extract regions of moving objects. Several experimental results show that the approach is very effective.