Establishing motion-based feature point correspondence

摘要

Given a sequence of frames corresponding to a dynamic scene, multiframe correspondence of a set of selected image points, called feature or interest points, determine the motion trajectories of those points. This paper describes a new polynomial time algorithm for finding feature point trajectories from a large sequence of frames. The proposed approach utilizes a constraint, called motion uniformity constraint. According to the motion uniformity constraint, the changes in the motion characteristics of a point are small over a short period of time and are temporally continuous. This implies that the trajectory of a point over a given time interval is as “smooth” and “short” as possible. The proposed approach permits point trajectories to terminate and start at intermediate frames of the sequence. For a given sequence, the trajectory set is obtained by considering every set of three consecutive frames and finding the locally best local trajectories (in terms of motion uniformity) that extend some of the existing point trajectories and/or define new point trajectories. Experimental results are presented to demonstrate the effectiveness of the proposed approach.