Hierarchical segmentation-based approach to motion analysis

摘要

This paper presents an algorithm for motion analysis that has been used for identifying moving targets in a sequence of images. The algorithm is based on first generating a hierarchical segmentation of each image and then identifying corresponding regions from image to image. Using hierarchical image descriptions allows regions to be matched from frame to frame even if scene changes cause the regions to appear at different scales in the segmented image. Hierarchical segmentation also allows the algorithm to isolate the stable regions in each image which are most effective for image registration and to focus computation on the range of resolutions that are relevant for a particular application. Performing registration at the region level has severtal important advantages over traditional optical flow or feature point matching algorithms. These advantages include better performance in the presence of scene changes and noise and the capability for identifying successfully large motions between frames. Following segment registration, the algorithm estimates the motion of the background and of candidate targets. Moving targets are identified based on their motion relative to the background. The effectiveness of the algorithm is demonstrated by experiments on image sequences containing large motions.