Motion field refinement and region-based motion segmentation

Abstract

In this paper, we propose a method to refine a motion field from image sequences and region-based motion segmentation using the motion information. An initial motion field is generated by a block matching algorithm. We compute the motion profile at each block and define the motion confidence measure from the motion profile. In the refining process, we regulate the motion vectors with low confidence to those with high confidence. In the segmentation stage, each frame of the image sequence is partitioned into regions by a watershed algorithm and a motion vector is assigned to each region. After constructing a region adjacency graph, the graph is segmented by the normalized cuts algorithm. The experiments show that the proposed method provides satisfactory results in motion segmentation from image sequences with or without camera motion. © Springer-Verlag Berlin Heidelberg 2005.