Compressed-domain moving region segmentation with pixel precision using motion integration

In this paper, we present an efficient algorithm for moving region segmentation which operates exclusively in the compressed domain, and yet provides region boundaries with pixel precision. Block-based motion vectors (MVs) are the only information used in the segmentation. First, MVs are-homogeneous regions. Second, the boundary regions are identified, and the density of motion field in boundary regions is increased by integrating MVs from previous frames on a pixel basis. Finally, the region boundaries are refined via contour evolution. Experimental results demonstrate that MV integration can greatly improve the segmentation sensitivity, and achieve pixel-precision boundaries in the compressed domain.