Automated moving object segmentation framework with chained optical flow analysis

Abstract

This paper proposes a general framework to apply an interactive image segmentation method (i.e. MRF modeled graph-cut segmentation) to virtually any video sequence for moving object segmentation without human intervention. Although computed motion information of an object between two consecutive frames is an important cue to differentiate the object from its background, applying error-prone motion cues as initial seeds for graph-cut segmentation results in inacceptable errors. In this work, a novel method to estimate the background seeds using estimated object contours and distance transformation is proposed. More importantly for accurate foreground seeds, a novel Chained Optical Flow Analysis (COFA) effectively differentiates incorrect foreground seeds from true initial seeds. Not only structured errors caused by the nature of frame difference motion cue computation, but incorrect seeds due to arbitrary noises are also eliminated by the analysis. With refined seeds, the graph-cut segmentation method produces robust segmentation of a moving object in the image frame. By calibrating the seed data, proposed method enhances the accuracy of segmentation and derives the maximum utilization of MRF modeled graph-cut segmentation method at frame level.