A novel motion estimation technique: patch-matching

The emerging MPEG-4 and MPEG-7 content-based facilities demands segmentation of moving semantic objects in video sequences. Motion is one of main information sources that segmentation algorithms can rely on. In this paper, the authors propose a novel motion estimation technique: patch-matching (PM). The motion of image patches of piecewise smooth intensity is modeled by two parameters and estimated using SAD-based matching. Simple-parametric-motion-model-and-small-support-region is preferred here to complex-parametric-motion-model-and-large-support-region. Both forward and backward PMs are performed to discriminate occlusion regions. Experiment results show the performance of the technique.