Robust 2D shape estimation of moving objects considering spatial and temporal coherency in one MAP detection rule

An algorithm for 2D shape estimation of moving objects is proposed, which reduces the estimation error compared to the ISO/MPEG-4 reference. The improvement is achieved by four approaches: firstly, the memory for object masks is motion compensated. Secondly, the robustness against camera noise is enhanced due to averaging the noise in succeeding frames. Thirdly, several steps of the algorithm are combined in one MAP detection rule, which leads to spatially more accurate and temporally more coherent object masks. Fourthly, the consideration of results from a colour segmentation is moved to a more suitable position in the algorithm, in order to avoid estimation errors at object boundaries