Abstraction Moving Objects from H.264 Bit Stream with MAP-MRF

In order to abstract moving objects from H.264 bit stream directly, a novel algorithm is proposed in this paper. Under MAP-MRF(Maximum A Posterior-Markov Random Field) framework, we model the problem by utilizing the MVs(Motion Vector) and inter-prediction modes in P frames to generate energy function. First we retrieve the MVs and the inter-prediction modes of identical 4 × 4 pixels block and make the MVs through low-pass filter to restrain noise. The amplitude of MV and the subMB´size(Macroblock) of the mode are used in our algorithm. The experimental results show that the algorithm can detect moving objects robustly. It also proves that detect moving objects just in compressed domain is feasible.