The H264 video compression standard is extensively applied thanks to its excellent compression efficiency and coding flexibility. A moving object segmentation approach in H264 compressed domain is proposed in this paper. The motion fields are first extracted from the compressed video

in which the noise vectors are removed by weighted median filter. Then the predicted motion fields reconstructed by backward estimation are used to accumulate the motion field. After that

the modified statistical region merging is exploited to segment the moving object based on three motion characteristics magnitude

divergence and curl. Experimental results demonstrate that our approach can efficiently extract the moving objects from H264 compressed and as the segmentation quality is good.