An efficient approach to extract moving objects by the H.264 compressed-domain features

Abstract

A new efficient algorithm on H.264 compressed-domain is proposed for extracting moving objects. The block entropy feature from the compressed-domain video is proposed to handle the bitstream with variable quantization QP values. Each video block builds two background entropy models to reflect the different coding efficiency between INTER and INTRA block types. Through background subtraction, the foreground energy frame is built after a normalization process. The hysteresis thresholding, where the thresholds can be easily obtained by an efficient unsupervised approach, is adopted with slightly modification to overcome the over-segmentation in extracting moving objects. At the final stage, morphological filters and connected components labeling(CCL) are applied to extract the list of moving objects. In our experiments, the proposed approach is able to operate faster than real time and the accurate object extraction can be achieved for huge H.264 surveillance videos.