Compressed domain moving object extraction algorithm for MPEG-2 video stream

In this paper, a compressed domain moving object extraction algorithm is proposed for MPEG-2 video stream. It is mainly based on the histogram analysis of motion vectors, which can be easily obtained by partially decoding the MPEG-2 video stream. The whole algorithm framework can be divided into three key steps: motion vector pre-processing, histogram analysis of motion vector and motion vector similarity based region growing for final mask generation. A piecewise cubic hermit interpolation is utilized to form a dense motion field. The outputs of region growing algorithm based on similarity matching are the final segmentation results of moving object. These final segmentation results are further smoothed and interpolated by B-spline curve estimation. Experimental results on several test sequences demonstrate that desirable segmentation results are obtained. The accuracy of segmentation results is improved obviously, nearly to pixel level accuracy because of B-spline curve representation of segmented object. For segmentation efficiency, the processing speed is about 30ms per frame, which can meet the requirements of real time applications.