Streaming Video Extraction And Segmentation Of Moving Objects

Moving object segmentation(MOS) aims to partition the moving targets among a continuous sequence of images. MOS is the basis for video analysis, provides important data for content-based coding, video retrieval and video summary, and is widely used in video surveillance, video-making and other fields.After the announcement of MPEG-4, which is typical of content-based coding, persons have paid more attention to the technology of MOS. It includes two directions, one is based on the pixel domain, and the other is based on compressed domain. In the pixel domain, objects are segmented by detecting changes of moving object, and the segmented objects are accurate. But the segmentation algorithm was complicated, more data need to calculate and results depend on specific setting. With the development of coding techniques, the technology of MOS in compressed domain got more attention. The motion vectors (MV) and DCT coefficients generated during video coding in compressed domain can be used, and with them the moving objects could be extracted quickly. It could be real-time, but the segmented objects could be coarse because of coding in blocks. For the above reasons, now few studies combined the compressed domain with pixel domain to achieve a balanced performance.At present, the H.264/AVC(MPEG-4 Part 10) is the prevailing video coding standard, because of its excellent network adaptability and high compression ratio. This paper focused the moving object segmentation on H.264 video stream. As above advantages and shortage of segmentation technologies in compressed domain or pixel domain, this paper adopted the method of extraction in compressed domain and segmentation in pixel domain. Firstly, moving object region was extracted by the motion vectors in H.264 stream, then in pixel domain the region was segmented by MRF (Markov Random Field) or Grabcut to obtain the accurate moving object. The main contents of this paper include:1) The research of motion vectors preprocessor in compressed domain. The filtering method was used, which combined the weighted average filter in time domain with weighted vector median filter in space domain. The accumulation of motion vector, 6 parameters global motion estimation, interpolation was used to enhance the reliability of motion vectors.2) The method of MV differences in amplitude and angle was used to extract the region of moving object in compressed domain.3) Proposed the method of that the region extracted in H.264 compressed domain was used to be the initial marking field in MRF model and then moving objects were segmented accurately using MRF in pixel domain.4) Proposed the method of that Grabcut was used in automatic objects segmentation with the region extracted in compressed domain as the initial interaction of Grabcut in pixel domain.Experimental results show that the algorithms in this paper could segment moving objects automatically and effectively. The algorithms can be used in the sequence of static background and dynamic background, and moving objects segmented was complete and accurate.