Visual attention based video object segmentation in MPEG compressed domain

A novel approach of visual attention based video object segmentation in MPEG compressed domain is proposed in this paper. DCT coefficients and motion vectors (MVs) are firstly parsed from compressed streams. Analysis of the scene texture is then proposed to decide the best appropriate information for region growing. MVs are exploited to perform region growing if texture is complex, otherwise DC coefficients and MVs are used together to perform region growing. Meanwhile, MVs of I frames are calculated by backward projection of MVs of subsequent P frames, then global motion compensation is performed using the MVs of I frames to obtain local MVs. Subsequently statistical region growing is exploited to segment the image into homogeneous regions. Finally an improved attention model is proposed to extract visual attention objects, which is based on position, clearness and local MVs. The attention model makes the segmentation results more accordant with human´s perception. The experimental results conducted on standard test sequences demonstrate the efficient, real-time, robust performance of the proposed approach.