立体视频对象分割是交互式多视点视频应用的关键技术。为了提高对象分割的时效性和精确性,提出了一种利用压缩域视差和运动信息的立体视频对象分割算法。该算法首先对运动矢量场和视差场进行提取和修正处理,然后对视频帧进行分割作为初始值,最后用均值偏移算法聚类得到最终的对象分割结果。实验结果表明,对于纹理复杂的场景有很好的分割效果,可以获得与语义一致的对象。
Stereo video object segmentation is a key technology in the multi-view video interacting. In this paper, based on the compressed domain information a novel stereo video object segmentation algorithm is proposed to improve the validity and accuracy of the segmentation. Firstly, we extract motion vectors and disparity field to segment the initial video object. Then we used the mean-shift algorithm to cluster the video object. Experimental results shows that the proposed algorithm could obtain the object with semantic confornity from the scene with complex texture.