Abstract: A fast moving object extraction method working in MPEG compressed domain is proposed in this paper. The embedded motion vectors and DC coefficients of DCT transformation in an MPEG-like coded video stream are acquired by partially decoding and used as input to the algorithm; the moving object mask for each P frame is extracted. Firstly, the global motion between current frame and the last reference frame is estimated by robust regression analysis, and blocks non-conforming to the estimated global motion are marked as potential moving blocks; Then, taking the upsampled motion vector field as the temporal space feature, and the reconstructed DC image converted to LUV color space as the spatial space feature, a fast mean-shift based clustering procedure is performed to find regions with similar temporal and spatial characteristics and get finer region boundaries; Finally, the moving object mask is obtained by an MRF-based statistical labeling procedure. The experimental results show that the proposed algorithm can effectively suppress the influence of motion vector noises, and have a very fast processing speed (For CIF video streams, the algorithm can run at a speed of 50 frames per second).

Key words: statistical labeling, moving object extraction, compressed domain, fast mean-shift clustering

摘要： 本文提出了一种工作于MPEG压缩域的快速运动目标提取算法。算法以通过部分解码得到的运动向量和亮度分量的直流DCT系数作为输入，提取P帧的运动目标。首先采用鲁棒性回归分析估计全局运动，标记出与全局运动不一致的宏块，得到运动块的分布；然后将运动向量场插值作为时间域的特征，将重构的直流图像转换到LUV颜色空间作为空间域的特征，采用快速平均移聚类找到时间和空间特征具有相似性的区域，得到细化的区域边界；最后结合运动块分布和聚类分析的结果，通过基于马尔可夫随机场的统计标号方法进行背景分离，得到运动目标的掩模。实验结果表明该算法可以有效地消除运动向量噪声的影响，并有很高的处理速度，对于CIF格式的视频码流，每秒可以处理约50帧。

关键词: 统计标号, 运动目标提取, 压缩域, 快速平均移聚类