基于块特征的H.264压缩域对象分割

Block Characteristic Based Object Segmentation in H.264 Compressed Domain

针对H.264压缩域,提出一种改进的运动对象分割方法。首先提取运动矢量场,对该矢量场进行矢量中值滤波,然后基于块进行前向矢量累积获得更加显著的矢量场,最后采用改进的K均值和EM的混合分层聚类方法,基于宏块层、块层两次聚类将运动对象分割出来。实验表明,该方法通过分层聚类,进一步提高了分割效率,能实现满足实时要求的较精确分割。

An improved moving object segmentation approach which extracts motion field from H.264 compressed domain is proposed. Pre-treatments such as vector median filtering and forward block vector accumulation are used to obtain more obvious motion field. Then mix and hierarchical clustering algorithm based on improved k-means and EM is exploited to segment the moving object on the macro-block level and on the block level. Experimental results demonstrate that this approach further improves the efficiency of the segmentation by hierarchical clustering. It provides satisfactory real-time segmentation results and good segmentation quality.