Spatiotemporal segmentation of moving video objects over MPEG compressed domain

Abstract

We propose an approach to unsupervised segmentation of moving video objects (VOs) over the MPEG compressed domain. The proposed algorithm utilizes the homogeneity property of the spatiotemporally localized VO's information (macroblock motion vectors (MVs) and DCT's DC coefficients) in order to achieve segmentation with an accuracy of 8/spl times/8 DCT block size. First, macroblock MVs are utilized to identify the locations of moving VOs. DC coefficients are then exploited to achieve finer boundary segmentation. For achieving both objectives, a maximum entropy fuzzy clustering algorithm is proposed to classify MVs and DC coefficients into homogeneous regions, respectively. Experimental results show that the developed algorithm can accurately segment VOs with an accuracy of 8/spl times/8 DCT block size without any user intervention.