Research On Moving Object Segmentation In H.264/AVC Compressed Domain

Video object (VO) segmentation aims to partition a video sequence into moving objects and track the evolution of the moving objects along the time axis. Many applications related to video index,intelligent monitor and pattern recognition rely on video object segmentation. Video object segmentation techniques are also important tools for content-based video coding,manipulation and interactive multimedia applications.At present, most methods about VO segmentation are carried out in the pixel domain which cost much time and memories. With advances in multimedia and video coding standard, more and more video information is compressed format, so it is significant to research the technique of VO segmentation in compressed domain that is especially useful in real-time environment, because there is no need to fully decode compressed streams and it can decrease a mass of datum compared with the methods in pixel domain.The research about VO segmentation in compressed domain is a new direction in the area of object segmentation and the existing technique isn't much. Some research of the VO segmentation in compressed domain is done in this dissertation. Firstly, all kinds of methods of VO segmentation in compressed domain are introduced,classified in detail,and a performance evaluation is given. Secondly, a new spatiotemporal method of VO segmentation in H.264 compressed domain is proposed which mainly uses DCT,motion vector(MV) in video streams , and the technologies quantized DCT coefficient difference,MV normalization,weighted extended vector median filter,morphological processing,motion projection are combined to abstract objects. Thirdly, aiming to eliminate the effect of camera motion in VO segmentation, an algorithm of based on foreground and background non-iterative global motion estimation (GME) in H.264 compressed domain is brought forward, it utilizes DCT and MV to partition background foreground to remove outliers, and adopts a non-iterative method by choosing least MV samples in background and integrating Singular Value Decomposition to estimate parameters of global motion model. Lastly, a prospect of VO segmentation in compressed domain is put up based on the analysis of existing segmentation techniques. The contrastive experiment is designed to test the proposed methods: spatiotemporal VO segmentation in H.264 compressed domain, based on foreground and background non-iterative GME in H.264 compressed domain on different standard sequences, the results proved the validity; meanwhile, the methods are transplanted easily to compressed domain with other video compression standards.