The Research On Moving Object Segmentation In The Compressed Domain

Video segmentation refers to partitioning the video frames into a series of spatially correlated regions in the temporal axis. This kind of region-based image representation makes it possible for flexible video processing. It has great potentials in video coding, video browsing and multimedia interaction. However, video are usually stored and transported in compressed form because of its high data capacity, and video segmentation in the compressed domain is usually faster than that in the spatial domain. Therefore, research on this topic has focused on compressed domain approaches.Video object segmentation is a difficult problem in the field of video processing and computer vision. The main difficulties lie in two aspects. First, the extreme complexity and proliferation of scenes in the real world is hard to be modeled with a uniform model. Second, the high-level semantic of video object is hard to be accurately defined and described with low level features such as motion and texture. Therefore, current video object segmentation puts emphasis on moving object segmentation, which is considered for practice, as it is considered in this dissertation.After briefly introducing the concept of video object and the fundamental principles involved in it, currently available moving object segmentation algorithms in the literature are reviewed, and their performances are evaluated and discussed. Two compressed domain moving object segmentation algorithms are proposed. For the most widely used MPEG streams, the segmentation algorithm is based on gradient model. The DCT coefficient, are utilized to get all the object edges, and then they are synthesized with the information of edge motions to obtain the interested object edges. For the video streams, which are compressed by the latest video coding standard H.264/AVC, a region-tracking based segmentation approach is proposed. It utilizes the new characters of H.264 by partial decoding the motion vectors as preprocessing. Region segmentation is performed on the reliable motion vector field for every frame, and finally tracking these regions to get the desired moving objects. The simulation results demonstrate that the proposed two algorithms can obtain satisfactory segmentation results, and are of high application value.