The Research On Moving Object Extraction Algorithm In MPEG Compressed Domain

Extraction of video object is a hot research topic in the field of video signal processing, which are applied in many fields, e.g. object based coding, intelligent video surveillance system, face detection, target identification, video database indexing, video summary etc. However, most video data is stored and transported in compressed format because of the huge memory occupation. The pixel-domain based extraction approaches will fail because decoding the compressed data to un-compressed form, which is usually computational intensive. Therefore, it is necessary to extract the video object directly from compressed stream. As a result, the current research hot topic is transferred from pixel domain to compressed domain. But the information utilized in the compressed-domain extraction process is not abundant and noisy, which results in the coarse object boundary. So the key to compressed domain motion object extraction is to trade off between real-time performance and high accuracy requirement.The thesis focuses on the extraction of moving object from the MPEG-2 compressed stream. Main contributions are summarized as follows:(1)Global motion estimation (GME) and compensation (GMC) in compressed domain are investigated. According to the resemblance and consistency between the background motion vectors and global motion, global motion can be estimated from the background motion vectors. Thus, the foreground motion vectors can be acquired through the global motion compensation on the original motion vector fields. After global motion estimation and compensation, the following extraction scheme is feasible for almost any sequence including the static scene and the dynamic scene.(2)A spatial-temporal information fusion based scheme for motion object extraction in compressed domain is proposed. Firstly, the DC+2AC image of I frame is re-constructed, which is used to obtain the contour feature by image segmentation. Secondly, the sparse motion vector field of P frame, which is directly extracted by partial entropy decoding, is intensified to make it dense and then projected to I frame iteratively. Coarse moving object is extracted by the fusion of spatio-temporal segmentation. Then tradeoff is made between the executive time and results'accuracy. High accuracy is acquired at the cost of partially decoding time and region-growing processing time, which is worthwhile at the expense of such a little more time to refine the result. Experimental results demonstrate that the proposed approach can obtain satisfactory segmentation results and can be used in real-time system because of its superiority in time complexity.(3)An improved compressed-domain motion object extraction algorithm for MPEG video stream is proposed. The motion vector field directly obtained from the video stream is firstly processed by reliability test, densification and filtering, and block-based region merging is used to obtain the rough region of every motion object. These motion regions are partially decoded to get the original pixel values, which are modeled with Gaussian mixture model by statistical method. The final accurate motion objects are extracted by threshold decision. Simulation results demonstrate that the proposed approach can get rid of the noisy motion vectors, and the accuracy of segmentation is greatly improved. It can also achieve a high processing speed.