Research On Sports Video Abstraction In The Compressed Domain

Multimedia analysis and retrieval is one of the hottest issues of the information research. With the development of the computer, communication and multimedia, the amount of multimedia work including the digital video is growing lager and wider. Traditional text-based information retrieval technology cannot analysis the structure of the multimedia effectively. While multimedia analysis and retrieval can provide impactful framework of retrieving the multimedia by extracting the lower feature and obtain the semantic content and become useful both for research and application. Sports video is one of the most desirable video that is enjoyed by lots of funs in the world. But often the length of the sports video is more than an hour. It is tedious and boring for fans and professional to search the highlights in the sports videos. Video abstraction is the summary of video content; it aims to help people to locate the important information quickly and precisely. Hence, it is an important work for the research of multimedia analysis and retrieval. Moreover, video abstraction system can be produced in real time in the compressed domain.This dissertation addresses key techniques of sports video abstraction in the compressed domain. The research focuses on static, dynamic and semantic video abstraction, including interactive model of key video objects selection, camera motion estimation in the MPEG domain, interactive key frame selection model in the compressed domain, replay scene detection, replay scene based sports video summary in the MPEG domain, ontology based sports video annotation and video abstraction. The impressive contributions are as follows.1. An interactive model of key video objects selection (IKVOS) is presented as the result of improving the model of key video objects selection (KVOS) in the compressed domain. The model of KVOS in the compressed domain is proved to satisfy the criterion of induction. Also, the course of IKVOS in the compressed domain is formalized and the model of IKVOS is proved to satisfy the criterion of coinduction. Experimental results are given, showing that the compact and comprehensive scheme is an open system and generates desirable summarizations. Moreover, the distortion rate of compact and comprehensive scheme is lower by 0.8% and 3.8% than the method of Kim [Kim2002], respectively.2. Propose the model of interactive key frame extraction to extend the method of traditional key frame extraction methods in the compressed domain. At first, A M-estimator and epipolar line distance restrained algorithm is presented since a lot of outliers have to be expected in the motion vector fields. Additionally, we used BFGS method to optimize the camera parameters. Moreover, key frame is found by the strategy of shortest path representation and minimal distortion of frames. Finally, interactive key frame extraction is introduced. Experiments show that the distortion rate of shortest path representation and minimal distortion scheme is lower by 8.5% and 9.1% than the method of Porter [Porter2003], respectively. Moreover, the test data validate the extensive applicability of the proposed approach.3. Present a framework of replay scene based video abstraction in MPEG sports video. At first, we address a new technique of identifying the replay boundary by using MPEG feature including macroblock and motion vector that is easy extract from MPEG video, then modify...