A Study Of Topic Modeling And Content Analysis For Soccer Video

With the rapid development of technologies of computer and multimedia, there is anexplosive growth in the amount of available multimedia information. Video is one of the mostimportant components of the multimedia data, which has huge quantity and complex semantic.As an important genre of video document, sports video has attracted increasing attention invideo content analysis due to its wide viewership and tremendous commercial potential.The research of this dissertation focuses on the problem of soccer video content analysis.By solving some problems such as scene analysis, detection of field line, camera calibration,extraction of Real-world trajectory, and tactic analysis, this dissertation proposes a novelapproach for soccer video analysis based on the research on multimedia analysis, patternrecognition and computer vision, such as topic modeling. The detailed description of theresearch content in the dissertation is as follows:(1) A novel soccer scene analysis method based on topic model is proposed, which canstructuralize the soccer video from the perspective of semantic. First, the concept of semanticshot segmentation is introduced. By automatic extraction of semantic color, the video can besegmented and classified into different scenes according to the focal length. Secondly, thelong-view shots are further classified into six different scenes according to which area of thesoccer playfield it contains. To improve the performance of scene classification, a novel topicmodel is proposed by incorporating spatial contextual information.(2) A novel field line detection method is proposed based on the directional informationof field line. First, the special character of modern soccer playfield is considered, and a novelsoccer field model is introduced which includes offside assistant lines. By designing adirectional filter, these field lines can be accurately detected. Moreover, the field linedetection and tracking for a clip is developed, which can further improve the performance offield line detection.(3) A novel continuous camera calibration algorithm is proposed, which can address thecalibration for non-goal area of soccer video. By considering the fact that the position of maincamera is stable, a novel continuous camera calibration algorithm is proposed for soccer videoanalysis. There are two stages for calibration: the first stage addresses the calibration only forgoal area and the position of camera is computed and recorded, and then the second stage canaddress calibration for any area including non-goal area where the camera position is given atfirst stage. Thus, the proposed algorithm can address the calibration for non-goal area ofsoccer video, which cannot be addressed by traditional calibration algorithms since the inadequate of calibration objects.(4) A novel Real-world trajectory based tactic analysis method is proposed, which canautomatically analyze the attack tactics of soccer game from the perspective of professionals.First, a real-world trajectory extraction method is proposed, and then an attack patternrecognition method is described based on the definition of the ball states. Our experiments onuser study by soccer professionals demonstrate that the defined attack patterns can be used toanalyze soccer tactics effectively in terms of conciseness, clarity and usability.In this paper, we integrate the basic theories of multimedia analysis, pattern recognitionand computer vision into content analysis for sports video analysis. Through exploring severalcomponents such as scene analysis, object detection, camera calibration, trajectory extraction,and tactic analysis, a systematical system is built up. The proposed algorithms can be not onlyused for soccer video analysis but also used for other sports video analysis. It provides apromising method for general multimedia analysis.