Sports Video has being a hot research topic for its wide viewer-ship and enormous application potential in recent years. The objectives and applications of sports video processing and analysis include: highlight extraction, video summarization, browsing, semantic event detection and retrieval, video service customizatino, video content editing, enhancement and richment, etc. This thesis focuses on mid-level feature extraction in sports video. The main content consists of:(1) Decision tree based seamntic shot classification is proposed. Firstly, several important shot types are prior defined. Then, color, texture and shape features are combined to perform view classification. The final shot type is voted by views contained in the shot. (2) Automatic replay scene detection based on replay-logo and shot context is proposed. Automatic logo detection algorithm is used to locate the boundaries of replay scenes. Then, motion and shot context cue are utilized to discriminate replay scene. This method can not only accurate locate the replay boundaries, but also robust recognize replay scenes. Based on shot classification and replay detection, shoot and red/yellow card events are detected in soccer videos.(3) A unified semantic shot description framework is constructed. The certain scenes of shot generated are analyzed, and the three-factor model is used to characterize a shot. Then, a unified field-ball semantic shot description framework is constructed. Finally, the framework is used to shot clustering and retrieval, video segmentation and semantic analysis.(4) An efficient soccer ball detection and tracking method is proposed. Firstly, the play-field is segmented, and a coarse-to-fine criterion is used to locate the soccer ball. The Monte Carlo based Condensation algorithm is applied in the procedure of ball tracking.(5) Motion analysis and swimming style classification is performed as viewed from local motion. Local player motion is firstly detected through color and motion cues. Then, motion periodicity is estimated based on motion energy. A motion salient frame in each period is selected and finally the motion periodicity and motion feature in salient frame is integrated for swimming style classification.
修改评论