Sports video content has been proliferating with the development of the multimedia and communication technology. As a result, sports video analysis has attracted significant attention from both academia and industry. Most existing work on sports video analysis mainly focuses on semantic analysis which aims to detect semantic events in the game. Different from semantic analysis which is totally audience oriented, tactic analysis serving for professionals has been paid more attention to recently. Professionals such as coaches and players are more interested in the tactics frequently used in the games rather than “objective” events. Presenting professionals with tactics in the game can assist them to establish strategy and improve training effect. Besides the introduction of direct inference from low level features to tactics, the thesis mainly focuses on the ball and player trajectory based tactic analysis. In ball games, tactics can be deduced from the movement of the ball and players. Therefore, temporal and spatial trajectories of the ball and players become the appropriate feature for tactic analysis. A robust decomposition-integration framework for ball and player trajectory extraction is proposed, which is specially designed to tackle the challenges in broadcast sports video. Trajectory similarity is defined and tactics are discovered with the classification of ball and player trajectories. The contribution of the thesis consists of: 1)An inference scheme is proposed to discover tactics with low-level features and semantics. The inference is based on the model of the view direction sequence and the rule of the game. 2)A robust decomposition-integration framework with collaborative ball and player trajectory extraction is proposed. The novel framework uses decomposition to contain error in local area and uses a fault-tolerant integration scheme to achieve global optimal trajectory. 3)Motion information is extracted from the trajectories by taking ball and player trajectories as mid-level features. Local feature based trajectory similarity is defined and semi-supervised and supervised tactic detection and classification is conducted with the trajectory similarity measurement. 4)Suffix tree is introduced to describe the local feature of trajectories. Key sub-trajectories are selected with the suffix tree model and most irrespective samples are discarded to improve the effectiveness and efficiency of the training process.
修改评论