Real-time Tracking of Player Identities in Team Sports

Sports video analysis is a common tool in team sports. It requires keeping track of player positions with corresponding identities during the game and analyzing the resulting trajectories at an abstract level. This thesis proposes a distributed cognitive system for automating this task. The tracking process includes information fusion and building as well as adapting models of the tracked players online; the system supports automated team behavior summarization and offers further analysis in a conceptualization framework. Our contributions are (1) an innovative, general multi-target tracking approach that outperforms current state-of-the-art algorithms, (2) adaptive methods for identifying players based on appearance and/or spatial relations with the help of extended self-organizing neural networks, and (3) the implementation of a concrete real-time tracking system for recorded soccer games. Research results are validated in various challenging domains, including full-length soccer match videos and broadcasted material.