Recognition of soccer players after occlusions using temporal color signatures

This paper deals with the automatic assignment of players' identities in images of sport scenes. The task to correctly assign a player id to one unique player over many frames remains a challenging task as the players usually wear nearly identical jerseys and therefore generate visually similar appearances. If low image resolutions prohibit a direct jersey number recognition, tiny visual differences between the players like shoe, hair or skin colors can still support the recognition process. We propose a novel system which gathers such visual differences of players over multiple frames and allows for a successful identification of players. Results are presented in the context of a merge and split handling module of a player tracking system where the recognition can significantly improve the correct assignment of players after ambiguous merge and split situations.