Mean shift tracking of soccer-players using motion and color likelihood images

To analyze a soccer game, it is useful to gather statistics of the players. In order to do this automatically, players can be tracked on the field. This paper presents an approach for tracking multiple players from a video of a soccer match. Players are tracked using a mean shift algorithm which operates on player likelihood images. These images are created using a combination of motion likelihood, implementing background subtraction, and color likelihood, obtained through classification with a neural network. With this mean shift algorithm, the system is able to track all players ~100% accurate in typical situations with a controlled environment. Two player occlusions are automatically resolved and the algorithm is able to detect whether a player’s track is lost after which a human operator is asked to assign the correct position