Long-term Player Tracking in the RoboCup Soccer SPL

In order to evaluate progress in RoboCup Soccer, automated video analysis tools are needed to compute game statistics such as ball possession and field coverage. To that end, the players have to be detected, identified, and tracked reliably for the entire duration of the match. Player tracking in the Standard Platorm League is especially challenging since all the players look exactly the same, except for the color of their jersey, so state-of-the-art apperance-based approaches for pedestrian or vehicle tracking would fail. The proposed tracking framework uses non-linear optimization to fuse information from different sources - visual short-term tracking, self-localization from the players, game events, and jersey color detection - and is capable of reliably tracking players over 5 minute sequences with >90% accuracy.