Automatic Tracking of Indoor Soccer Players Using Videos from Multiple Cameras

Indoor soccer has been of tactical and scientific interest, with applications dedicated to analyze tactical and physiological factors and also physical training. In both cases, the analysis is based on player tracking, done with human supervision. This paper presents an automatic tracking method which shows the trajectories of indoor soccer players during the game and saving skilled labor during the process. For this, we use a predictive filter to model the motion and the observation of multiple stationary cameras, strategically positioned around the court. We associate a particle filter to a robust probabilistic observation model with the measurement in court coordinates. The observation model proposed is based on data fusion across multiple camera coordinates and projected onto the court plane, creating a multimodal and bidirectional probability function, which represents the potential localization of players in the court plane. The probability function uses an appearance model to observe player´s location, distinguishing very close players and yielding good weights in the observation model. The experimental results show tracking errors below 70 centimeters in most cases and indicate the potential of the method to help sports teams.