Sift-based multi-view cooperative tracking for soccer video

This paper presents a SIFT-based multi-view cooperative tracking scheme for multiple player tracking in soccer games. We assume that future sports events will be captured by an array of fixed high-definition cameras which provide multi-view video sequences. The imagery will then be used to provide a free-viewpoint networked experience. In this work, SIFT features are used to extract the interview and inter-frame correlation among related views. Hence, accurate 3D information of each player can be efficiently utilized for real time multiple player tracking. By sharing the 3D information with all cameras and exploiting the perspective diversity of the multi-camera system, occlusion problems can be solved effectively. The extracted 3D information improves the average reliability of tracking by more than 10% when compared to SIFT-based 2D tracking.