Automatic Acquisition of Motion Trajectories: Tracking Hockey Players

We address the problem of automatically analyzing hockey scenes by estimating the panning, tilting and zooming parameters of the broadcasting cameras, tracking hockey players in these scenes, and constructing a visual description of the scenes as trajectories of those players. Given quite fast and non-smooth camera motions to capture highly complex and dynamic scenes of hockey, tracking hockey players that are small blob-like, non-rigid and amorphous becomes an extremely difficult task. We suggest a new method of automatically computing the mappings to represent the globally consistent map of the hockey scenes by removing camera motions, and implement a colorbased sequential Monte Carlo tracker to track hockey players to estimate their real world position on the rink. The result demonstrates a quite successful performance on both objectives. We make two new contributions in this research. First, we introduce a new model fitting algorithm to reduce projection errors. Second, we use an adaptive model to improve the current state-of-art color-based probablistic tracker. Our approach is also applicable for video annotation in other sports, surveillance, or many other situations that require obect tracking on a planar surface. Since there have not been any hockey annotation systems developed in the past, we hope that our system would become a stepping stone for automatic video annotation in hockey.