Player tracking using prediction after intersection based particle filter for volleyball match video

Player tracking plays a key role in volleyball tactics analysis. Therefore, many tracking algorithms have been proposed to track and locate the players´ positions. However, in a volleyball match, intersection of players wearing the same uniform occurs frequently. Conventional tracking algorithms can hardly handle such complicated situation with high accuracy. This paper proposes a prediction after intersection based Particle Filter to track players distinctively in intersection events of volleyball. As for likelihood models in the proposed particle filter, a labeling likelihood to distinguish each tracking object and a distance likelihood to prevent particles from concentrating in only one tracking object are employed. As for prediction models, two re-detection models, Crash Pattern Model (CPM) and Others Pattern Model (OPM), are used to detect each tracking object after intersection. Experimental results reveal that the proposed algorithm attains high tracking accuracy for various kinds of intersection scenes. The tracking accuracy of player using whole scenes of a volleyball match is around 80% while conventional method is only 30%.