Particle filter with least square fitting prediction and spatial relationship based multi-view elimination for 3D Volleyball players tracking

Multiple players tracking in volleyball video analysis is very important for developing applications such as tactical analysis system. To obtain a high success rate of tracking, frequent occlusion among players is a problem to be solved. This paper proposes a least square fitting prediction model and a spatial relationship based multi-view elimination method based on a particle filter scheme in 3D space. The prediction model applies a least square fitting to positions in several previous time steps, which can predict player's position during occlusion accurately. The elimination method eliminates other players' regions based on distances between camera and players' positions, which distinguishes players separately and avoids feature loss in severe occlusion. Experiments conducted on videos of the Final Game of 2014 Japan Inter High School Games of Men's Volleyball in Tokyo Metropolitan Gymnasium show that this multiple players tracking algorithm achieves an average tracking success rate of 97.05%.