Spatial pixels selection and inter-frame combined likelihood based observation for 60 fps 3d tracking of twelve volleyball players on gpu

Abstract

3D players tracking plays an important role in sports analysis. Tracking of players contributes to high level game analysis such as tactic analysis and commercial applications such as TV contents. Many services like sports live and broadcasting have strict limitation on processing time, thus real-time implementation for 3D players tracking is necessary. This paper proposes a particle filter based 60 fps multi-view volleyball players tracking system on GPU platform. There are three proposals: body region constraint prediction, spatial pixels selection and inter-frame combined likelihood. The body region constraint prediction uses player’s body region as limitation in prediction to increase tracking accuracy. The spatial pixels selection method selects pixels for likelihood calculating to reduce calculation amount in spatial space. The inter-frame observation method does particle filter algorithm with two frames each time to reduce calculation amount in temporal space. Our experiments are based on videos of the Final and Semi-Final Game of 2014 Japan Inter High School Games of Men’s Volleyball in Tokyo Metropolitan Gymnasium. On the GPU device GeForce GTX 1080Ti, our tracking system achieves real-time on 60 fps videos and keeps the tracking accuracy higher than 97%.