Model Selection based Parallel Prediction and Image-trajectory-independent Estimation for Real-time Ball Data Acquisition in Volleyball

Real-time automatic acquisition of the game data plays key roles in the applications of sports analysis such as the broadcasting TV contents and on-line strategy development. In the volleyball analysis, the acquisition of the ball physical data (position and velocity) and the event data (what kind of play is done, such as serve and spike) are quite involved in the further game analysis. Besides the 3D ball tracking, which is a part of the target algorithm based on particle filter, the large time consuming are caused by high-dimensional state vector and adaptive changing system models. To achieve the real-time ball physical and event data acquisition as well as keep a reasonable detection rate, the model selection based parallel prediction and the image-trajectory-independent event estimation are proposed. Firstly, the model selection based parallel prediction solves the problem of different tasks between threads. Secondly, the image-trajectory-independent event estimation separates the observation of event change into the image based observation and the trajectory based estimation, which reduces the observation tasks. The experiments are based on videos of the final game of 2014 Japan Inter High School Games of Men's Volleyball in Tokyo Metropolitan Gymnasium, which are recorded from 4 corners of the court. On the GPU device GeForce GTX 1080Ti, the presented data acquisition system achieves real-time on 60fps videos and keeps the detection rate higher than 90%.