Automated Event Detection Based on Ball Trajectory Analysis for Tennis Match

Event detection plays a significant role in tennis match analysis. Current methods for event detection mostly rely on cues from audience sounds and players’ positions and are confined to videos captured by professional facilities. The performance of these methods suffers from audio interference and the obscuration of players. In this paper, we propose a new event detection method based on ball trajectories for tennis matches. Our method learns event-related features from the trajectory, displacement, and flight direction of a tennis ball in successive frames and uses the contextual information of the ball trajectory to mitigate quality degradation of the trajectory under certain conditions. We also manually annotate each frame of video with a label of events in the dataset of GridTrackNet, which encompasses tennis videos captured from both professional and amateur footage, and designate it as GridTrackNet-A. Experiment results show the proposed method obtained superior performance with macro F1-score of 91.65% and MAE of 0.63 on the GridTrackNet-A dataset.