Foul Detection for Table Tennis Serves Using Deep Learning

Abstract

Detecting serve fouls in table tennis is critical for ensuring fair play. This paper explores the development of foul detection of table tennis serves by leveraging 3D ball trajectory analysis and deep learning techniques. Using a multi-camera setup and a custom dataset, we employed You Only Look Once (YOLO) models for ball detection and Transformers for critical trajectory point identification. We achieved 87.52% precision in detecting fast-moving balls and an F1 score of 0.93 in recognizing critical serve points such as the throw, highest, and hit points. These results enable precise serve segmentation and robust foul detection based on criteria like toss height and vertical angle compliance. The approach simplifies traditional methods by focusing solely on the ball motion, eliminating computationally intensive pose estimation. Despite limitations such as a controlled experimental environment, the findings demonstrate the feasibility of artificial intelligence (AI)-driven referee systems for table tennis games, providing a foundation for broader applications in sports officiating.