BDTA: events classification in table tennis sport using scaled-YOLOv4 framework.

This article focuses on a lightweight deep learning approach for detecting the position of the table tennis ball and analyzing its trajectory to classify in-game events such as table bounce, bounce, and net hits. The proposed methodology employs a super-resolution technique to enhance video frames, followed by a Scaled-YOLOv4 detector to precisely locate the ball, and then uses trajectory analysis based on 2D ball coordinates to identify events related to the table region. Evaluated on a custom dataset, the approach achieved high precision (97.8%) and F1-score (98.1%) for ball detection and similarly strong results for event classification (97.47% precision, 97.8% F1-score), while operating at real-time speeds (73.7 FPS for detection). Limitations include inability to classify serve events requiring 3D trajectory analysis and sensitivity to camera angle changes, with future work aimed at extending to 3D reconstruction and adapting the method to other racket sports.