Rally Scene Detection Using Time of Score Change in a Table Tennis Video for Broadcast

This study presents an automatic method for rally scene detection and score estimation in a broadcast table tennis video. Pre-match analysis typically requires extensive manual recording of player actions, shot trajectories, and scores from videos. This time-intensive process impacts preparation efficiency. The proposed methodology integrates YOLOv8 for player and table detection with Segment Anything Model (SAM) for scoreboard area segmentation. Rally scenes are identified through pixel change rate analysis combined with Mean Square Error calculations for frame similarity assessment. The research introduces a score recognition technique functioning independently of scoreboard design variations. Experimental evaluation demonstrates a 9 percentage point increase in precision compared to existing methods, with reduced false positives and missed detections. By incorporating score change timing data, the system overcomes challenges presented by variable camera angles and scene transitions. This work contributes to automated sports video analysis by demonstrating the feasibility of detection systems that enhance pre-match preparation while reducing manual analytical workload in professional table tennis contexts.