Computer-vision-based tactical decision support system for pickleball doubles

Pickleball doubles, with its small court size, fast pace, and diverse tactical combinations, places extremely high demands on athletes' ability to make split-second decisions. Current training methods rely on subjective experience and offline replays, suffering from poor real-time performance and insufficient quantification. To address issues such as insufficient robustness against dynamic occlusion, missing tactical logic, and system latency, an end-to-end computer vision decision support framework has been developed: 1) A lightweight multi-object tracking pipeline has been designed, integrating dual-branch occlusion compensation and motion blur suppression techniques, improving tracking accuracy in the net-front conflict zone to 0.82; 2) Established a finite state machine-driven tactical rule engine, integrating spatio-temporal potential field models and expert knowledge bases, to achieve millisecond-level tactical generation. The system was validated through 120 hours of video data from the Asian Championships, achieving a multi-object tracking MOTA of 0.89 (with a latency of 105 ms), a tactical state recognition F1-score of 0.90, and a correlation coefficient of 0.93 between gap prediction and coach annotations. In practical applications, AR heatmaps improved athletes' tactical understanding efficiency by 38.7% and increased the success rate of third-shot net attacks by 22.7%. This framework marks the first real-time perception-decision feedback loop for all elements of pickleball doubles, providing a verifiable technical pathway for intelligent training in competitive sports.