Vollytics: A Computer Vision System for Single-Camera Event Spotting and Rally Reconstruction in Beach Volleyball

Beach volleyball video analysis is difficult because contacts are brief, the ball is small and fast, occlusions occur, and broadcast footage contains cuts. This paper presents Vollytics, a modular single-camera framework for spotting volleyball actions, reconstructing rallies, and inferring match progression from broadcast and fixed-camera videos. It combines a You Only Look Once (YOLO)-based detector with court geometry estimation, geometry-aware filtering, lightweight player–ball association, and volleyball-specific temporal reasoning. Outputs use a unified frame-level timepoint representation for validation, correction, visualization, and export. We curate 23,651 annotated frames with player, ball, and action bounding boxes, plus an event dataset with 231 contact events and 44 rally intervals. Vollytics achieves a mean Average Precision of mAP50:95=0.576 over Intersection over Union (IoU) thresholds from 0.50 to 0.95, with F1-scores of 0.735 for ball detection and 0.951 for player detection. Event-level macro F1-scores are 0.851 with a ±2 s tolerance and 0.826 with ±0.5 s. At match level, Vollytics reconstructs 43 of 44 rallies, with F1-score 0.966 and mean temporal IoU 0.937. These results show that domain-specific reasoning converts noisy detections into coherent match-level information for scalable beach volleyball analysis.