CourtVision: Next-Generation AI Framework for Real-Time Basketball Dynamics

This work introduces CourtVision, a real-time computer vision framework for basketball analytics. The system integrates YOLOv12 for multi-class detection of players, referees, and the ball, a Kalman filter for robust identity tracking and trajectory prediction, and YOLOv8-based pose estimation to extract fine-grained player movement statistics. A custom dataset of 415 annotated basketball images was used for development and evaluation. The framework achieved a mean Average Precision of 0.93, precision of 0.92, and recall of 0.90, with an average inference speed of 32 frames per second, confirming its suitability for live deployment. Beyond detection, CourtVision automatically generates player-level metrics such as speed, distance covered, and ball possession, while providing trajectory predictions under occlusion. These results highlight its potential to enhance tactical evaluation in basketball and extend to broader applications in sports analytics.