Enhancing Tennis Player Tracking Accuracy Using a Vision-Based Framework with YOLOv8, Adaptive Kalman Filtering and Homography-Based Court Mapping

Abstract

Manual player monitoring in sports is error-prone, and existing automated systems are economically constrained by the need for high-speed cameras. This paper presents a cost-effective tennis player tracking system leveraging YOLOv8, which achieves 94.90% training accuracy and 97.87% validation accuracy. The system employs a vision-based framework combining a key point-based approach for court landmark detection, Homography transformation to map image coordinates to real-world court positions, and a Kalman filter for robust tracking during fast movements and occlusion. Together, these components enable quantitative analysis including trajectory visualization, distance coverage, speed estimation, and heatmap generation.