Application of a Multiple Object Tracking System to Professional Football Training Sessions

Abstract

The application of computer vision research in football has experienced significant growth in recent years, with a focus on detecting and tracking key elements during matches, such as the players and the ball, which provides spatio-temporal data. This data type is becoming invaluable for clubs as the demand for advanced data-centric solutions in sports increases. This dissertation aimed to develop a system that achieves detection and tracking performance on par with industry standards, such as GPS- systems, but in the context of football training sessions. A YOLOv8 model, pre-trained on the MS-COCO dataset, fine-tuned with data from professional matches and manually annotated data from professional training sessions, combined with BoT-SORT tracking, achieved high performance in player detection and tracking. Perspective transformation was then applied to create a top-down projection of the training area, enabling the mapping of player positions, player teams and motion vectors. This data is used to extract physical metrics such as total distance covered and the distribution of movement types by each player (walking, jogging, running, and sprinting). Additionally, an advanced metric, Pitch Control, is computed and visualized on the original video, offering tactical insights throughout the exercise. The proposed system demonstrates the feasibility of automated football analysis in training sessions, while also highlighting the need for further research to address current limitations in ball detection and real-time implementation.