This thesis presents a comprehensive system for the automatic analysis of football matches from a single-camera video recording. The system combines object detection using RF-DETR, player tracking using SAM2, ball tracking using a Kalman filter, team assignment based on SigLIP visual descriptors, pitch keypoint detection, and homographic projection of object positions onto a two-dimensional pitch model. Based on player positions and ball movement, the system recognises passes, interceptions, shots, and goalkeeper saves.
The system was evaluated on six manually annotated video clips comprising a total of 3,554 frames. The proposed approach achieved an IDF1 score of 0.96 for player tracking and preserved the correct identity in 95.8% of the matched tracks. For ball tracking, it achieved a tracking rate of 97.0% and a mean location error of 6.2 pixels. Team assignment accuracy reached 97.6%, while event recognition achieved an overall F1 score of 79.1%. The results demonstrate that modern computer vision methods can be integrated into a unified system for reconstructing the course of a football match and obtaining data for further sports analytics. The main limitations are the computational complexity of SAM2 and evaluation on a custom test set. Future work therefore includes improving processing speed and evaluating the system on publicly available benchmark datasets.
|
