Authors:
Abdelrahman Mostafa
1
;
Muhammad Rushdi
1
;
2
;
Tamer Basha
1
and
Khaled Sayed
3
Affiliations:
1
Department of Systems & Biomedical Engineering, Faculty of Engineering, Cairo University, Giza, Egypt
;
2
Department of Computer Science, New Giza University, Giza, Egypt
;
3
Department of Electrical & Computer Engineering and Computer Science, University of New Haven, West Haven, CT, U.S.A.
Keyword(s):
Multiple Object Tracking, YOLO, Soccer Player Tracking, Football Analytics.
Abstract:
Data analytics have had a significant impact on tactical and workload planning in football. Football data is divided into two categories: event data, which captures on-the-ball events like passes and shots, and tracking data, which captures off-the-ball movements. However, traditional methods of collecting tracking data are expensive and inconvenient. Recently, AI solutions have emerged as low-cost and user-friendly alternatives to track players’ movements from video streams. This paper introduces FOOTBALLTrace, an end-to-end AI system for tracking multiple football players from a panoramic game view. The system also incorporates a novel algorithm that detects potential occlusion events and ensures trajectory continuity for occluded players. The workflow involves five stages: panoramic view creation, player detection, player ID association, occlusion detection, and trajectory correction. The system utilizes YOLOv7 for multiple object detection and employs a pre-trained deep affinity
network to assign unique IDs to players throughout the game. Occlusion detection and trajectory correction are achieved by extracting geometric features from discontinuous player trajectories. The system’s performance was evaluated on full-length video data of a football game, with occlusion events manually extracted for training and testing the occlusion detection and trajectory correction algorithm. The system achieved an 87.5% trajectory correction rate for occluded trajectories.
(More)