Kinart, Adrien
[UCL]
Van Oirbeek, Robin
[UCL]
Andrea Pennisi
[UCL]
With the advent of new technologies such as Artificial Intelligence (AI), the Internet of Things (IoT), and Cloud Computing, many industries are being revolutionized. These advances offer new low-cost solutions to complex problems. Despite this, amateur sports, such as field hockey, have seen limited research due to constraints such as budgets and a lack of public data for performance benchmarking. Throughout this study, the perspective of a field hockey coach wishing to make the most informed decisions related to in-game tactics and injury risk prevention is adopted. This study is two-fold. It first proposes a Multi-Player Tracking (MPT) framework that guides in selecting appropriate devices and techniques for insight extraction. To reflect the reality of the field, the following framework has been chosen: real-time and fully automatic insight extraction from a single stationary camera. Secondly, it paves the way for Computer Vision techniques by creating a first public annotated dataset of field hockey footage and applies current techniques before assessing their viability. Finally, YOLO detection algorithms were trained on the newly annotated dataset to classify between the goalkeepers, referees, and field players. This is possible in field hockey by exploiting the visual cues relative to the three classes. YOLOv5n, which can run almost in real-time on a CPU, was compared to YOLOv7-X and the newly released YOLOv8s that run in real-time on a GPU. Custom YOLOv5 rapidly converges, while the learning curves of Custom YOLOv7 and YOLOv8 still indicate a need for more training and data, although their performance is already satisfactory. As expected, YOLOv8 performance is better than YOLOv7 and should become the reference in MPT. Overall, this study shows that the visual cues of field hockey actors can be leveraged to turn the team classification task into a hard clustering task with two well-balanced groups, without having to deal with the noise introduced by the difference in appearance of the referee and goalkeeper.


Bibliographic reference |
Kinart, Adrien. Multi-Player Tracking in field hockey : Framework analysis and visual cues leverage for Referee, Goalkeeper and Field player detection. Faculté des sciences, Université catholique de Louvain, 2023. Prom. : Van Oirbeek, Robin ; Andrea Pennisi. |
Permanent URL |
http://hdl.handle.net/2078.1/thesis:40878 |