A Review of YOLO Models for Soccer-Based Object Detection

The application of AI-driven computer vision techniques to sports footage for the purpose of automatic insight generation is a growing area of research and development. The ability to detect players of various teams along with other entities is a foundational component of such activities. This paper describes the machine learning-based detection of soccer players, balls, goalkeepers, referees, assistant referees, and other attendees from multiple teams which form the initial steps for soccer analytics. This study performs an empirical analysis of various versions of the computer vision model YOLO (You Only Look Once) using the open source soccer dataset named SoccerNet. The images contained within are trained on YOLO V3, V5, V8 and V9 with epoch configurations 1, 5, 10, 25, and 50 to facilitate a broad review of the capabilities of YOLO models. The results are evaluated using mAP50 and mAP50-95 (Mean Average Precision) metrics. YOLOv9-(9c, 9e) weights exhibited a similar level of performance to those of YOLOv8(8l, 8x), which both outperformed versions V3 and V5. However, YOLOv9 resulted in a higher true positive rate than YOLOv8, which suggests that a multi-metric analysis is pertinent for model ranking. The results of each model variant utilised are reported and visualized for reproducibility and to form a benchmark supporting future studies. The implementation was performed using the Python-based Ultralytics library.