Comparative Analysis of YOLOv5/v8/v9 for Object Detection, Tracking, and Human Action Recognition in Combat Sports

YOLO models are widely used object detectors in computer vision (CV). This study investigates the relative performance of YOLOv5, YOLOv8, and YOLOv9 for object detection, tracking, and human action recognition in combat sports. The models were evaluated using curated datasets encompassing various combat scenarios, athlete movements, and equipment configurations. Pre-processing protocols and augmentation techniques were applied to improve model accuracy and generalizability, including automated orientation correction, image dimension standardisation, contrast enhancement, and methods such as zoom, rotation, shear, and grayscale conversion. The key findings provide insight into the comparative performance of the models across various evaluation metrics, such as precision, recall, and mean average precision. Each model's ability to detect, track, and recognise human actions in dynamic combat sports environments is evaluated. Computational efficiency and real-time performance were assessed as these are important indicators for practical applications in coaching, training, and competitive scoring systems. The findings suggest that YOLOv8 offers the best balance of precision and recall, making it particularly suitable for real-time applications in combat sports analytics. This study contributes to advancing CV technologies in combat sports analytics, with potential implications for improving athletic training methods, facilitating personalised coaching interventions, and enhancing objectivity and consistency in competitive scoring processes in combat sports.