Automated player tracking in 3x3 basketball using a deep learning framework: Comparison of YOLO architectures and limitations in identity maintenance

INTRODUCTION: Deep learning, a subset of machine learning based on artificial neural networks, has become central to computer vision. YOLO (You Only Look Once), a family of object detection architectures, processes an image in a single pass and has evolved through successive generations with gains in accuracy and efficiency, being increasingly applied to player tracking in team sports. However, 3x3 basketball poses unique tracking challenges, such as a smaller court and identical uniforms, and dedicated methods for this modality remain scarce. OBJECTIVE: To compare different YOLO architectures for player detection in 3x3 basketball and to evaluate the performance of the selected model in both detection and tracking tasks. METHODS: Local ethical approval was obtained. The sample comprised 48 professional 3x3 basketball players from nine matches, whose frames were manually annotated for player detection. Seven YOLO generations were trained and compared using standard detection metrics, and the best-performing model was combined with the BoT-SORT tracking algorithm. Tracking performance was assessed by comparing generated player identities against the actual number of players per match. RESULTS: YOLOv8m achieved the best overall detection performance among the seven architectures compared and was selected for the tracking pipeline. Detection performance was satisfactory, with precision and recall above 96% in eight of the nine matches. Tracking, however, was markedly unstable, with generated identities exceeding the actual number of players by 4 to 9 new IDs per player per minute, possibly related to occlusions, similar uniforms, and non-optimized hyperparameters. CONCLUSION: YOLOv8m provided the best detection performance among the compared architectures and, combined with BoT-SORT, reliably detected players but failed to maintain stable identities. Since tracking was assessed only through identity counts, without formal metrics such as IDF1 or HOTA, these findings are preliminary. Future work should apply standardized tracking metrics and systematic hyperparameter tuning to improve identity maintenance.