Evaluating YOLOv8-pose variants for soccer keypoint detection and homography mapping

Abstract

Soccer analytics intensely relies on accurate pitch keypoint detection and homography mapping, yet soccer broadcast data poses challenges caused by camera angle variation and partial pitch coverage. This study systematically evaluates you only look once version 8 (YOLOv8)-pose variants (nano, small, medium, large, and x-large) for keypoint detection and homography mapping. These models were retrained on broadcast and scouting datasets, while homography mapping was computed using direct linear transform (DLT) and random sample consensus (RANSAC). Three metrics were used for evaluation: detection accuracy, computational efficiency, and homography reliability. Results showed that the medium variant achieved the best balance, with the lowest mean pixel error (35.67 px), competitive inference speed 49.6 frames per second (FPS), and moderate model size (52.2 MB). Furthermore, it also reached the highest homography success rate (96.2%), surpassing all other variants. These findings suggest that the medium variant is the most effective balance for real-time soccer analytics, providing robust performance in accuracy, efficiency, and reliability under diverse broadcast data. Future studies would focus on enlarging the dataset with additional keypoints, incorporating temporal smoothing, and integrating a player tracking system. For reproducibility, the dataset and code are openly available on GitHub.