FusionPose: a MediaPipe and keypoint R-CNN fused model for enhanced cyclist pose estimation and injury risk prediction

Cyclist posture has a major impact on both injury risk and performance efficiency. This paper suggests a markerless pose estimation method based on deep learning to estimate injury risk by examining cyclist posture. Conventional marker-based systems, although precise, are expensive and not feasible for real-world use. To validate the markerless approach, knee angle measurements were compared with those from a marker-based system, revealing minimal deviation. Keypoints were detected from cyclist images with four models—MediaPipe, YOLOPose, MoveNet, and Keypoint-RCNN—and tested against ground truth data with metrics such as Euclidean Distance, PCK, and RMSE. The two best models were merged to improve accuracy, and the ultimate FusionPose model predicted joint angles to evaluate injury risk with an accuracy of 97.68%. The findings show the efficacy of deep learning models in providing a precise and cost-effective solution to cyclist posture assessment and injury mitigation.