Prosthetic Keypoint Estimation for Alternative Exercise Recommendation in Lower Limb Disabilities

초록

Traditional motion analysis for individuals with lower limb disabilities often relies on expensive equipment and controlled environments, limiting accessibility. Moreover, standard pose estimation datasets do not adequately represent prosthetic users, leading to suboptimal performance in prosthetic keypoint estimation. To address these challenges, this study proposes a deep learning framework for accurate prosthetic keypoint estimation using a monocular camera, enabling cost-effective motion analysis. This framework integrates a Vision Transformer backbone for enhanced feature extraction, a classification-based coordinate prediction head for precise keypoint localization, and a prosthetic-specific loss function that prioritizes keypoints through weighted optimization. In conjunction with this framework, a novel prosthetic image dataset, comprising real-world and synthetic images, is developed to improve learning of prosthetic-specific features. Experimental results indicate that models trained on this dataset outperform those trained solely on the MS COCO dataset, achieving a 16.4 AP improvement in prosthetic keypoint estimation. Furthermore, a prosthetic-specific training strategy elevates performance to 97.2 AP, significantly surpassing baseline models. By providing a specialized dataset and a targeted training approach, this framework facilitates accessible motion analysis for prosthetic users without requiring complex or costly equipment. Potential applications include rehabilitation, sports analysis, and clinical assessments, with the potential to advance prosthetic design and improve rehabilitation practices.

키워드