Vision Transformer-based Pose Estimation for Automated Gait Analysis in Ankle-Foot Prosthetic Design

Gait abnormalites arising from neurological illnesses can have severe impact on the quality of life. This research bridges the literature gap by conducting a rigorous comparative analysis of YOLO v8, DeepPose, and the novel vision transformer-based RTM model for automated gait analysis and diagnosis in the context of ankle-foot prosthetic design. Using the 3DPW dataset with existing 2D pose annotations, quantitative performance evaluations demonstrated RTM Pose's superiority, with the lowest mean errors of 19.75 (MAE), 732.62(MSE), and 15.61 (RMSE). The model adequacy (R 2 ) for Yolov8, Deep Pose, and RTM Pose was calculated as 99%, 99.3%, and 99.5%, respectively. The novel contribution lies with proving whether vision transformer models like RTM Pose can precisely extract intricate body movements, enabling automated diagnosis of gait abnormalities and personalized rehabilitation programs. Such capabilities are invaluable for improving the treatment outcomes in stroke, traumatic brain injury, Parkinson's, and other neurological conditions. This research also provides the maiden comparative analysis of multiple pose estimation models (YOLO v8, DeepPose, and RTM) on a common annotated video dataset, offering new performance insights. Overall, the vision transformer-based RTM model achieved the highest accuracy for gait analysis while maintaining efficient computation, evidenced by the lowest 107.7ms mean inference time. The potential is immense starting from tracking subtle motion patterns to analyzing complex body mechanics. This methodology paves the way for AI-based tools that objectively evaluate gait impairments and provide personalized corrective interventions to enhance neurological rehabilitation and inform ankle-foot prosthetic design.