PoseRx: A Transformer-Based Remedy for Precision in Physical Rehabilitation Monitoring

Abstract

Accurate and reliable human pose estimation plays a vital role in physical rehabilitation, where therapists depend on precise joint tracking to evaluate posture, monitor range of motion, and assess patient progress. While traditional pose estimation models such as AlphaPose, MediaPipe, and HybrIK have shown varying levels of performance, they often struggle in scenarios involving occlusions and diverse body positions commonly encountered in rehabilitation settings. Their limitations—particularly in terms of temporal consistency, occlusion robustness, and joint angle accuracy—undermine their clinical applicability. To address these challenges, this study introduces PoseRx, a transformer-driven pose estimation framework built on the TokenPose architecture, specifically tailored for physical rehabilitation monitoring. PoseRx processes RGB video inputs and employs a Vision Transformer-based joint attention mechanism to estimate 2D keypoints, which are subsequently lifted to 3D using temporal models such as VideoPose3D. The framework is evaluated across rehabilitation-specific postures, including supine, seated, and standing positions, and benchmarked against state-of-the-art methods using metrics such as 2D localization error, joint angle mean absolute error (MAE), model complexity, and occlusion handling capability. Results demonstrate that PoseRx achieves superior performance, with a 2D localization error as low as 5.9 pixels and a joint angle MAE of 5.4°, outperforming existing models across all evaluated positions. Moreover, it exhibits the highest resilience to occlusion and provides enhanced support for custom joints, both of which are essential in real-world rehabilitation scenarios. PoseRx delivers a robust, efficient, and clinically relevant solution for human pose tracking in rehabilitation environments. Its transformer-based design and modular architecture make it a promising next-generation tool for improving physiotherapy feedback, tracking patient progress, and advancing digital health interventions..