A Posture-Wise Comparative Analysis of Human Pose Estimation Models for Physical Rehabilitation Monitoring

Abstract

Introduction: Marker-less human pose estimation (HPE) is increasingly proposed for monitoring physical rehabilitation, yet most models are trained on upright, everyday poses, whereas physiotherapy involves lying, sitting and slow, occluded movements.

Objectives: To determine, posture by posture, how widely used 2D and 3D HPE models behave on rehabilitation exercises, how stable their accuracy is across postures, and which requirements a rehabilitation-specific estimator must meet.

Methods: A secondary quantitative analysis of the published UCO Physical Rehabilitation benchmark (27 subjects, OptiTrack ground truth) covering seven models: AlphaPose, MediaPipe, KAPAO, StridedTransformer, HybrIK, PoseBERT and VideoPose3D. Beyond the reported 2D keypoint error, 3D joint-angle mean absolute error (MAE) and throughput, we derive a Posture Sensitivity Index, rank stability, rotation gain for supine frames, an accuracy–throughput Pareto front and the gap to a ~5° clinical reference.

Results: No model leads in every posture. For 2D keypoints AlphaPose is best on supine frames (3.86 ×10⁻² couch-normalized units) and KAPAO on seated (3.24) and standing (3.34) frames; supine error is 14–103% higher than the best upright error for every model. For 3D joint angles HybrIK is best for supine (6.92°) and seated (7.29°) exercises and MediaPipe for standing (11.96°). The best 3D error exceeds the ~5° reference by 1.9°, 2.3° and 7.0° in the three postures, and the most accurate models are the slowest (13–16 FPS).

Conclusions: General-purpose HPE models are not yet sufficient for posture-agnostic rehabilitation monitoring. Estimators that reason jointly over all body joints, remain robust to non-upright orientation and run in real time are needed; joint-token transformer designs are a promising direction.