Benchmarking Markerless Vision-Based Human Pose Estimation with an Upper-Limb Rehabilitation Exoskeleton

Reliable Human Pose Estimation (HPE) is essential for vision-based monitoring of upper-limb movements in exoskeleton-assisted rehabilitation. However, the robotic structure, camera placement, and arm orientation may challenge pose-estimation performance, and widely used markerless models remain insufficiently benchmarked under these conditions. This paper benchmarks 10 representative HPE models on video sequences of a healthy participant wearing an upper-limb exoskeleton (idle, with the collaborative manipulator locked), recorded with an OAK-D camera from frontal, top-down, and lateral placements, covering elbow flexion and extension under shoulder flexion and shoulder abduction. Performance is assessed across accuracy, robustness, temporal stability, and speed. RTMO-M achieved the highest accuracy (Percentage of Correct Keypoints (PCK) = 0.921 ± 0.081), while YOLO26m offered the best accuracy-throughput trade-off (PCK = 0.896 ± 0.125, 27.0 ± 3.5 frames per second). Camera placement influenced accuracy more than model selection: frontal placement achieved a mean PCK of 0.947 versus 0.719 for lateral, a gap exceeding the largest inter-model difference. An ablation across five YOLO size variants and three recording resolutions further characterizes accuracy–speed trade-offs and resolution sensitivity. Results highlight the importance of model and camera configuration for vision-based HPE with upper-limb exoskeletons in laboratory scenarios representing rehabilitation conditions.