Deep learning-based single-camera video pose estimation algorithms have shown promise for enabling large-scale movement studies in real-world environments, free from the limitations associated with marker-based and multi-camera markerless motion capture systems. However, there is limited research regarding the accuracy and applicability of human pose estimation in out-of-laboratory conditions. In single-camera systems, changes in camera projections can reduce joint visibility, increase occlusion, and lead to issues in observing multi-axial movements, which can affect the accuracy of calculating joint-angle estimates. This research aims to assess the performance and sensitivity of 2D (OpenPose) and 3D (GVHMR) reconstructions in estimating knee joint angle across eight camera angles compared to traditional 3D marker-based motion capture. 2D pose estimation (OpenPose) shows high variability in joint-angle calculations between camera projections, as oblique views, particularly rear view (45°, 315°), yielded the most accurate knee joint-angle estimates. For this algorithm, careful camera placement should be a consideration when collecting data. The addition of a 3D mesh model (GVHMR) helps generalize estimates across orientation, leading to more consistent results across camera placements. However, joint centers in these mesh models are not anatomically defined, introducing bias in the results that requires calibration. Single-camera markerless motion capture shows strong potential for clinical and sport settings, where rapid movement assessment is often preferred. However, important gaps remain in understanding how certain factors, such as configuration and occlusion, can affect joint angle-estimation accuracy. By quantifying how joint-angle outputs vary with camera angle across systems, this study offers practical guidance for system setup and clarifies the limitations and capabilities of single-view markerless motion capture for kinematic assessment.
Ishi Arora (Thu,) studied this question.