Comparison of computational pose estimation models for joint angles with 3D motion capture

Tools to calculate human movement patterns can benefit musculoskeletal clinicians and researchers for rehabilitation assessments. The research objective of this study was to compare two human pose estimation models (HRNet, MediaPipe) against the laboratory marker-based reference standard for joint angles and range of motion (ROM) for several movement parameters. Twenty-two healthy volunteers (Female n = 16, Male n = 6), participated to compare outputs for knee and elbow kinematics. Joint angles were calculated by selecting three marker points defining the joint and angle between them in Qualisys Track Manager software. Using predicted key points, pose estimation model calculations for the same musculoskeletal kinematic outputs were computed. Coefficient of Variation (CoV) was used as a variation statistic for joint angle during movements. All comparison results were under 10%, implying that both models compute reliable joint angle data during the five tested activities. When comparing ROM as a discrete parameter, CoV values remain low, though not all below 10%. Intra-class Correlation Coefficients were computed across the ROM data as a measure of statistical similarity. Each exercise displayed good-excellent and significant correlations for both models compared to Qualisys apart from left knee sit-to-stand. Exploration from this data sampling imply that flexion/extension exercises give stronger consistency results than full sit-to-stand movements when compared to 3D motion analysis, and there is little distinction between these two models. Finer tuning of models will give further reliability for in-depth analysis as these results are restricted, but valuable for a rehabilitative setting with limited objective analysis alternative.