Benchmarking Pose Estimation and Two-Camera Angle Configurations for Markerless Musculoskeletal Multibody Dynamics Prediction

Abstract

Low-cost RGB video-based markerless motion capture is attractive for estimating lower-limb biomechanics, but the influence of implementation choices on the resulting kinematic and kinetic estimates remains uncertain. This proof-of-concept study developed a two-camera video-based musculoskeletal multibody dynamics workflow and assessed two pose estimation algorithm, RTMPose-x with the HALPE_26 keypoint convention and OpenPose with the BODY_25 convention, and three camera placement angles (20°, 40°, and 60°). Ten healthy male participants performed walking, running, and squatting while barefoot and wearing minimal clothing. In the video-based workflow, kinetic outputs were computed with ground reaction forces (GRFs) estimated using a foot–ground contact model and subsequently used to calculate joint contact forces (JCFs). A force-plate-informed Vicon-driven musculoskeletal model served as the reference. Relative to OpenPose-BODY_25, RTMPose-x–HALPE_26 produced 2.8%–26.4% lower root mean square errors (RMSEs) and 1.0%–46.3% higher Pearson correlation coefficients (ρ) for knee flexion–extension across the three activities, as well as 1.5%–38.2% lower RMSEs and 1.0%–16.4% higher correlations for most JCF outcomes. The 60° configuration showed favorable performance for selected biomechanical outcomes within the tested range. Overall, RTMPose-x–HALPE_26 showed lower numerical errors and/or stronger waveform correlations than OpenPose-BODY_25 for several outcomes, although these differences varied across activities and biomechanical variables.