Evaluation and Error Characterization of a Multi-View 3D Markerless Motion Capture System for Human Kinematics

Markerless motion capture using RGB cameras offers an accessible alternative to marker-based systems, but its measurement properties require further evaluation. We evaluated a three-camera MediaPipe Pose system against a Vicon marker-based reference during deep squat and in-line lunge exercises in six healthy adults. Analyses examined calibration, landmark visibility, camera configuration, systematic joint position offsets, and residual variability. Calibration root mean square error (RMSE) was below 3 mm and small relative to joint position disagreement. For the tested arrangement, three cameras increased data availability and reduced errors relative to camera pairs when compared at time points with sufficient landmark visibility in all three views. The mean trial-level median joint position error was 45.4 mm, mean segment orientation error was 9.6°, and mean overall angle error was 6.4°. Subject-specific offset correction in marker-based segment frames reduced these values to 28.6 mm, 7.4°, and 4.6°, respectively, in the same trials used to estimate the offsets. Cross-validation yielded similar improvements within subjects and smaller improvements across subjects, while still requiring marker-based segment frames. Residual variability showed subject-dependent directional patterns within this cohort. These findings characterize agreement with the reference and underscore the need to consider visibility, landmark definitions, and participant variability when interpreting performance.