3D kinematics of human falls: evaluating opencap and SAM3D body against inertial motion capture

Abstract

Abstract Detailed 3D kinematics provide crucial insights for evaluating fall injury mechanisms and impact severity; nevertheless, their application remains understudied. While traditional 3D motion capture systems (marker-based and inertial) are highly effective for laboratory research, their real-world applicability is limited. Consequently, real-world monitoring frequently relies on alternative systems that prioritize usability over kinematic accuracy. Recently, advanced AI algorithms have enabled 3D human pose estimation from standard RGB cameras, potentially extending the applicability of 3D fall kinematics analysis in real-world settings. However, their reliability in highly dynamic, occlusion-prone fall scenarios has yet to be established. This study validates two RGB vision-based motion capture frameworks, the multi-camera OpenCap (OC) and the monocular SAM3D Body (3DB), against a reference OpenSim musculoskeletal pipeline driven by an inertial system (i.e., Xsens) during simulated falls. Nine healthy participants performed forward, backward, and lateral falls, and their kinematic trajectories were compared using the Mean Absolute Error (MAE) across 24 rotational and 3 translational degrees of freedom. Results indicate highly comparable tracking capabilities, both yielding a global rotational MAE of 13°. Lower-body kinematics demonstrated moderate errors (<10°), whereas upper-body tracking exhibited higher discrepancies (>15°). Regarding pelvic translation, OC consistently outperformed 3DB (6.1 cm vs. 8.5 cm). Despite non-negligible error margins, these findings confirm that these vision-based algorithms hold great potential in 3D fall kinematics research. Remarkably, the monocular 3DB system performed comparably to the multi-camera OC framework, highlighting the feasibility of further streamlining RGB markerless motion capture through single-camera setups.