Learning-based 3D human kinematics estimation using behavioral constraints from activity classification
Daekyum Kim, Yichu Jin, Haedo Cho, Truman Jones, Yu Zhou, Ameneh Fadaie, Dmitry Popov, Krithika Swaminathan, Conor J. Walsh
Abstract
Inertial measurement units offer a cost-effective, portable alternative to lab-based motion capture systems. However, measuring joint angles and movement trajectories with inertial measurement units is challenging due to signal drift errors caused by biases and noise, which are amplified by numerical integration. Existing approaches use anatomical constraints to reduce drift but require body parameter measurements. Learning-based approaches show promise but often lack accuracy for broad applications (e.g., strength training). Here, we introduce the Activity-in-the-loop Kinematics Estimator, an end-to-end machine learning model incorporating human behavioral constraints for enhanced kinematics estimation using two inertial measurement units. It integrates activity classification with kinematics estimation, leveraging limited movement patterns during specific activities. In dynamic scenarios, our approach achieved trajectory and shoulder joint angle errors under 0.021 m and $$6.5^\circ$$ , respectively, 52% and 17% lower than errors without including activity classification. These results highlight accurate motion tracking with minimal inertial measurement units and domain-specific context. Inertial measurement units offer a cost-effective, portable alternative to lab-based systems for measuring human motion. Here, the authors predict human motion using inertial measurement units combined with machine learning, leveraging limited movement patterns and reduced motion variability during specific activities.