Learning Biomechanically Plausible Human Motion from Sparse Radar Point Clouds

Traditional radar-based human pose estimation (HPE) typically represents the body as a set of unconstrained 3D keypoints. This approach often results in biomechanically implausible movements, such as fluctuating bone lengths or feet that penetrate the floor. This paper addresses these limitations by integrating a full-body skeletal model into a differentiable, end-to-end trainable framework, aiming to recover clinically interpretable biomechanical descriptors from low-cost radar sensors.

The authors propose a pipeline that maps temporal sequences of radar point clouds to generalized coordinates (joint angles) rather than raw Cartesian coordinates. Key components include:

• Biomechanical Modeling: A kinematic tree of rigid segments with anatomically defined rotation axes, which ensures consistent limb proportions by construction.
• Subject-Specific Scaling: A multi-task Lasso regression model that predicts individual body segment proportions directly from radar features to scale the skeleton.
• Differentiable Forward Kinematics: A module that maps predicted joint angles to 3D positions, ensuring the output adheres to the skeletal structure.
• Contact Classification: A loss function that encourages physically plausible foot-ground interaction, allowing the model to recognize when feet are in contact with the floor.

The framework was evaluated using leave-one-subject-out cross-validation on 11 participants performing various rehabilitation exercises.

The proposed method achieved a mean per-joint position error (MPJPE) of 6.456 cm and a mean per-joint angle error (MPJAE) of 8.083 degrees, outperforming the unconstrained mmMesh baseline. By construction, the model maintained rigid bone lengths, significantly reducing the variability seen in unconstrained models. Furthermore, the framework successfully classified foot-ground contact with an F1 score of 0.935, demonstrating that a single low-cost radar sensor can provide reliable, interpretable biomechanical data without the privacy concerns associated with cameras.