A Vision-Based marker-less measurement system for Real-Time Estimation of Human Body Inertial properties

Accurate real-time estimation of the inertial properties of the human body is essential in biomechanics, human–robot interaction, and immersive environments. Traditional methods based on wearable markers or sensors suffer from drawbacks such as invasiveness and signal drift. This work introduces a vision-based, marker-less system for real-time estimation of the human body's inertial parameters, specifically the center of mass and inertia tensor. A multi-camera 2D setup is employed to reconstruct the 3D human posture using 13 key-points extracted via 2D human pose estimation model. A simplified geometric model discretizes the body into basic volumetric shapes, allowing analytical computation of segment-level inertial properties, while ensuring real-time compatibility. The pipeline runs on a desktop system (Intel Core i9-13900HX and NVIDIA RTX 4090). Evaluation on the Human3.6M dataset shows a mean processing time of 15.5 ms per frame. Experimental validation was performed on 15 healthy participants using a load cell-equipped platform during five predefined motor tasks. The system estimated the torque generated by the subject around the fixed axis of rotation of the instrumented platform, corresponding to the hinge axis of the force plate, with a median mean absolute error of 5.88 Nm and a median Pearson correlation coefficient of 0.98. A Monte Carlo analysis propagated measurement uncertainty from keypoint localization and body mass through the model. The resulting standard deviations were below 6.12 mm for center of mass position and below 0.095 kg m2 for inertia tensor components resulting consistent across tasks and subjects. Overall, the proposed method provides physically-coherent human body inertial properties estimation without markers, wearables, or subject-specific calibration. Its non-invasiveness, computational efficiency, and adaptability make it suitable for real-time motion analysis in clinical, ergonomic, and robotic contexts.