A computer-vision method to estimate joint angles and L5/S1 moments during lifting tasks through a single camera

Area of Science:

• Biomechanics
• Ergonomics
• Computer Vision

Background:

• Weight lifting is a significant risk factor for work-related low-back musculoskeletal disorders (MSD).
• Accurate measurement of worker body motion and estimation of low-back joint moments are crucial for assessing biomechanical loading and preventing injuries.
• Advancements in deep neural networks and computer vision offer new tools for human pose estimation from video data.

Purpose of the Study:

• To evaluate the accuracy of a computer-vision-based system for estimating 3D human poses during lifting tasks.
• To calculate joint angle trajectories and L5/S1 joint moments using a computer-vision approach.
• To compare the accuracy of computer-vision-derived biomechanical data against a laboratory-grade motion tracking system.

Main Methods:

• Utilized VideoPose3D, an open-source library, for 3D pose estimation from single RGB camera videos of lifting tasks.
• Employed a top-down inverse dynamic biomechanical model to compute joint angle trajectories and L5/S1 joint moments.
• Conducted an experiment with participants performing various lifting tasks, capturing motion with both an RGB camera and a motion tracking system for validation.

Main Results:

• Demonstrated strong correlations (r > 0.9) between computer-vision and motion tracking for shoulder flexion, trunk flexion, trunk rotation, and elbow flexion (RMSE < 10°).
• Achieved good estimates for total L5/S1 moment and sagittal plane L5/S1 moment using the computer-vision method (r > 0.9, RMSE < 20 N·m).

Conclusions:

• Computer vision, specifically using VideoPose3D, provides a reliable and accurate method for analyzing biomechanics during lifting tasks.
• This technology can empower safety practitioners to efficiently identify high-risk jobs through field survey videos, thereby mitigating musculoskeletal disorder risks.