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

Weight lifting is a risk factor of work-related low-back musculoskeletal disorders (MSD). From the ergonomics perspective, it is important to measure workers' body motion during a lifting task and estimate low-back joint moments to ensure the low-back biomechanical loadings are within the failure tolerance. With the recent development of advanced deep neural networks, an increasing number of computer vision algorithms have been presented to estimate 3D human poses through videos. In this study, we first performed a 3D pose estimation of lifting tasks using a single RGB camera and VideoPose3D, an open-source library with a fully convolutional model. Joint angle trajectories and L5/S1 joint moment were then calculated following a top-down inverse dynamic biomechanical model. To evaluate the accuracy of the computer-vision-based angular trajectories and L5/S1 joint moments, we conducted an experiment in which participants performed a variety of lifting tasks. The body motions of the participants were concurrently captured by an RGB camera and a laboratory-grade motion tracking system. The body joint angles and L5/S1 joint moments obtained from the camera were compared with those obtained from the motion tracking system. The results showed a strong correlation (r > 0.9, RMSE < 10°) between the two methods for shoulder flexion, trunk flexion, trunk rotation, and elbow flexion. The computer-vision-based method also yielded a good estimate for the total L5/S1 moment and the L5/S1 moment in the sagittal plane (r > 0.9, RMSE < 20 N·m). This study showed computer vision could facilitate safety practitioners to quickly identify the jobs with high MSD risks through field survey videos.