Towards Biomechanical Analysis in Workplace Ergonomics Using Marker-Less Motion Capture: 3D Human Pose Estimation for Lifting/Lowering Tasks

Computer vision-based human pose estimation has a high potential for applications in the prevention of musculoskeletal disorders among workers. However, there is a lack of accurate datasets on the motion of workers that can be utilized for human pose estimation in the industrial environment. In this work, we collected a 3D annotated multi-view high-accuracy image dataset for human pose estimation while lifting/lowering a cardboard box, with all the participants’ faces blurred to protect their privacy. Furthermore, a previously validated deep learning model was trained on 6 participants and tested on 6 additional participants. Additionally, the effect of the size of cardboard box used for lifting task in the test set was subsequently examined. The average MPJPE (Mean Per Joint Position Error) of the neural network for human pose estimation was 11.97 ± 2.10 mm while the larger cardboard box yielded an increase of MPJPE to 22.56 ± 12.02 mm. Our findings suggest that the joint estimation could be highly accurate when environment is similar between training and testing. Increasing training dataset by various environments and scenarios would be useful to increase robustness of the model.