Evaluation of a markerless motion capture to measure 3D joint kinematics during occupational lifting tasks using mobile devices

Recent advances in human pose estimation (HPE) have enabled markerless motion capture (MoCap) techniques as a promising alternative to traditional marker-based MoCap systems. However, most HPE algorithms only provide sparse video keypoints, which are insufficient to estimate joint angles in all anatomical planes according to biomechanical guidelines. OpenCap, an open-source smartphone-based markerless MoCap platform, addresses this limitation using a deep learning model (named the marker augmenter) that predicts dense anatomical markers from sparse video keypoints. However, it has shown lower performance for activities not included in its training dataset, such as occupational lifting tasks. In this study, we adapted the original marker augmentation model of OpenCap and proposed a task-specific model for occupational lifting, trained on a large and diverse dataset of manual lifting tasks. The proposed model reduced both kinematic errors (mean RMSE = 9.45° vs. 15.04°) and error variability (SD = 7.26° vs. 16.13°) compared to the original model. These findings suggest that OpenCap can be adapted for occupational lifting tasks, offering a low-cost, easy-to-use, and field-viable solution to collect 3D lifting kinematics for ergonomics applications.