A Deep Learning-Based Marker Augmentation Model for Markerless Motion Tracking During Manual Lifting

Computer vision is increasingly used to extract movement data from videos and perform ergonomic analyses. However, open-source computer vision algorithms often only detect major body keypoints, which are insufficient to calculate three-dimensional (3D) joint kinematics. This study developed a deep learning marker-augmentation model to predict the 3D positions of 33 anatomical markers from 20 video keypoints, along with participant height and weight, and lifting style. Data for model training and testing were collected from 21 adults performing lifting tasks recorded simultaneously using an optical motion capture system and smartphone cameras. Both video keypoints and predicted anatomical markers were used to drive inverse kinematic analysis. The mean kinematic error decreased from 18.5° to 9.5° when inverse kinematics was driven by augmented markers compared to using original video keypoints. These findings suggest that marker augmentation can improve video-based kinematic analysis and may support biomechanical evaluation of lifting tasks outside laboratory settings.