3D ergonomics parameter measurement using video-based deep learning

In contrast to strapping sensors or scanning with a 3D scanner, obtaining ergonomic parameters by taking a video of a single person is undoubtedly one of the simplest, most efficient, and most comfortable measurement methods. In this paper, we propose an ergonomic parameter measuring method based on extracting the keypoints of the human body in the video using deep learning techniques,aiming to achieve high accuracy under simple hardware conditions. A YOLOv8-FSC-Pose network model is developed for human target detection and joint point localization in binocular video. Kalman filtering compensates for joint point outliers caused by occlusion or noise in each frame. Then the optimized DeepLabv3+ semantic segmentation network is used to extract a complete and fine human motion contour from the target region. We extract the head and foot points from the contour required to measure human parameters. Based on ergonomic principles, we designed specific calculation methods to determine walking height, upper and lower limb range of motion (ROM), and gait parameters by analyzing the three-dimensional coordinates extracted from human keypoints. The error analysis results show that this method can fully apply to the real-time measurement of human parameters in complex outdoor scenes, with the advantages of good real-time, ease of operation, and high accuracy.