2020-2021 PPRT Project

A Deep-Learning-Based Computer Vision System for Body Asymmetry Angle Estimation

Research Trainee: Zhengyang Lou, PhD Student, Electrical and Computer Engineering, University of Wisconsin-Madison

Principal Investigator: Robert Radwin, PhD, Professor, Biomedical Engineering, Industrial & Systems Engineering, and Orthopedics & Rehabilitation, University of Wisconsin-Madison

Zhengyang Lou, PhD student

Robert Radwin, PhD

Overexertion during manual lifting is one of the costliest and prevalent occupational injuries and among the most challenging to analyze in today’s dynamic workplace. The revised NIOSH lifting equation (RNLE) developed by Waters, et al. (1993), provides the recommended weight limit based on lifting frequency, location of the object, and position of the subject. Conventional methods for applying this equation typically involve manual measurements, which is time-consuming and impractical, particularly for multiple lifting tasks. This is further complicated by the fact that manual materials handling tasks in manufacturing and growth

industries such as warehousing, distribution centers, and shipping, involve numerous lifting tasks that are continuously changing. Thus, it is becoming increasingly important to develop a system that can monitor worker physical stress exposures continuously using techniques such as computer-vision-based methods. Previously our group demonstrated the feasibility of extracting RNLE parameters including lifting height, distance, frequency, and lifting instant using a single video camera. However, our previous approach is only capable of accurately estimating the asymmetry angle, an important factor of the RNLE, using two cameras and traditional computer-vision-based methods viewing the sagittal plane of the subject. Recent advancements of deep learning based single view human pose estimation algorithms has made it possible to estimate 3D skeletal joint coordinates from a monocular RGB camera, making it promising to measure body asymmetry angle directly. Thus, this proposal will extend our previous work by investigating a body asymmetry angle estimation algorithm based on single-view 3D human pose estimation algorithms. This pilot project will create a dataset of 40 subjects performing lifting tasks with different twisting angles in a range of 135° using cameras shooting at different viewing angles. A systematic analysis will be carried out using this dataset to provide recommendations for the optimal placement of cameras for estimating body asymmetry angle.