Vision-Based Ergonomic Risk Estimation: Deep-Learning Strategies.

Ergonomic risk assessment is traditionally human-assisted and performed filling forms or through on-site inspection by ergonomists. This frequently leads to inaccuracies due to the subjective bias. Also, it is inefficient and costly given the time and technical knowledge required. Computer-based alternatives are slowly emerging, but so far there is few consensus or uniformity between the underlying practices and technologies. A standardization of data collection in video takes through computer vision offers the opportunity to obtain considerable replication levels, which would increase the reliability of the results and the quality of the data available to ergonomists. In this work we propose a workflow that employs two open-source neural networks: STAF for workers’ body joints detection and tracking, and VIBE for 3D movement estimation. Finally, the ergonomic risk is calculated based on REBA, which is one of the most widespread standard in industrial settings. As data collection may be a bottleneck (as usual in deep learning) we propose the use of virtual scenarios generated in Unity3D. This allows to evaluate and quantify several problems associated with actual video takes, including self-occlusions, camera positioning, illumination, noisy backgrounds, and many others. The results are positively conclusive about both the use of this workflow for actual risk assessment, and the feasibility of virtual environments for controlled experimentation.