Visual-based ergonomics analysis using artificial intelligence

Abstract

Workplace ergonomics issues, especially those driving musculoskeletal disorders, have drawn the attention of both workers and managers due to the resulting harm to health and loss of revenue. Thus, to evaluate undesirable posture and alert individuals to the risk, ergonomists have developed manual worksheets to work out a total risk score. However, these worksheets are labour-intensive and relatively biased, instead, AI(Artificial Intelligence)-based techniques such as computer vision have been developed.

The presented study have introduced two pipelines to address human pose assessment in video datasets, a transfer learning model and a multi-task learning model. The core of the transfer learning method consists of 3D joint estimation implemented on pre-trained Video Pose 3D architecture or VIBE model, and offline REBA (Rapid Entire Body Assessment) risk prediction which was programmed based on off-the-shelf mapping guideline. The performance of this paradigm shows the capability to generalized applications for real-world problems, and failure cases concerning particular body parts and jitters caused by different viewpoints are analysed. In the multi-task learning model, ergonomics risk analysis is considered a scene-dependent task where human pose assessment and human action segmentation are combined. This model uses modified targeted skeletal information and an ergonomics risk labelling approach. Consequently, the estimation accuracy outperformed state-of-the-art research, and occlusion matter corresponding to specific actions has been identified. The reliable and insightful analysis conducted in this study could be helpful for future visual-based human posture evaluation.