Ergonomic risk assessment based on computer vision and machine learning

Abstract

We develop a novel method that performs accurate ergonomic risk assessment, automatically computing Rapid Upper Limb Assessment (RULA) scores from snapshots or digital video using computer vision and machine learning techniques. Our method overcomes the limitations in recent developments based on computer vision or in wearable measurement sensors, being able to perform unsupervised assessment handling multiple workers simultaneously, even under sub-optimal viewing conditions (e.g., poor illumination, occlusions, and unstable camera views). The processing workflow uses open-source neural networks to detect the workers’ skeletons, after which their body-joint positions and angles are inferred, with which RULA scores are computed. The method was tested with computer-generated, controlled real-world image datasets, and with freely available videos taken in outdoor working scenarios. The computed RULA scores were in close agreement with the assessments of seven specialists in the field, achieving a Cohen´s κ over 0.6 in most real-world experiments.