Automating Ergonomic Evaluation in Harsh Environments: A Real-Time Computer Vision Framework for Underground Mining

Abstract

Musculoskeletal disorders (MSDs) represent a critical problem in underground mining due to biomechanical demands and adverse environmental conditions. Traditional observational methods, such as RULA, are subjective and do not allow for continuous monitoring. This article proposes a real-time machine vision framework for automated ergonomic assessment in harsh mining environments. The methodology comprises: (i) a video acquisition protocol adapted to dusty conditions, poor lighting, and PPE occlusions; (ii) frame preprocessing and filtering using MediaPipe Pose; (iii) construction of a dataset labeled with RULA risk levels (Low, Medium, High) based on joint angles; and (iv) a custom convolutional neural network (CNN) for postural classification. The model was trained and validated with data from real mining operations, employing cross-validation and metrics for accuracy, completeness, and F1 score. The results demonstrate the system's viability in providing objective, continuous, and real-time assessments, overcoming the limitations of traditional methods and offering a scalable tool for the proactive prevention of musculoskeletal disorders in underground mining