Enhancing Industrial Environmental Safety Using Computer Vision

Abstract

Safety is crucial in industrial settings, ensuring both operational continuity and the well-being of workers. Safety in industries directly contributes to several Sustainable Development Goals (SDGs) outlined by the United Nations. Industries, particularly in sectors like steel production, face a spectrum of risk factors spanning forging, machining, and smelting processes, emphasizing the potential for economic, financial, and human losses. Traditional safety protocols, typically reliant on human oversight, often encounter constraints such as time inefficiencies, error susceptibility, and restricted visibility. To mitigate these challenges, one promising avenue involves implementing automatic monitoring systems leveraging computer vision technology. Accordingly, this paper introduces an approach aimed at detecting unsafe behaviors, including monitoring personal protective equipment (PPE) compliance and identifying hazardous areas through distance-based segmentation, specifically Customized for steel manufacturing facilities. To accomplish the objective of the paper, we introduce a benchmark dataset sourced from the ALASHRY steel factory. Through rigorous evaluation employing various computer vision models, including YOLOv8l, RT-DeTr, Faster R-CNN, and YOLOv8l for segmentation, we evaluate the dataset's efficacy. Our experimental findings reveal that YOLOv8l surpasses other models, achieving a mean Average Precision (mAP-50) score of 0.919. Moreover, our distance-based segmentation model, Customized for detecting hazardous regions, demonstrates YOLOv8l's effectiveness with an mAP-50 score of 0.857 on the provided dataset.