EAFvision: Real-Time Automated Safety Surveillance in Electric Arc Furnaces Using Deep Learning Models

Ensuring safety and operational efficiency in Electric Arc Furnace (EAF) steel manufacturing is critical due to the extreme hazards such as intense heat, toxic emissions, and heavy machinery present in these environments. We propose EAFvision, a real-time automated pipeline for safety surveillance EAFs, leveraging advanced deep learning architectures. EAFvision enables real-time detection of critical safety-related situations, including personnel, electrode clamps, and smoke emissions, to enhance situational awareness and operational safety in industrial environments. We collected and carefully annotated a comprehensive image dataset from an active EAF facility to benchmark a variety of models, including YOLO versions 8 through 11, RT-DETR, and established two-stage detectors like Faster R-CNN and Mask R-CNN. Our results demonstrate that lightweight, single-stage detectors deliver superior accuracy and faster inference times compared to more complex models, enabling efficient real-time testing on edge devices for immediate hazard detection and automated response. This approach highlights the transformative potential of AIpowered real-time monitoring systems to enhance workplace safety and optimize steel production processes.