Steel Manufacturer Achieves Zero Accidents with Computer Vision-based Edge AI Safety System

Our team spent two weeks on-site, mapping camera positions, analyzing existing alarm systems, and understanding the exact workflow of crane operations. We identified critical blind spots in the current monitoring setup and documented the connectivity limitations in deep production zones.

We also worked closely with union representatives to address privacy concerns upfront – ensuring facial blurring would be non-negotiable from day one.

Generic PPE detection models weren’t accurate enough for the client’s needs. Their hard hats used specific color codes – yellow for floor workers, orange for supervisors, white for engineers – and the system needed to distinguish between them.

We collected site-specific training data, including challenging scenarios: poor lighting, steam obstructions, workers partially obscured by machinery. The resulting YOLO-based model achieved over 97% detection accuracy across all PPE types.

Each hazardous zone received a dedicated FPGA processing unit, eliminating network latency entirely. The detection-to-alert pipeline runs in under 50 milliseconds – 40 times faster than the cloud alternative the client had initially considered.

We integrated the system with the plant’s existing MQTT infrastructure and PLC controls, enabling automatic strobe activation and machinery slowdown without requiring operator intervention.

Industrial environments demand OT-grade security. We architected the system to meet IEC 62443 standards from the start – isolated network segments, encrypted edge-to-server communication, and strict access controls.

The system passed a rigorous third-party audit on the first attempt, giving the client full confidence to expand deployment across additional facilities.