Mining Eyes: Real-Time Safety Monitoring Across 15+ Mining Sites

Context

Berau Coal operates open-pit sites where heavy dump trucks, light vehicles and people share the same haul roads. Safety supervision relied on manual CCTV monitoring and after-the-fact incident reports. By the time a violation was reviewed, the risk had already passed.

I joined the Mining Eyes Analytics project as Lead ML Automation Engineer and stayed on it for close to three years, from the first proof of concept to a platform that ran across 15+ sites.

Problem

Three things made this harder than a standard detection project:

1. Nobody watches 24 camera feeds. Operators had the screens, not the attention. Detection had to be automated and the alert had to arrive fast enough to act on.

2. Mining footage is hostile to models. Dust, backlight at sunrise, night shifts under floodlights, and objects that appear as a handful of pixels at the far end of a haul road.

3. An alert nobody trusts is worse than no alert. Early false positives would have killed adoption in the first month.

Approach

I shipped seven separate solutions on one shared pipeline rather than seven separate systems:

• Human detection in active pits and restricted areas

• Unsafe distance monitoring between people and heavy dump trucks

• Light vehicle detection

• Vehicle counting

• Vehicle violation detection at the Coal Processing Plant

• Vehicle dwell time calculation

• Consolidated analytics and evidence reporting

The pipeline pulls RTSP streams, runs YOLOv4 and later YOLOv8 inference, applies zone and rule logic on top of raw detections, and pushes events over WebSocket to a live dashboard. Rules live outside the model, so a change in restricted-area boundaries is a configuration edit, not a retraining cycle.

Datasets were built from the client's own footage and labelled through Roboflow, weighted deliberately toward night shifts and dusty conditions because that is where the early versions failed.

Results

• 7 AI safety solutions delivered and deployed across 15+ mining sites

• 24/7 processing of 24+ camera streams

• 200ms response time for critical safety alerts

• 40% reduction in workplace incidents

• 60% improvement in operational compliance

• Web platform serving 200+ concurrent users with real-time dashboards, WebSocket notifications and automated reporting

• 65% database performance gain through query optimisation, with automated housekeeping over 10TB+ of evidence data

Notes

The engineering lesson was that model accuracy was never the bottleneck. Stream stability, alert precision and storage discipline decided whether the system stayed in use. The 10TB of evidence video would have buried the database inside a year without an automated retention policy, and that policy took as much design thought as the detection model.