Computer-vision safety monitoring for the factory floor

The challenge

On a manufacturing floor, safety runs on attention, and attention is the thing that runs out. Supervisors cannot watch every zone at once. Someone steps into a red zone for a few seconds. A machine keeps running while a technician reaches in to service it. Elsewhere, a cell fills with one person too many. Most of the time nothing happens. The one time it does, the result is a serious incident that, afterward, everyone agrees was foreseeable. The manufacturer wanted to give its safety team a second set of eyes that never blinks, and a record it could actually learn from.

The cost of getting this wrong is not counted in dollars first. It shows up in someone going home injured, and in a safety team that only learns about near-misses once they have become incidents. They wanted to move from reacting to what already happened toward seeing it as it happens.

What we did

We built a computer-vision system on Ultralytics YOLO to watch the floor for the situations that matter. It detects red-zone incursions and counts how many people are in a cell. Machines still running while maintenance is underway get flagged too, and new violation types slot in as the safety team defines them.

Detection is only useful if it goes somewhere. Every event the system flags is surfaced to a data lake, so a safety violation stops being a passing moment on a camera feed and becomes a durable, queryable record, something to trend, audit, and learn from.

On top of that data, we built the tools people actually use. A custom review portal lets the safety team work through detected events with the context to triage and confirm them. And a Grafana-based dashboard gives safety management a live picture of the floor, with real-time alerting so supervisory staff know the moment a situation needs a human, rather than an hour later in a report.

Why it mattered

Safety programs tend to run on lagging indicators, on incidents that already happened and audits after the fact, held back by the honest limits of human attention. This flips that. When a red zone is breached or a machine runs while it ought to be locked out, supervisors hear about it in time to act. Because every event is captured in the data lake, the safety team can finally see where risk clusters and steer their effort straight to it. Safety turns from a story you tell after something goes wrong into something you manage before it does.

The range it took

This work brings together computer vision, data engineering, and the daily reality of a factory floor. We trained and tuned the detection models against the specifics of the site. We built the pipeline that carries events into the data lake and designed the portal and dashboards around how a safety team actually works. A model that detects a violation is a demo. The system a supervisor trusts to alert them and a safety manager can stand behind later is the product. We built the second kind.