Computer vision for site safety: what works on real sites

What cameras can reliably detect

Computer vision is software that watches camera feeds and flags what it has been trained to recognise. On a construction site, a factory floor or a warehouse, it works best on things that are visible, distinct and happen in a known place.

Good candidates include a person without a helmet or high-visibility vest, someone entering a restricted area, a person standing too close to a moving machine or vehicle, and a blocked exit or walkway. On a production line, cameras can check for a missing part, a wrong label, a surface defect of a known type or a count that does not match.

Weaker candidates depend on intent or fine judgement, such as whether a lift is being done carefully. A system can flag a posture or a position. It cannot know what someone was thinking.

Site conditions decide accuracy

A model that performs well in a demonstration can struggle on your site, because sites are hard places for cameras. Light changes through the day and drops at night. Low sun and welding create glare. Dust, rain and fog reduce contrast, and lenses get dirty. Scaffolding, stacked material and vehicles hide people from view.

Camera placement matters more than most buyers expect. A camera mounted too high sees only the tops of helmets. One mounted too far away gives the model a person only a few pixels tall. Each detection needs a clear view of the place where the risk occurs, at a suitable distance and angle. Existing security cameras can often be reused, but check each one against the specific detection you want before you count on it.

• Lighting across day, night and glare
• Dust, rain and dirty lenses
• Occlusion by material, vehicles and structures
• Camera height, distance and angle per detection

The false-alert problem

The largest risk in a safety vision project is a system that alerts too often. If a supervisor receives dozens of alerts in a shift and most are wrong, the supervisor soon stops looking. From then on a true alert is ignored along with the false ones, and the site believes it is monitored when nobody is watching. That is worse than having no system, because the cameras create confidence that nothing supports.

So judge a system by the alerts people act on. Ask how many alerts a site will see per shift and what share of them are correct on your own footage. Favour fewer, better alerts at the start. Add rules that cut noise: alert only when the condition lasts several seconds, only inside marked zones, and once per event instead of on every frame.

• Alerts per shift, measured on your own footage
• Share of alerts that are correct
• Minimum duration before an alert fires
• Zones that limit where detection applies
• One alert per event

Human verification

A camera system should inform a person who decides. For most detections, send the alert with a short clip or image to a named supervisor or control room, and let that person confirm or dismiss it. Each confirmation or dismissal is feedback that can be used to retrain the model on your conditions.

Agree the response before go-live: who receives each alert type, how quickly they must respond, and what happens next, such as a call to the area supervisor, a stop of work or a note for the next safety briefing. Reserve automatic responses such as a siren for a small number of high-risk cases where detection has proved dependable, and test them thoroughly. An alert with no owner and no agreed response is only a recording.

Privacy boundaries for people on camera

These cameras film people at work, so set the boundaries early and write them down. State the purpose plainly: safety and quality. Then keep the system to that purpose. Using safety footage to measure individual productivity will lose the trust of the workforce, and the programme depends on their cooperation.

Practical measures help. Detect conditions and events without identifying individuals wherever that meets the safety goal. Avoid face recognition unless there is a clear and agreed need. Keep clips only as long as needed and limit who can view them. Keep cameras out of rest areas and changing rooms. Tell workers and contractors what is monitored and why. Check your plans against the privacy and compliance requirements that apply where you operate.

• Stated purpose: safety and quality only
• Events detected, individuals not identified by default
• Short retention and limited access to clips
• No cameras in rest or changing areas
• Clear notice to workers and contractors

Quality checks follow the same rules

Data Prophets has built computer vision for logistics, construction, manufacturing and retail, including site safety monitoring, restricted-area monitoring and quality checks. One lesson repeats across them: conditions on the ground shape the result as much as the choice of model.

On a production line, conditions are easier to control than on an open site. You can fix the lighting, the camera position and the distance, and that control is worth using. Define each defect with example images your inspectors agree on, since a model cannot learn a standard that people apply differently. Keep an inspector reviewing flagged items and a sample of passed items, so that misses are found.

How to pilot on one site or one line

Start small. Choose one site or one production line, and two or three detections where incidents or defects are already a known problem. Record footage across day and night shifts and different weather before anyone quotes an accuracy figure. Have your own safety or quality staff label a sample, so the model learns your helmets, your vests, your products and your layout.

Run the system silently first, with alerts going only to the project team. Compare its alerts with what supervisors and inspectors recorded in the same period. When alert volume and accuracy are acceptable, switch on live alerts for one shift, then for all shifts. Agree in advance what result would justify a second site and what result would stop the project.

• One site or line, two or three detections
• Footage across shifts and weather
• Labels from your own safety or quality staff
• Silent run before live alerts
• Written criteria for expanding or stopping

What to do next

Walk one site or line with your safety or quality lead and list the three incidents or defects you most want to catch. For each, note where it happens and whether a camera can see it clearly. That list is the brief for a pilot. Our free AI readiness report can help you place it among your other priorities.