Safety Monitoring on Construction and Warehouse Sites With Computer Vision

Near misses are the data nobody records

A distribution warehouse runs 12 forklifts across two shifts. The accident book has three entries this year. Anyone who has stood on that floor knows the real number of close calls is far higher, because nobody fills in a form when a truck swings round a racking end and a picker steps back just in time.

Safety monitoring with computer vision is at its most useful here. It sees the near misses that never get reported, shows where and when they cluster, and gives the safety manager evidence to redesign a walkway or change a pick route.

It is least useful, and most corrosive, when used as a camera that emails a supervisor every time someone lifts their hard hat to wipe their forehead.

What vision can detect on site

Alerts, trends or both

Two quite different systems hide under the same name.

Real-time alerts fire when something is happening now: someone walked under a suspended load, or a pedestrian entered a forklift-only aisle. They need low latency, usually processing on site, and very careful tuning, because a system that cries wolf ten times a shift gets muted by lunchtime.

Trend analysis processes footage in batches and produces a weekly picture: 40 proximity events this week, 28 of them at the junction by dock four, mostly between 2pm and 4pm. That is the version we recommend starting with. It creates no alert fatigue, involves no individual being singled out, and produces the kind of evidence that changes layouts.

Both can be connected to your existing reporting through workflow automation, so a pattern becomes a task on the safety manager's list rather than a chart nobody opens.

Getting the workforce on side

1. Consult the site team and any union or safety representatives before installing anything
2. Be explicit in writing about what the system detects and what it will never be used for
3. Blur faces by default in stored clips and reports
4. Report hazards by area and time, not by named individual
5. Share the trend reports with the people on the floor, not only with management

If the workforce believes the cameras are there to catch them, they will find the blind spots. If they believe the cameras are there to fix the junction by dock four, they will point out the next one.

There are also legal reasons. In the UK, workplace monitoring requires a data protection impact assessment and transparency with staff. Under the EU AI Act, systems used to monitor and evaluate workers' behaviour can fall into the high-risk category, which brings real documentation and oversight obligations.

Construction sites are harder than warehouses

A warehouse has fixed lighting, fixed racking and fixed camera mounts. A construction site changes weekly. Scaffolding goes up, the camera pole moves, a new floor blocks the view, and the exclusion zone you drew in March is inside a stairwell by May.

• Use mobile camera towers with their own power and connectivity
• Budget for re-drawing zones and re-checking angles as the build progresses
• Expect weather, dust and low winter light to reduce accuracy
• Keep detection simple, such as zones and PPE, rather than ambitious posture analysis

For firms already digitising site paperwork, pairing vision trends with digital site forms for near-miss reports gives a much fuller picture than either alone.

When not to use computer vision for safety

If the site does not yet have basic segregation, marked walkways and a working near-miss reporting culture, fix those first. Cameras do not replace barriers. They also should not become a reason to skip a physical control; a pedestrian gate on a forklift aisle beats any alert.

Small sites with a handful of workers and a supervisor present all day rarely justify the cost. And if the plan is to discipline individuals from camera footage, we would decline the work at SpiderHunts; it tends to destroy the reporting culture that safety depends on.

A realistic first deployment

Pick one hazard in one area. For most warehouses that is forklift and pedestrian interaction at the busiest junction; for most construction sites it is people inside a plant exclusion zone. Install or reuse two cameras, run batch analysis for six weeks, and review the findings with the site team every Friday.

Success is not a detection accuracy figure. It is a physical change made because of what the data showed, such as a moved walkway, a mirror, a revised shift pattern, followed by fewer events at that spot. If that happens once, extending to other areas is easy to justify. If six weeks pass and nothing changes on the ground, the problem was never a lack of data. Our wider write-up on computer vision use cases covers the other places cameras tend to pay back.

Frequently asked questions