Draft: under review before publication.
When we tell safety managers that software can check PPE compliance on their cameras, the good ones push back with the right question: what exactly can it verify? Here's the honest answer.
What vision does well
- Presence of a hard hat or hi-vis vest on a person. This is a well-studied detection problem. On a decent camera, in decent light, modern models handle it reliably, including at the small pixel sizes a wide site camera produces.
- Position relative to a zone. Whether a person is inside a drawn region of the image is geometry, and geometry is cheap. Tracking a person across frames to establish they've stayed in a zone is harder but solved well enough to be useful.
- Persistence. Software never blinks. The violation that happens for the fourth time at 4:47pm, when everyone's tired, looks identical to the model as the one at 8am.
Where it gets harder
- Worn versus worn properly. A hard hat carried under an arm is easy to distinguish from one on a head. A chin strap left undone at height is often genuinely invisible at camera resolution. Anyone claiming to verify strap compliance from a gate camera should be asked to demonstrate it.
- Occlusion and clutter. Workers behind rebar, half-inside an excavation, or clustered around a pour will sometimes be missed or double-counted. Good systems report their blind spots instead of pretending coverage.
- Rare PPE classes. Hearing protection and safety glasses are small, low-contrast objects. Detecting them dependably usually needs closer cameras than sites actually have.
The camera matters as much as the model
The most common failure mode we see isn't the model; it's expectations set without looking at the camera. A feed that's backlit at golden hour, or mounted for gate security rather than work-area coverage, has a ceiling no model can break through. An honest deployment starts with an honest map: which cameras can support which checks, and where the gaps are.
The practical standard
The right question for any vision-based safety tool isn't whether it's perfect; human spot checks aren't either. It's whether the system catches real violations a walking manager would have missed, with few enough false alarms that people keep listening, and whether it's honest about the difference between what it verified and what it inferred. That's the standard we hold our own models to.