Edge AI for PPE: from a camera frame to the supervisor’s inbox, without anyone asking
A worker in a pink polo walks past a moving forklift. No hard hat, no hi-vis vest. Twelve seconds later the site supervisor has an email with a clip of it. Nobody pressed a button.
Most “AI safety camera” demos end with a bounding box. A person is drawn in green or red, and the audience is left to imagine the rest. The rest is the hard part: deciding whether the red box matters, explaining why, and getting it to the one person who can do something about it, before they’ve moved on.
This post shows a short clip from our PPE compliance and forklift safety console, and walks through what happens between the frame and the inbox.
Video 1: the walker and the forklift
Recorded from the console · 40 seconds · the clip, the verdict, the events as they land, and the email that went out
What you’re watching, left to right:
- Left pane. The warehouse clip with the system’s own overlay. Person 1, the walker, is boxed red and labelled non-compliant. Person 2, the forklift operator, is also red, but for a different reason: hard hat on, no vest. The forklift is labelled, and the label changes to “moving” when it is.
- Right pane, top. The verdict for the whole clip, decided before you press play: 2 critical: non-compliant person close to a moving forklift.
- Right pane, as the clip plays. Rows appear at the moment they happen. At one second, “Person 1 close to moving Forklift 1” turns red. Under it, a green card slides in: the supervisor has been emailed, with the subject line, the time, and the 15-second clip that went out as the attachment.
- The end of the recording is the email itself, as the supervisor received it. Verdict, the people involved, what each is missing, a plain-language brief, and the clip playing.
One thing worth noticing: the console never asked a question. The chat box at the bottom is there for follow-ups (“what should the supervisor do first?”), but every finding on screen was surfaced by the system on its own. That is the difference between a safety tool and a chatbot with a camera attached.
What happens in between
Four stages, and only two of them involve a model.
Two detectors look at every frame
A general-purpose person detector finds people and follows each one through the clip, so “Person 1” at second one is the same Person 1 at second eleven. A specialist model, trained on PPE and forklifts, finds hard hats, hi-vis vests, and the truck.
We split the job on purpose. A PPE model trained mostly on workers wearing PPE turns out to be weakest at spotting the person who isn’t — which is exactly the person you’re looking for. The general detector doesn’t have that blind spot.
No model decides anything here
A short, readable set of site rules does: compliant means hard hat and vest; a person whose box sits inside the forklift’s for most of the clip is the operator, not a pedestrian; the forklift is moving when it shifts or changes size over a one-second window; a pedestrian is “close” when the gap is under six-tenths of a forklift length, and they’re at the same depth in the scene.
Severity follows from those facts. Critical is reserved for one combination: a non-compliant pedestrian near a forklift that is actually moving.
Because the rules are code, a safety manager can read them, argue with them, and change them for a site without retraining anything.
A small vision-language model, running locally, doing two narrow jobs
It zooms in on each person and answers a strict question: is that a rigid hard hat or a beanie, a fluorescent vest or a dark jacket? And it writes the situation brief and answers follow-up questions, working only from the event list the rules produced.
It reports the verdicts. It never re-judges them.
One action, automatically
A critical event cuts a short clip around the moment and emails the supervisor with the verdict, the brief, a still, and the clip. Warnings on their own don’t email anyone.
That threshold is the whole point of the next section.
The operator who did everything right
A forklift moving a load. The operator wears a hard hat and a reflective vest. The verdict is green: All people compliant; no proximity alerts. The forklift moves, the operator is on it, and no email is sent.
This is the case we’d show a safety manager first. Every camera system that has ever been switched off was switched off for the same reason: it cried wolf. A system that stays quiet when a compliant operator drives a forklift, and only speaks when a pedestrian without PPE is in the path of a moving one, is a system people leave on.
Why “edge”
Everything above ran on one machine on the site side of the network. Video never leaves the site. What leaves is a few kilobytes of events — who, what was missing, when, how close — and, on a critical event, a one-megabyte clip.
That matters for bandwidth, it matters for privacy, and it matters for latency: the decision is made where the camera is, not after a round trip to somebody’s cloud.
The vision stage is designed for the class of edge accelerators we deploy on poles and in cabinets. The rules cost nothing. The explaining stage uses a small local language model today, and can be swapped for a larger hosted one where a site wants richer briefs.
Send us sixty seconds of your own site footage
Fixed camera, dashcam, drone — whatever you already have. We will run it through the same pipeline and send back the annotated clip and the brief, so you can see how it behaves on your site rather than ours.
Get in touchPPE compliance and forklift safety is one module on the Edgegenix platform — the same edge core that runs fire and smoke detection, intrusion and asset inspection, against the cameras a site already operates. More on the detection layer in Edge AI, and on the wider set of operations in Solutions.