A PPE violation is a record with an owner, not a red box on a screen
In short
A PPE compliance event is a violation record, not a frame. It carries the camera and zone it happened in, the rule and rule version that judged it, a start and an end that give it a duration, an evidence clip, the supervisor who owns it, and the state it is in. Detection without those fields produces a screen nobody works and an audit trail nobody can defend.
Key takeaways
- The detection is 1 field. Zone, rule version, duration, clip, owner and state are the 6 that make it workable.
- Zone and duration separate someone crossing a bay from someone working in it unprotected. Both read as "no helmet".
- Events attach to a tracked person, not a frame. At 25 frames per second, an unaggregated 10-second breach is 250 alerts.
- Store the rule version on the event. Change a dwell threshold without it and last quarter's trend is rewritten.
- A closure outcome class — valid, false detection, rule misapplied — is the only feedback the detector will ever get.
A PPE compliance event is a record with an owner, and the detection is one field on it. The rest is what makes it workable: which camera and zone, which rule judged it and at which version, when it started and stopped, the clip that proves it, the supervisor it belongs to, and the state it is in. A system that emits detections and stops there gives a plant a wall of red boxes, which is a different product from a safety programme.
The two fields most often left out decide everything. Zone and duration separate a fitter crossing a press bay on the way to the canteen from a fitter working four minutes under a suspended load with no helmet. Both read as "no helmet detected", and no amount of model accuracy tells them apart.
The fields, and the argument each one ends
| Field | What it settles | What happens without it |
|---|---|---|
| Event id and track id | That 250 frames of one person are one event | A 10-second breach at 25 frames per second arrives as 250 alerts |
| Camera, area and zone id | Where on the site it happened | Every alert needs a human to place it |
| Rule id and rule version | Which policy judged it, as it stood that day | Last quarter gets re-read under this quarter's rule |
| Start, end, dwell threshold used | Passing through, or working unprotected | A 2-second detection weighs the same as a 6-minute one |
| Classes detected and confidence | What the model saw, and how sure it was | A borderline call and a certain one look identical |
| Evidence clip, pre-roll and post-roll | Whether the flag was right at all | Every dispute resolves into seniority |
| Owner, state, closure outcome class | Who must act, whether they did, and whether the flag was valid | Repeat findings look like repeat inaction |
The first row decides whether anyone opens the screen twice. Detection runs per frame; a person exists across frames. In the on-premise PPE and intrusion system we run, person detection feeds a tracker holding a short identity history, so PPE state attaches to a track: an event opens when the track crosses the rule's dwell threshold and closes when it stops. That is the difference between eight alerts a shift and eight hundred — the subject of three hundred alerts a shift and nobody opens them.
Walking past the bay is not working in it
A rule with no dwell fires on the walkway; a rule with no zone fires in the car park. The dwell threshold is a safety judgement rather than a technical one. Entry alone is the hazard in a robot cell, so 1 to 2 seconds is right there. A helmet rule over an assembly bay people legitimately walk through needs 10 to 15 seconds before the record means anything.
The zone half is harder than it looks: a polygon drawn on a camera image is a claim about floor area made from a flat picture. Someone standing well behind the taped line can project inside the shape and produce a genuine record of a violation that never happened, which is why what a restricted zone polygon really means has to be settled first.
Open, acknowledged, actioned, closed — and who owns each
- Open. The system creates the record and owns it. An event that never leaves this state is a measurement, not a control.
- Routed. A named recipient and an acknowledgement window attach to it, from the table in who gets the alert and what they must do.
- Acknowledged. A person and a timestamp. The gap between open and acknowledged is the most useful operational number here, and it degrades long before anyone admits the alerts are being ignored.
- Actioned. Area cleared, worker coached, machine stopped, contractor called, or nothing because the flag was wrong. A short controlled list, not free text alone.
- Closed with an outcome class: valid and corrected, valid but nothing could be done, invalid detection, or rule misapplied. The last two are the engineering backlog.
That outcome class is the only feedback the detector gets from the floor. Sort a fortnight of invalid closures by hour and by camera before changing anything: a cluster after dusk is one bug rather than eighty, covered in when the cameras switch to infrared. A daylight cluster on one camera is usually the model or the geometry, traced in wearing a hard hat and still flagged.
A detection is a claim about a frame. A violation is a claim about a person, a place, a rule and a length of time — and only the second kind can be actioned, argued with, or defended a year later.
What this record deliberately does not decide
- It is not a near miss. A missing helmet is a rule breach, not a measured proximity — the boundary drawn in what a camera may call a near miss.
- It is not a named person. Whether the record identifies an individual, and for how long, is a policy decision argued in coaching signal or disciplinary evidence.
- It is not a report on detector quality. Precision is measured against a labelled sample, not against how many events were closed invalid.
- It is not evidence of coverage. An area with no camera produces no violations and looks immaculate on the trend — the reasoning behind auditing coverage after the fact.
Everything downstream reads this record and nothing else: the daily area summary, the trend, the conversation with a contractor about their crew. Settling the record before choosing a model is the instinct behind what breaks when AI agents run in production, and where our automation work starts. The rest of this silo sits under safety, PPE and site monitoring, part of our industrial vision practice.
Frequently asked questions
Short answers to the follow-ups this page tends to raise.
What should a PPE detection system output for each violation?
One record per person per continuous breach, not a stream of frames. It carries the camera, area and zone, the rule id and version, start and end timestamps with the dwell threshold applied, the classes detected with confidence, an evidence clip, an owner, a state and a closure outcome class.
Should a PPE violation be one event per frame or one per person?
One per person per continuous breach, keyed on a track id. Frame-level output at 25 frames per second turns a 10-second breach into 250 records, which is how alert screens die in their first fortnight. The event opens when the track crosses the rule's dwell threshold and closes when the state stops or the track is lost.
How long should someone be unprotected before it counts as a violation?
It depends on the hazard, and the threshold belongs to the safety function rather than the integrator. Where entry itself is the hazard — a robot cell, a press guard — 1 to 2 seconds is defensible. Where people legitimately cross a bay on foot, a helmet rule needs 10 to 15 seconds of dwell. Whatever the figure, store it on the event.
Does an evidence clip have to be attached to every event?
Yes, or the record settles no arguments. A few seconds of pre-roll and post-roll let a supervisor confirm a flag in about ten seconds rather than debating it, and turn a closure of "invalid" into something engineering can act on.
- ppe compliance
- safety events
- video analytics
- data model
The work behind this page
Builds from our portfolio that this page draws on.
Open Vision PPE Monitoring
Boundary surveillance, PPE compliance monitoring, and intrusion detection via real-time video analytics. Runs fully on-premise — no cloud required.
Safety & ComplianceFactory OS
Production planning and task management for a tier-1 apparel manufacturer — replacing Excel with automated milestone planning, SOP gate enforcement, and real-time visibility.
ManufacturingRead next
- Drawing a restricted zone on a camera view: what the polygon meansThe polygon is the smallest part of a zone. The entry rule, the dwell condition and the assumption about where the floor is are what decide who gets alarmed on.definition
- What a camera may call a near miss, and what still needs a human reportCameras measure distance, speed, dwell and crossings. A near miss is a claim about what nearly happened — so camera events belong in the process as candidates, not as counts.definition
- Wearing a hard hat and still flagged: tracing a false PPE violationOpen the evidence clip and ask whether the head is visible at all. That one question splits false PPE flags into three groups that need three unrelated fixes.diagnostic
- Three hundred alerts a shift and nobody opens themA shift raising 300 alerts and closing 6 has an aggregation fault, not a discipline problem. Five causes, ranked, and the event model that turns hundreds of detections into one incident.diagnostic
- Everything works until the cameras switch to infrared at duskThe collapse tracks the camera's day-night switch, not sunset. Once the infrared cut filter drops out, a colour-keyed vest class is scoring on a monochrome frame it has never seen.diagnostic
- The incident nobody has footage of: auditing coverage after the factMissing footage is a finding, not a dead end. Five causes look identical in the moment and carry completely different remedies, and only one of them needs a new camera.diagnostic
Working on something in this space?
Tell us where you are in a sentence or two. We'll tell you honestly whether we're the right team, and what a sensible first slice of the work looks like.
Start the conversation