Spotting a missing hard hat is a solved problem on a clean image. Making it useful on a working floor is not, and the gap is almost never the model. It is the rule layer: whether the person is in a zone that requires that PPE, whether they have been there long enough for it to matter, and whether this event has already been raised. Skip that and you get a camera producing thousands of technically correct detections a day that nobody can act on. Then the system gets switched off, and the accuracy figure in the report does not explain why.
Every PPE detection pilot demos well. Someone walks into frame without a hard hat, a red box appears, the room nods. Six weeks later the alerts are muted and the dashboard is a browser tab nobody opens.
This is such a consistent pattern that it is worth treating as the design problem rather than an implementation failure. The detector was fine. What was missing was everything between the detection and a human being asked to do something.
The detector is the easy part
Object detection for safety equipment is well-trodden research ground, and published work on hard hat and vest detection reports strong benchmark accuracy for modern detectors. That is genuinely useful — it means you are not inventing anything.
It is also not transferable. Benchmark figures come from curated datasets: reasonable angles, cooperative lighting, people mostly unobstructed. Your floor has racking, steam, a shutter that floods one bay with afternoon sun, and a camera mounted in 2019 for a reason that had nothing to do with this.
So treat published accuracy as evidence the approach works, and your own footage as the only evidence of whether it works here. And measure two numbers, not one:
- Missed violations — what the safety case cares about.
- False alerts — what decides whether the system is still running in three months.
Teams report the first and are surprised by the second. A system that is right most of the time but cries wolf twice an hour has an excellent accuracy score and a short life.
Not all PPE is equally detectable
There is a hierarchy here, and it is mostly about size, contrast and occlusion:
- Reliable: hard hats and hi-vis vests. Large, brightly coloured, usually unobstructed, visible from a distance and from above.
- Harder: safety glasses, gloves, ear defenders. Small, frequently occluded by hands, posture or equipment, and sometimes only a few pixels across at real camera distances.
- Hardest: anything requiring you to distinguish a type — the right glove for the chemical, the right boot for the zone. That is fine-grained classification at low resolution, and it is where confident-sounding projects quietly underdeliver.
Adding classes makes the problem harder rather than simply broader — more ways to be wrong, more confusion between similar items, more labelled data needed per class. Start with the one or two items that carry the real risk on your floor. "Detect all PPE" is a specification that sounds thorough and produces something unusable.
From detection to something a human can act on
This is the part that gets skipped, and it is where these systems are actually built or lost.
Each filter is doing real work:
Zone. PPE rules are spatial. A hard hat is required under the crane and not in the corridor to the canteen. Without zone geometry, every person in every frame anywhere is a potential violation — and the corridor generates most of your alerts.
Dwell. Someone crossing a bay to reach a door is not the risk the rule exists for. Requiring a detection to persist for a meaningful interval removes an enormous share of noise at almost no cost to the safety case.
Deduplication. A person standing in a zone for ten minutes at fifteen frames a second is one event, not nine thousand. Collapsing frames into events is what turns a stream into a list.
What actually wrecks accuracy here
Industrial floors are hostile to vision in specific, diagnosable ways:
- Occlusion. Machinery, racking, pallets, forklifts, other people. A clean line of sight at installation is often gone after the next layout change.
- Mixed lighting. Daylight through a shutter plus sodium or LED overhead. Strong shadows and blown highlights in the same frame are harder than uniform dim light.
- Airborne interference. Dust, steam, spray, coolant mist.
- Motion blur. At the frame rates security cameras usually run, anyone moving quickly smears.
- Camera angle. Steep downward mounts compress a person into very few vertical pixels and hide exactly the head-and-face region PPE detection needs.
Note that none of these are model problems, and most are cheaper to fix by moving a camera or changing a lamp than by retraining anything. This is also why the camera survey matters more than the model selection — the same point that governs whether you can retrofit analytics onto cameras you already own.
The part that is not technical at all
A system that watches people working is regulated, and the rules differ by jurisdiction — that needs local legal advice, not a paragraph from us.
What we can say from building these is that the deployments that survive share a shape, and it is a deliberate one:
- A leading indicator, not a disciplinary tool. The output is "compliance in the press bay drops on the night shift", not a file on an individual.
- Alerts to a supervisor, not to a record. Someone goes and has a word. Nothing is stored against a name.
- Identity not retained. You need to know a person is in a zone without PPE. You do not need to know which person, and choosing not to know is both a legal simplification and the reason the floor tolerates it.
- Told in advance. What is measured, why, where the cameras are, what is kept, who sees it. Discovered monitoring ends projects regardless of accuracy.
Designed as individual surveillance, these systems generate resistance that no accuracy figure overcomes. Designed as a safety instrument, they are usually welcomed by the people who have been asking for better data on near-misses for years.
How to scope the first phase
- One zone, one or two PPE items. The zone with the real risk, not the one with the best camera.
- Draw the zone geometry before writing any detection code. If you cannot state where the rule applies, you cannot filter, and without filtering nothing else matters.
- Test on your own recorded footage, including the night shift, the dusty hour and the afternoon sun.
- Set the false-alert budget explicitly. How many alerts per shift will a supervisor tolerate before ignoring them? Tune to that number, not to a benchmark.
- Agree the escalation path with the safety lead first. Who is told, how, and what they are expected to do. If that is undefined, you are building a dashboard.
Done in that order, the first phase answers a question worth answering: does this surface things your existing process misses, at an alert volume someone will actually absorb? That is a far more useful result than a high accuracy number on a camera chosen because it was convenient.
Questions we get asked
How accurate is PPE detection in practice?
Published research reports strong benchmark accuracy for hard hat and vest detection, but benchmarks use curated data and do not transfer to a production floor. The figure that matters is measured on your own footage — as both missed violations and false alerts, because the second number decides whether anyone keeps using it.
Which PPE can realistically be detected?
Hard hats and hi-vis vests are the most reliable — large, high-contrast, usually unobstructed. Glasses, gloves, ear protection and correct boot type get progressively harder: small, often occluded, sometimes near-invisible at factory camera distances. Adding classes makes the problem harder, not just broader.
What's the difference between a detection and an alert?
A detection is per-frame model output; an alert is something a human must act on. Between them sits zone logic, dwell logic and deduplication. Without that layer you get a continuous stream of technically correct detections nobody can use — the most common way these deployments fail.
What wrecks accuracy in an industrial environment?
Occlusion by machinery and people, mixed lighting and hard shadows, dust and steam, motion blur at low frame rates, and steep camera angles that compress a person into few pixels. These are environmental, not model problems — usually cheaper to fix with a camera move than with retraining.
Is it lawful, and how do we handle the workforce side?
Workplace monitoring is regulated and varies by jurisdiction, so take local advice. In practice the systems that survive are framed as a leading safety indicator: aggregate trends by zone and shift, alerts to a supervisor rather than a personnel record, identity not retained, and the workforce told in advance.
Can we use our existing cameras?
For compatibility, usually yes — most IP cameras expose a stream an analytics server can read. Whether a given camera is useful for PPE is separate, and depends on angle, distance and light where the PPE must be visible. Security cameras are mounted high and wide, which is poor for seeing eye protection.
We build computer vision for industrial environments, running on the client's own network so footage never leaves the site. If you are scoping something like this, the most useful thing you can send us is footage from the awkward camera in the zone that actually worries you.