Glossary / Safety analytics
What is PPE compliance detection?
PPE compliance detection uses computer vision to determine whether people in a designated zone are wearing required protective equipment such as helmets, vests or safety glasses. The useful systems apply persistence and zone rules, so a finding reflects sustained non-compliance rather than one bad frame. That distinction governs whether operators trust the alerts.
Two things get called PPE detection, and confusing them is why so many deployments end with the alerts switched off.
Detection is a per frame question: does this frame contain a person without a helmet? Compliance is an operational question: did a person work in a designated zone, for a meaningful period, without required equipment? The first is a model output. The second is what a safety officer can act on, and getting from one to the other is rule design, not model quality.
Why per frame accuracy is the wrong metric
Vendors quote per frame detection accuracy because it is the number the model produces. Here is what it means in operation.
Worked example: a worker crossing a zone
Per frame accuracy = 95%, so per frame error = 5%
Probability of at least one wrong frame:
1 − (0.95 ^ 30) = 1 − 0.215 = 0.785
A model that is 95% accurate frame by frame will misclassify at least one frame in 78.5% of three second appearances. If a single non-compliant frame raises an alert, roughly four out of five compliant workers generate a violation. The model is not broken. The rule is.
Require five consecutive non-compliant frames instead. The chance of five consecutive errors on a compliant worker is 0.05 to the fifth power, around 0.00003% per position, which effectively removes that failure mode while delaying a genuine detection by half a second.
This is why persistence rules are the single highest-value configuration in a PPE deployment, and why a vendor quoting only per frame accuracy has not told you the thing that determines whether the system is usable.
The rules that turn detection into compliance
| Rule | What it does | Typical setting |
|---|---|---|
| Persistence | Requires sustained non-compliance before a finding is raised | 3 to 5 seconds of continuous detection |
| Zone | Applies the requirement only where the PPE is actually mandated | Polygon per requirement, not per camera |
| Item mapping | Different zones require different equipment | Helmet in the yard, helmet plus vest plus glasses on the line |
| Grace period | Allows time to don equipment after entering | 5 to 10 seconds inside the boundary |
| Occlusion handling | Suppresses findings when the relevant body part is not visible | Suppress rather than assume non-compliance |
| Deduplication | Collapses repeated findings on one person into one event | One finding per person per zone entry |
Occlusion handling deserves emphasis. When a worker's head is hidden behind equipment, the honest output is unknown, not violation. Systems that treat not visible as not wearing produce most of their false findings from behind machinery, which is exactly where supervisors then stop believing the alerts.
What cameras genuinely detect well
- Reliable. Hard hats and high visibility vests. Large, high contrast, designed to be conspicuous, which is precisely what makes them easy for a model.
- Moderate. Safety gloves and boots, where the item is smaller and often occluded by posture or equipment.
- Difficult. Safety glasses and hearing protection, which are small, frequently occluded and easily confused with ordinary eyewear at typical camera distances.
- Not feasible visually. Whether equipment is correctly fitted, whether a harness is clipped to an anchor point, whether a respirator is sealed. A camera sees presence, not correct use.
The pixel density point from plate recognition applies here too. Safety glasses at 25 metres are a handful of pixels. If the requirement includes small items, the camera placement has to be designed for it rather than inherited from a general surveillance layout.
An operational tool, not an enforcement system
PPE analytics supplements a safety programme. It does not replace supervision, training, toolbox talks or statutory duties, and it should not be presented internally as doing so. Its most common failure is organisational rather than technical: a system that generates findings nobody owns, reviews or closes produces a log of unaddressed violations, which is a worse position to be in than having no log at all.
Before deployment, settle who reviews findings, on what cadence, and what happens when one is confirmed. A weekly compliance rate by zone that a supervisor discusses with a team is worth more than real-time alerts that nobody answers.
Monitoring workers and DPDP
PPE analytics observes employees at work, which brings duties even where the purpose is plainly protective. Reasonable practice is to state the purpose in your notice and posted signage, retain evidence images only as long as the review process needs, report at zone level rather than building individual records where the safety objective can be met that way, and consult worker representatives before deployment rather than after. Rule 4 commences on 13 November 2026. This is general information rather than legal advice.
Frequently asked questions
How accurate is PPE detection?
For helmets and high visibility vests on well placed cameras, per frame detection is generally strong. But per frame accuracy is not the number that matters. What determines usability is the persistence rule: requiring several consecutive seconds of non-compliance before raising a finding removes the overwhelming majority of false alarms that a single frame trigger produces.
Can a camera tell whether PPE is worn correctly?
Only in limited ways. A camera can see whether a helmet is present and roughly whether it is on the head rather than carried. It cannot verify chin strap fastening, harness clipping to an anchor point or respirator sealing. Requirements that depend on correct fitting need physical inspection, not video.
Does PPE detection work on existing CCTV cameras?
Usually yes for helmets and vests, which are large and high contrast. Smaller items such as safety glasses need closer placement and higher pixel density than a general surveillance layout provides. Assess the existing estate per zone against the specific items required, rather than assuming it works everywhere or nowhere.
Why does PPE detection produce so many false alarms?
Almost always because it is configured to alert on single frame detections rather than sustained non-compliance, and because occlusion is treated as violation instead of unknown. Both are configuration problems. Adding a three to five second persistence rule and suppressing findings when the relevant body part is not visible typically removes most of the noise.
Should PPE alerts go to supervisors in real time?
It depends on whether anyone can act within the window. Real-time alerts are valuable where a supervisor is present and can intervene. Where they are not, a daily or weekly compliance summary by zone is more useful, because it feeds the safety conversation rather than accumulating unanswered notifications.
Is monitoring employees with PPE analytics permitted under DPDP?
Workplace safety monitoring is a recognised purpose, but the usual obligations still apply: state the purpose, post notice, limit retention and avoid using the data for purposes you did not declare. Reporting at zone level rather than building individual records is good practice where it meets the safety objective. This is general information, not legal advice.
Running PPE analytics on an existing camera estate?
IndoAI adds PPE compliance detection to the cameras you already have, with persistence and zone rules tuned to your site so supervisors get findings worth acting on rather than an alert stream they learn to ignore.
Talk to an adviserReviewed by Dr. Vivek Gujar, Chief Strategy Officer at IndoAI Technologies Pvt. Ltd., a Pune-based edge AI camera platform founded in 2021. He reviews IndoAI's published technical claims for accuracy. Profile