Industrial safety
PPE detection in Saudi Arabia, on the cameras you already own
This guide is about detection software: AI that watches existing CCTV for missing hard hats and vests and raises alerts your HSE team can act on. (Buyers looking for physical PPE supply are on the wrong page, cheerfully.) It covers what is honestly detectable, what your cameras must deliver, what the first weeks of tuning involve, and how privacy is handled on a Saudi site.
Solution guide Published 7 min read
01The mechanism
What PPE detection actually does
A detection model watches camera streams for people, and for whether the people it sees are wearing the protective equipment the zone requires. When the answer is no, someone in a hard-hat area without one, a missing high-visibility vest on an active apron, it raises an event: timestamp, camera, zone, event class, evidence frame. That event becomes an alert to whoever the site says should act, and a record the safety system can count.
The honest way to think about it: not a guard that never blinks, but an observation programme whose coverage stopped depending on where the supervisor happened to be standing. The unsafe-act category your observation cards already track, “PPE non-compliance”, starts being recorded at 3am with the same reliability as 3pm, which changes both the volume and the honesty of the leading indicators built on it.
02Honest physics
Detectability differs by PPE class, and a vendor should say so
The classes below are the ones industrial PPE programmes most commonly monitor, with the engineering truth about each. For any specific deployment, ask the vendor, Interkey included, for the currently supported class list in writing, and pilot the classes that matter on your own cameras. A supplier who claims everything detects equally well is describing their hopes.
| PPE class | Detection difficulty | Why |
|---|---|---|
| Hard hat / helmet | The strong case | Large, geometric, high-contrast, worn on the most visible part of the body. This is where PPE detection earns its keep, and where pilots should start. |
| High-visibility vest | Strong | Designed to be conspicuous, which helps the model for the same reason it helps the crane operator. Degrades when vests are faded, filthy or worn under other layers, which is a real Saudi site condition, not an edge case. |
| Coveralls / workwear | Moderate | Detectable as presence/absence of compliant workwear; subtler distinctions (correct type, closed versus open) are harder and need conservative expectations. |
| Gloves, eye protection, hearing protection | The honest “it depends” | Small objects at CCTV distances, frequently occluded by the work itself. Detectable in favourable close-range views; anyone promising reliable glove detection across a general camera estate should be asked to demonstrate it on yours. |
| Fall-arrest harness | The honest “it depends” | In close, dedicated views of work-at-height access points a worn harness is detectable; at general CCTV distances it reads as straps among tools, and whether it is anchored is not a camera question. Pilot it only where cameras were placed for the height work. |
03The ceiling
Your cameras set the ceiling on everything
Detection performance is decided before any model runs, by placement, resolution and light. A camera mounted high on a pole, looking steeply down at sixty metres, produces heads-and-shoulders blobs no software can responsibly judge; the same camera at a gate, face-height at eight metres, is a different instrument. The working test is unsentimental: if a person freezing the frame cannot tell whether the worker is wearing a hard hat, neither can the model.
This is why deployments start with a survey rather than a licence count, and why the survey report divides the estate into three: cameras usable as-is, cameras usable after adjustment (angle, zoom preset, a cleaned dome), and views that need new hardware if they matter enough. Saudi conditions belong in that survey explicitly: dust accumulating on domes, midday glare and heat shimmer, steep sun angles at the hours work actually happens, and night areas lit for people rather than for cameras. None is disqualifying; all of them move the numbers, and a pilot that avoids them is rehearsing a site that does not exist.
04The operational truth
False positives are the real deployment work
Every camera-analytics system produces false alerts at first: a white cap read as a helmet, a worker briefly occluded by equipment flagged as non-compliant, a contractor walking a permitted corridor the zone map drew too generously. What decides whether the deployment survives is not the initial rate but what happens next, because an alert channel that cries wolf gets muted within weeks, and a muted safety system is worse than none: it produces the feeling of coverage without the fact of it.
The first weeks after go-live are therefore a tuning period by design: thresholds adjusted per camera, zones redrawn against how the site actually moves, grace periods set for transitions (the walk from the gate to the changing room does not need an alert), and suppression windows agreed for the conditions that generate noise.
05Architecture
Where the model runs: edge, on-premise, or cloud
The inference location is usually decided by data rules rather than preference, which is exactly the factor most published comparisons skip.
| Option | How it works | When it is the answer |
|---|---|---|
| On-premise / on-site | Inference runs on servers inside the site network; footage never leaves | The default for Saudi industrial and government sites where footage leaving the network is prohibited or simply unacceptable. Costs site hardware; buys the cleanest data answer. |
| In-Kingdom central | Streams or events from several sites processed on infrastructure inside Saudi Arabia | Multi-site estates that want central operations while keeping data in-Kingdom; network capacity between sites becomes part of the design. |
| Cloud outside the Kingdom | Footage processed on infrastructure abroad | Rarely the answer for Saudi industrial video, and the option your data-governance owner should explicitly rule in or out rather than discover. If a vendor assumes it silently, that is your first finding. |
06The last mile
What happens when an alert fires
The question that separates deployments that change behaviour from deployments that generate dashboards. A detection is worth exactly what the response to it is worth, so the workflow is designed with the HSE team before go-live: which events go to the control room versus the area supervisor versus the duty phone, in which tools (the VMS operators already watch, the radio procedures that already exist, the incident system that already counts), in which language, and with what escalation when nobody acknowledges.
The record matters as much as the alert. Each event carries timestamp, camera, zone and class, which is what lets detections feed the observation and leading-indicator reporting the safety system already produces, per zone, per shift, per contractor, and what makes an event usable in an investigation afterwards. Wiring that into permit and incident systems is integration engineering, and it is work Interkey does in-house rather than a connector promise.
People are watching back
Worker privacy, handled like adults
Privacy is a design decision with named owners
A camera programme that surprises its workforce deserves the resistance it gets. The workable pattern is explicit and boring: the site decides, with its HSE and legal owners, what is retained (detection records can be kept without retaining continuous footage), for how long, who can view what, where masking applies, and how employees and contractors are informed before go-live. It helps to be precise that three different things exist, the live stream, the model’s inference, and the event record, because they can be governed differently, and usually should be.
Two honesty notes this page owes you. Face blurring is a reasonable measure and not a magic one: gait, build and clothing can still identify a colleague, so masking is a control within a governance design, not a substitute for one. And on the legal position, Saudi data-protection law applies to workplace footage, this page deliberately does not interpret it; bring your counsel into the design early, and treat any vendor offering legal conclusions as exceeding its brief.
Evaluating automated PPE monitoring at an industrial site? Interkey will scope a pilot around your camera environment, PPE requirements and HSE workflow, starting with a suitability review that tells you whether your estate can support it at all.
07Buyer questions
What buyers ask us
Direct answers to the questions that come up in real evaluations. Anything missing, ask us at the bottom of the page.
Can PPE detection work with our existing CCTV cameras?
Usually for a meaningful subset of the estate; the survey’s whole job is finding that subset. Placement, resolution at working distance and lighting decide it per camera; the freeze-frame test, can a human tell from this view?, is a fair predictor before anyone quotes anything.
What detection accuracy should we expect?
Whatever your pilot measures on your cameras in your conditions, and nothing else. Published accuracy figures, including the one on the SquintPRO page, are starting points for that conversation, produced under reference conditions that are not yours. Insist on measuring both false positives and missed detections; a system tuned to never miss will drown you, and one tuned to never cry wolf will miss.
How long does deployment take?
The honest structure: a survey measured in days, an installation and integration phase driven by how many cameras and systems are in scope, then a tuning period of weeks in which false positives are driven down against real site conditions. Fixed durations quoted before a survey are sales numbers; the sequence, and the fact that tuning is real work, are the commitments worth holding a vendor to.
Does it only do PPE?
PPE compliance is the usual starting point because the condition is clear and the consequence is direct, but the same deployment typically extends to restricted-zone entry and person-vehicle proximity, and on hydrocarbon sites to the permit-adjacent scenarios covered in the oil and gas guide. Start narrow, prove it, then extend: the platform is reused, the tuning is per case.
From the insights library
Continue exploring industrial safety
Scoping a pilot that proves something
The success criteria, measurement discipline and decision gates that make a pilot worth running.
Industry guideAI safety monitoring for oil and gas sites
Classified areas, permit-controlled work and contractor crews: what changes on a hydrocarbon site.
ProductSquintPRO, deployed by Interkey
The detection product and the deployment work around it, including the published accuracy figure and its caveats.
Next step
Scope a PPE detection pilot
Tell us the PPE classes that matter on your sites and what the camera estate looks like — the detectability question becomes specific very quickly.
Or directly
+966-11-2180999 info@interkey.com.saTawuniya Towers, North Tower, 7th Floor, King Fahad Highway, Olaya, P.O. Box 56835, Riyadh 11564, Saudi Arabia
Elsewhere