AI PPE detection is a computer vision system that analyses site camera feeds in real time to spot missing personal protective equipment (PPE), such as hard hats, hi-vis vests or safety glasses, and unauthorised entry into danger zones. It sounds a local alarm, notifies the person responsible and logs the event with its image, monitoring behaviour rather than identity.
The rules are followed while the inspector is on site. What about the rest of the day?
Every health and safety professional knows the pattern. During the inspection round, everyone wears their hard hat and vest. Once it ends, heat, fatigue or "it is only a two-minute job" means PPE comes off. People cut through the forklift aisle or step over the yellow line around the press. The safety specialist cannot be everywhere at once, and on a multi-shift plant or a sprawling construction site, they are rarely anywhere near.
Most sites already have a resource that could close this gap: cameras. Yet footage is usually reviewed only after an accident or incident. The recording explains what happened instead of preventing it. Yet because the data already exists, camera footage is also one of the most practical places for starting with AI in business: a measurable pilot can run on existing cameras without collecting anything new.
In Türkiye, Occupational Health and Safety Law No. 6331 places a duty on employers to ensure workers' health and safety; providing suitable PPE and checking that it is actually worn is part of that duty. AI does not replace the safety specialist in this. It extends their view across the whole site.
What the numbers say
Work-accident figures in Türkiye are heading in a worrying direction. According to the SGK Statistical Yearbook 2025, published by the Social Security Institution, 766,039 workers insured under Article 4-1/a had a work accident in 2025. Calculations from the same data show that manufacturing (NACE 10–33) accounted for roughly 314,000 of these, about 41% of the total, along with 420 deaths.
Of the 1,931 work-accident deaths recorded in Türkiye in 2025, 586, about 30%, occurred in construction. (SGK Statistical Yearbook 2025)
The cost to a business goes beyond compensation and fines. SGK records show 4,639,930 days of temporary incapacity for workers injured in 2025. Each one is a missing person on a planned shift, a task someone else has to cover and disrupted output. We show how to turn these items into an investment case in measuring AI project ROI.
The global picture is similar. The ILO's global estimate of work-related deaths shows that nearly 3 million workers die every year from work-related accidents and diseases, 330,000 of them from work accidents.
Not all of this is caused by missing PPE, but missing hard hats, missing vests and danger-zone breaches are among the most visible precursors of preventable accidents.
How AI PPE detection works
The system has four layers:
- Image source. The site's existing IP cameras (RTSP/ONVIF-capable) or cameras added at critical points. Angle, distance and lighting decide whether a small item of equipment can be seen at all.
- Detection model. The model first finds each person in the frame, then the equipment on them: hard hat, hi-vis vest, safety glasses, mask. Some projects add harder classes such as safety harnesses or gloves.
- Zone rules. Zones are drawn on each camera view and each gets its own rule: "safety glasses mandatory in the welding bay", "no pedestrians in the forklift aisle", "no entry around the press outside shift hours". The logic is the same as zone-based access in factories; the difference is that a camera, not a door, enforces the rule.
- Action. If a breach persists for a set time (to filter out momentary false detections), a local audible and visual alarm is triggered, the zone supervisor is notified, and the event is written to a report with its image.
Example rules by zone
| Zone | Required PPE / rule | Camera need | Action |
|---|---|---|---|
| General production hall | Hard hat, hi-vis | Wide angle, high mount | Warning on zone display, daily report |
| Welding and grinding | Glasses or face shield | Close view, good lighting | Immediate audible alert |
| Forklift aisle | No pedestrians | Along the aisle | Audible and visual alarm, supervisor alert |
| Scaffolding | Hard hat, harness | Outdoor, weather-resistant | Instant alert to site manager |
| Chemical store | Mask, gloves, authorised entry | At the entrance | Checked together with the door |
Know the limits
No model is right every time. A person partly hidden behind a machine, seen from behind, or in poor light or rain is harder to assess. Very small items such as earplugs may not be reliably visible on general-view cameras. That is why the system first runs in log-only mode for a few weeks, false alarms are measured and thresholds are adjusted. A system that keeps raising false alarms is soon ignored, which is worse than having no system at all.
A general PPE model may not recognise your site's helmet colours, company workwear or unusual camera angles. In that case, training a site-specific model on footage from your own cameras cuts false alarms markedly; our custom AI model training guide covers the process from data preparation to deployment.
What the reports should show
The real value lies less in individual alerts than in the data they build up. A good report answers questions such as: in which zones and at what times do breaches cluster? Do they rise at shift changeover, after breaks or during the hottest hours? After a change, such as a newly marked walkway or a switch to a lighter hard-hat model, did breaches fall? These questions ground safety committee discussions in data rather than impressions and make the effect of corrective actions measurable. Turning such reports into an automatic written summary is covered in AI management reporting.
Whether the model recognises a hard hat or vest on site depends on consistently labelled images taken in different light, from different angles and distances. We explain the method in data labelling for AI.
Privacy: monitor behaviour, not identity
Camera footage is personal data wherever it makes a person identifiable, and in Türkiye it falls under Law No. 6698 on the Protection of Personal Data (KVKK), the country's GDPR-style data protection law. PPE detection rarely needs to know who breached a rule; "working without glasses in the welding bay at 14:05" is enough to act on. We recommend these design principles:
- Avoid facial recognition unless it is truly needed. It processes biometric data, which KVKK treats as a special category, and that makes obligations much heavier.
- Blur faces in report images and keep raw footage retention short.
- Inform employees. Explain clearly, through a privacy notice and signage on site, what the cameras are used for.
- Restrict access. Breach reports should be seen by the safety team and zone supervisors, not everyone.
- Tie the purpose to improvement, not discipline. Use the data to see where and when breaches increase, and to decide on training, layout changes or equipment.
If you want expert support with this assessment, our data protection compliance consultancy can run alongside the camera project. Our article on biometric attendance and data protection covers biometric data in more depth. For the general rules on workplace cameras, see our guide to CCTV recording and KVKK.
How we do this at Digital Bridge
As part of our computer vision service, we build PPE and danger-zone monitoring around your site:
- Site survey. Together with your safety specialist, we identify the critical zones in your risk assessment and the rule for each. Scope, phases and cost come as a written proposal.
- Assessing existing cameras. If your RTSP/ONVIF cameras have adequate angle and resolution, we use them; where they fall short, we install cameras and processing units. Central management of the camera estate works with our IP camera monitoring solution.
- Adapting the model to your site. If your own equipment (hard-hat colours, vest types, company workwear) or site conditions reduce the performance of a generic model, we train a custom AI model on your footage.
- Log-only pilot. The system first runs without raising alarms; we measure the false-alarm rate with you and tune thresholds.
- Alerts and reporting. Sirens, beacons or information displays on site, supervisor notifications, and breach reports by zone and time. Because the same team builds the software and installs the hardware, you are never caught between two suppliers when an alarm device and the model disagree.
Construction and contracting projects differ from factories: the site keeps changing, work happens outdoors and subcontractor crews come and go. Our construction and contracting sector page looks at those differences, and our article on construction site access and attendance covers the check at the site gate.
Cameras alone are not enough: safety starts at the gate
PPE monitoring watches the site from inside, but workplace safety starts at the entrance. It is better for a subcontractor worker with missing training, a missing health certificate or no employment notification never to enter the site than to be caught inside without a hard hat. Card access control moves that check to the entry point. Our own product, SmartPass, works with a zone hierarchy on construction sites and factories, and in an emergency it lists, zone by zone, who is inside at that moment. Our article on contractor access and HSE compliance covers this in detail, and our guide to computer vision quality inspection shows how the same camera and vision infrastructure catches defects in production.
Next step
Start with one high-risk zone: which area comes up most often in your near-miss and accident records over the last two years? Take a few sample frames from the cameras covering it and write the zone's rule in a single sentence. Then contact us; we will walk the site with you and scope a log-only pilot. For other uses of AI across images, text and numerical data, browse our complete Artificial Intelligence guide.