Hard Hat and PPE Detection: Camera-Based Safety Monitoring and Privacy
Business

Hard Hat and PPE Detection: Camera-Based Safety Monitoring and Privacy

Start with hard hats and vests at checkpoints the camera sees well, send alerts to supervisors and trends to the safety specialist, and settle KVKK before going live.

Short answer

Hard hat and PPE detection uses a camera and an AI model to find each person, check whether they wear the hard hat or vest required in their zone, and alert the area supervisor when something is missing. At fixed gates and defined zones it often runs on existing IP cameras; gloves, glasses and earplugs are much harder to see. Spotting someone without a hard hat needs person detection plus association, not just helmet detection. Because employee footage is personal data under KVKK, purpose, retention, access and employee information must be settled first.

What is hard hat and PPE detection, and can existing cameras do it?

Hard hat and PPE detection is a vision system that finds each person in a camera image, checks for the protective equipment required in that zone, such as a hard hat or high-visibility vest, and alerts the right person when something is missing; for hard hats and vests at fixed gates and defined zones, it can often run on your existing IP cameras. The model is rarely the hard part; camera angles, alert routing and footage storage are.

Türkiye's Occupational Health and Safety Law No. 6331 requires the employer to monitor whether safety measures are followed and correct nonconformities (Article 4), and to determine protective equipment from the risk assessment (Article 10); Article 5 puts collective protection before personal protection. A camera helps with monitoring; it replaces neither the risk assessment nor collective measures.

Which PPE can a camera detect reliably, and which is hard?

Hard hats and high-visibility vests are the most reliable, being large, bright and predictably placed; gloves, safety glasses, hearing protection and whether a harness is clipped on are much harder.

  • Works well: Hard hats and vests at turnstiles, doors and corridors where people walk towards the camera.
  • Hard: Gloves (small, moving, often behind the workpiece), safety glasses (small, transparent), earplugs (usually invisible), harnesses (visible, but whether the lanyard is anchored is another question).
  • Difficult conditions: Distant and crowded scenes, people blocking each other, colour lost in night-time infrared, rain, dust, steam and backlight.

A hard hat detection study on railway construction surveillance footage describes hard hats in such footage as small and frequently obstructed targets. Start the pilot with hard hats and vests; add other PPE after measuring it on your own footage.

How do you detect a person without a hard hat?

Finding a person without a hard hat takes more than detecting hard hats: the system finds the person, looks at that person's head and ties the result to them. Counting helmets misleads, since one carried in the hand or left on a shelf counts as "worn". The model then separates a hard hat from a beanie or hood, follows the person over several frames, and decides once the head is clearly visible.

The outcome needs three states: compliant, non-compliant and unknown. The Ultralytics Construction-PPE dataset documentation states the same principle in its labelling rules: an occluded or out-of-frame body part does not by itself establish missing equipment. Unknown raises no alert; it is only logged. The tracking number is not an identity, just a few seconds of tracking in one camera.

Are our existing IP cameras good enough?

Existing cameras are often enough when the head covers enough pixels, the angle is right and a network stream is available; wide-angle surveillance cameras on the ceiling usually are not. If a 1,920-pixel-wide image covers 20 metres, a 25 cm head gets about 24 pixels: borderline for a hard hat, too few for glasses. Aimed at a 6-metre gate, the same camera puts 80 pixels on the head.

  • Angle: A camera looking straight down sees the hard hat but not the face or glasses; a view from the front and slightly above suits most PPE.
  • Frame rate: At walking speed, about 1.4 metres per second, 5 frames per second still gives over 15 frames across a 5-metre view; sharpness matters more.
  • Night: Infrared images are black and white; the vest's colour vanishes, only reflective tape remains. Night checkpoints may need lighting.
  • Edge processing: Processing on site, next to the recorder, saves bandwidth and keeps raw footage on the premises; only events leave.

How do you define zone rules?

Rules belong to zones, not cameras: the risk assessment says which PPE is mandatory where and when, and each rule is tied to an area drawn on the image. Türkiye's Regulation on the Use of Personal Protective Equipment at Workplaces ties equipment choice to the risk assessment (Article 7) and reserves PPE for risks that collective measures, work organisation or working methods cannot prevent (Article 5).

ZoneMandatory PPETime and conditionCamera note
Production entrance turnstileHard hat, vestEvery shiftFront view, slightly above; the most reliable checkpoint
Loading dock, forklift routeVest, hard hatLoading hoursIf people appear small, check the vest only
Grinding and welding cellGlasses or face shield, glovesWhile the machine runsNeeds a close camera; measure in the pilot first
Press lineHearing protectionWhile the line runs; off during maintenanceEarmuffs visible, earplugs not
Work-at-height platformHard hat, harnessWhile a work permit is openAnchor attachment stays with the permit check
Office, canteen, break areaNoneNoneAnalysis off
Changing room, toilet, shower, prayer roomNoneNoneNo cameras

Who should receive alerts, and how should the system support the safety process?

Real-time alerts go to the area supervisor and aggregate reports to the occupational safety specialist; the system shows where and why non-compliance repeats and is not there to punish individuals.

  • Real-time alert: A light or screen at the gate and a short message to the supervisor, at most one per person every few minutes. Too many alerts get ignored within days.
  • Aggregate report: Rates by zone, shift and PPE type; "vest use on dock B is low at night", not a ranking of people.
  • Confirmation: A person marks each alert true or false from its clip; the marks are the quality measure and new training data.
  • Integration: The video management system (VMS), a light or siren, a turnstile display and the safety reporting tool. No ERP link is needed.

Repeated non-compliance often points at the equipment: hard hats that clash with earmuffs, hot vests, no spares at the gate. Article 6 of the regulation requires PPE to be free, within easy reach in sufficient numbers, and backed by practical training; Article 18 of Law No. 6331 requires consulting workers on new technologies. EU-OSHA's report on smart digital monitoring systems lists the concern that such systems become a pretext for surveillance and performance measurement among the barriers, and recommends treating such systems as tools that improve safety through corrective measures and training, not as ends in themselves.

What does KVKK, Türkiye's data protection law, require?

Camera images of identifiable employees are personal data, so PPE detection is processing under Türkiye's Personal Data Protection Law No. 6698 (KVKK). This is information, not legal advice; the employer's lawyer or data protection adviser must assess the specific set-up.

  • Article 4 requires processing to be lawful and fair, for specified, explicit and legitimate purposes, relevant, limited and proportionate, with data kept only as long as necessary; Article 5 lists the legal bases.
  • In decision summary 2022/797, the Personal Data Protection Board assessed safety camera monitoring at a hazardous-class factory under legal obligation, establishing a right and legitimate interest, while stating that Article 4 applies in every case. It listed camera numbers, angles, areas, retention and access as factors, and stressed the expectation of privacy in changing rooms, toilets, showers, prayer rooms and rest rooms. It treated face recognition for entrance identity checks as biometric processing; under Article 6, biometric data is a special category.
  • Article 10 requires telling employees the controller's identity, purpose, recipients, collection method, legal basis and their rights; Article 12 requires security measures and makes the controller jointly responsible with a vendor processing on its behalf.
  • Article 11 gives a right to object to an adverse outcome of solely automated analysis; no alert should affect anyone without human review.

A checklist before installation:

  1. A written, narrow purpose: see non-compliance by zone and shift and give real-time alerts. No performance measurement.
  2. No face recognition or matching against staff lists unless separately justified.
  3. Faces in stored clips are blurred, or only the person crop is kept.
  4. A short, written retention period with automatic deletion.
  5. Role-based access: supervisors see their zone, the safety specialist sees aggregates; every access is logged.
  6. Footage is processed on site; a data processing agreement is signed with the vendor.
  7. Workers and representatives are informed in advance; privacy notice and camera signs are in place.

What data and access does the project need from us?

You need a camera list with a day and a night frame from each camera, a zone plan, the PPE matrix from your risk assessment, a safety lead who can decide, and stream access. Testing needs clips from your own cameras across shifts and weather; the dataset documentation above also recommends final evaluation on independent footage from the deployment site. Staff lists, face databases and HR records are not needed.

How do you run the pilot and decide on acceptance?

The pilot starts on a few cameras in shadow mode, where alerts reach nobody, and results are measured per PPE type and per camera.

  • Recall: Of 100 people who really passed without a hard hat, how many triggered an alert; the rest are missed violations.
  • Precision: Of 100 alerts raised, how many were real violations; the rest are false alarms that drive alert fatigue.
  • Confirmed alerts per shift: How many alerts arrive, and how many does a person confirm?
  • Unknown rate: The share of undecided passes; if high, move the camera.

Report by PPE type, camera, shift and day versus night, not as one average. Set targets from today's site-walk observations and the cost of misses and false alarms, not from a vendor brochure. The test set holds clips not used in training. Our article on acceptance tests for a visual inspection pilot covers shadow mode and shift-by-shift test sets.

Hypothetical example: how accurate are hard hat alerts at a warehouse gate?

This example is hypothetical; the numbers only illustrate the method. A logistics warehouse runs hard hat and vest checks in shadow mode for a week on existing cameras over three turnstiles. Over 21 shifts the system raises 160 hard hat alerts; a safety technician checks each clip.

ShiftAlertsConfirmedPrecision
Day and evening (14 shifts)907886.7%
Night (7 shifts)703448.6%
Total16011270%

Overall precision is 112 / 160 = 70%, but the average hides the problem: more than half of night alerts are false, because a white beanie and a white hard hat look alike in infrared. For recall, the team hand-labels 600 random passes from the same week; 30 show a person without a hard hat, and the system alerted on 26: 26 / 30 = 86.7%. The team adds night lighting, arranges single-file passage at shift change and measures again. About 7.6 alerts per shift (160 / 21) is a manageable load. The reports also show that subcontractors using the side gate find no spare hard hats; the fix is a locker, not a camera.

What drives the cost, and what should a quote itemise?

Cost is driven by the number of cameras and zones, the PPE types, camera condition, night and outdoor conditions, and integration scope. New cameras or lighting, gloves and glasses, many sites with different rules, blurring and turnstile integration make it bigger. Hard hats and vests at one gate are a small job where a ready-made analytic may suffice, if measured on your own footage. A quote should list separately: camera and edge hardware, data collection and labelling, software and model, integration, commissioning and pilot measurement, and monitoring and retraining.

Why do these projects fail?

  • Learning only in the pilot that heads get a few dozen pixels on surveillance cameras.
  • Counting hard hats without tying them to people.
  • Treating unknown as non-compliant and burying the supervisor in alerts.
  • Announcing the system as a disciplinary tool and losing workers' trust.
  • Going live before the KVKK work: no written purpose, retention period or informed workers.

Hatson Tech builds PPE detection systems of this kind as part of its image recognition and computer vision service. If you would like to review your cameras, zone rules and KVKK boundaries together, the first meeting and the discovery work are free.

Key takeaways

  • Start with hard hats and high-visibility vests at gates and corridors; measure gloves, glasses and hearing protection on your own footage before promising them.
  • Detect the person first, then check the head, and keep an unknown state so that occluded heads do not become alerts.
  • Check pixels on the head, camera angle and night images before reusing surveillance cameras; process the streams on site.
  • Write PPE rules per zone and time from the risk assessment, and keep changing rooms, toilets and prayer rooms out of view.
  • Send real-time alerts to the area supervisor and trends to the safety specialist, and measure precision and recall per PPE type, camera and shift.
  • Settle purpose, retention, access, blurring and the privacy notice before going live, and have your lawyer or data protection adviser review the set-up.

Frequently asked questions

Can a camera automatically detect workers without hard hats?

Yes, when the head is clearly visible. The system detects the person, checks the head region, follows the person over several frames and only then decides. Heads that are hidden or out of frame should give an unknown result, not an alert, otherwise false alarms quickly bury the supervisor.

Do we need new cameras for AI-based safety monitoring?

Not always. Existing IP cameras work when the head covers enough pixels, the angle shows people from the front and slightly above, and a network stream is available. Wide-angle ceiling cameras and infrared night images are the usual weak points; a closer camera or extra lighting at key gates often solves them.

Does a PPE compliance monitoring system need face recognition?

No. Checking PPE needs a person and a head region, not an identity. In decision 2022/797 the KVKK Board treated identity verification by face recognition as biometric processing, which Article 6 of Law No. 6698 makes a special category of personal data. Keep identity matching switched off unless it is separately justified and assessed.

Do we need employees' explicit consent for safety camera monitoring?

Article 5 of KVKK lists other legal bases besides explicit consent, and in decision 2022/797 the Board assessed safety cameras under legal obligation and legitimate interest. The same decision noted the power imbalance in employment and warned against relying on consent when another basis applies. Which basis fits your set-up is a question for your lawyer; employees must be informed under Article 10 either way.

Will the system replace the occupational safety specialist or site supervisor?

No. It counts and flags; people decide. Every alert should be confirmed by a person before it has any consequence, and the specialist uses the trends to fix causes such as missing spares, poor fit or unclear zone rules, alongside training and collective measures.

What determines the cost of a PPE detection project?

The number of cameras and zones, which PPE types are checked, whether new cameras or lighting are needed, night and outdoor conditions, labelling of your own footage, privacy features such as blurring, and integration with turnstiles, sirens or the video management system. A quote should itemise these after a site survey.

All articles