PPE Monitoring System: How AI Detects Hard Hats and Safety Vests

A PPE monitoring system checks hard hat, vest, glove and safety glasses use 24/7 with existing cameras. How AI PPE detection works and how 90% accuracy is measured.

Translated from the Turkish original · Türkçe aslı

A PPE monitoring system is a video analytics application that uses AI to analyse footage from existing IP cameras and checks hard hat, high-visibility vest, glove and safety glasses use per person, 24/7. The model first finds the person, then classifies the equipment and assigns it to that person; if missing equipment is confirmed over consecutive frames, the shift supervisor receives a notification with an image. No additional camera investment is required.

Why is manual PPE inspection not enough?

A large share of workplace accidents happen at moments when personal protective equipment (PPE) has been taken off or was never put on. Traditional inspection relies on site walks and visual checks; a sample of a few minutes per shift leaves the rest of the day in the dark. What is more, hard hats go on and vests get worn the moment the inspector appears. Whether a hard hat was being worn is always obvious when you look back after an accident; the point is to have that information before the accident, on every shift and for every person.

The scale of the problem is not small. According to ILO estimates, around 3 million people worldwide die every year from work-related accidents and diseases; in Türkiye, SGK (Social Security Institution) statistics record hundreds of thousands of workplace accidents each year. Law No. 6331 (Türkiye's Occupational Health and Safety Law) obliges the employer not only to provide PPE but also to monitor whether the measures taken are being followed. A PPE monitoring system moves this monitoring from sampling to continuity.

How does AI detect hard hats and vests?

In practice the work has three steps. The first is detection: the deep learning model marks every person in the frame with a body bounding box, then searches for hard hats, vests, gloves and glasses as separate classes. The second is association: detected equipment is assigned to the correct person's bounding box based on its position; at a crowded line entrance, the hard hat in the hand of the person next to you is not credited to the wrong person, and a vest left on a shelf is not counted as “worn”. The third is the rule: each zone carries its own PPE requirement, and missing equipment is evaluated together with its confidence score.

A single frame is never trusted. Footage is analysed at 25 frames per second, the person is tracked across frames, and a violation becomes an event only if it is seen consistently over consecutive frames; a worker adjusting their hard hat for a moment does not trigger an alarm. This temporal confirmation is the key to 90% accuracy and a low false alarm rate on site. Person tracking also prevents the same violation from being counted again in every frame; on the dashboard it appears as one event, one record.

Which types of PPE can be detected, and which are difficult?

Hard hats and high-visibility vests take up a large area in the image and have high colour contrast; the most reliable results come from these classes, and the hard hat detection accuracy measured on our solution page is 90%. Gloves and safety glasses, however, are small; in the frame they shrink to a few pixels and can disappear depending on the angle of the hand or face. For these classes the person's distance from the camera and the viewing angle directly determine the result; if necessary, the rule is applied only at points where the equipment is clearly visible.

Equipment such as ear protection, filter masks or safety harnesses is not visible from every camera angle; even when visible, the image cannot tell whether it is the correct type. An additional rule can be defined for masks in a chemical filling area; safety harnesses, however, are not among the ready-made classes. For site-specific equipment a project-specific video analytics model is trained; the typical timeline is 12–14 weeks. The honest approach is to separate the equipment that can be reliably verified by camera and keep the rest under manual inspection.

How are PPE rules defined by zone?

Polygon zones are drawn over the camera's field of view, and a list of mandatory equipment is assigned to each zone. A hard hat can be required at the line entrance, a vest in warehouse aisles with forklift traffic, glasses in the welding area, and glasses and a mask in the chemical filling area. If the same camera sees two different zones, a separate rule is written for each zone. The canteen or office corridor is left outside the rules; a worker who takes off their hard hat and walks into the break area is not counted as a violation.

A time dimension can also be added to the rules: requirements that apply only on certain shifts or during maintenance hours can be defined. No separate setup is needed for subcontractor staff; the camera looks not at who the person is but at what they are wearing. When the same zone logic is combined with area and zone intrusion detection, unauthorised entry and PPE rules are managed from a single dashboard.

Which site conditions determine accuracy?

Accuracy is not a laboratory value but a result that depends on the site itself. On the camera side we look at two conditions: 1080p resolution and the person's body appearing at least 80 pixels tall in the frame. Entry points, views over the line and corridor views give good results; a steep top-down angle is not recommended, because the hard hat is visible but the vest and glasses almost disappear. Distant, wide-angle cameras are enough for counting people but not for PPE classification. Our camera placement guide covers angle and distance decisions in detail.

Lighting is also decisive. Backlight at a doorway, weak lighting on the night shift or a high-visibility vest glaring in headlights make the model's job harder; a camera that switches to infrared mode loses colour information. A beanie or cap can resemble a hard hat, and workwear with reflective strips can resemble a vest. We note these per camera during the site survey; at some points turning the camera a few degrees is more effective than training a new model. Thresholds are calibrated on site during the pilot, and the measured accuracy is reported by comparing it with manually labelled real events.

What happens when a violation is detected?

A confirmed violation becomes a notification in under 2 seconds. The notification reaches the shift supervisor via the web dashboard, email/SMS or WhatsApp, with an image of the moment of the event, the zone name and the missing equipment. At line entrances, a turnstile warning or a light signal can be triggered via an IoT alarm relay; the worker notices the missing item at that moment. The event clip is written to the archive; because events are published over REST API and MQTT, they connect to your existing VMS, MES or OHS software.

The system works with existing IP/CCTV cameras that provide an RTSP stream. In the default setup, processing takes place on an edge server inside the facility; it can also be done in the cloud if preferred. A single GPU server typically handles 8–16 cameras; if heavy models such as pose analysis will also run on the same server, this drops to 6–10 cameras.

The value lies in the distribution, not in individual violations

The value of PPE data lies in its distribution. Once you can see which gate people enter without hard hats, on which shift compliance drops and which equipment is consistently skipped, the intervention turns into a training plan, a layout change or an equipment procurement decision. In the sample dashboard on our solution page, compliance over the last 30 days is 94% for hard hats and 91% for vests, while it is 78% for gloves and 64% for safety glasses. Glove and glasses compliance often turns out to be related to the comfort and accessibility of the equipment; at sites where the locker location was changed, the rate rises noticeably in the first month. In other words, the solution is not always an alert; sometimes it is where a locker stands.

The weekly violation curve on the same dashboard shows a −41% decline over 8 weeks. The first two weeks of the pilot are measurement only, followed by the intervention period; the decline lasts because the alert reaches the shift supervisor instantly and the topic enters the start-of-shift meeting with data. How this curve takes shape in your own facility is measured in the pilot. The shift-based compliance score also feeds into monitoring and measurement activities under ISO 45001.

How do cameras, turnstiles and manual inspection divide the work?

Camera-based PPE monitoring is strong on continuity: the night shift is inspected as closely as the day shift, and every violation is recorded with its zone and time. Its limits are also clear: it does not work in areas the camera cannot see, and it cannot reliably check whether a hard hat's chin strap is fastened or whether the equipment is intact. Turnstile and tag-based systems verify the presence of the equipment at the entrance but not its continuous use inside the area; a hard hat can be carried through the gate in the hand. Manual inspection evaluates context and corrects on the spot, but it is a sample. The best result comes from using them together: the camera shows where the risk accumulates, and the OHS specialist devotes their time to those points.

Beyond PPE: behaviour and site-specific rules

Accidents can happen even when PPE is complete; reaching into a machine, climbing a guardrail or lifting a load with your back are not solved by equipment. For these risks, safe behaviour and pose analysis measures posture by extracting 17 keypoints of the body and can run together with the PPE module on the same camera infrastructure. We described how hard hat and vest checks are set up on a site whose layout changes every week, and why safety harnesses are monitored with a project-specific model, in the construction site safety article, and the monitoring of missing aluminised aprons and face shields in casting areas in the OHS in metal and foundry plants article. How camera data connects to OHS processes as a whole is the subject of the digital transformation in OHS article.

What are the common mistakes in PPE monitoring?

The first mistake is switching on all rules in all areas from day one. A system that starts with uncalibrated thresholds floods the shift supervisor's phone with alerts and is soon muted; in PPE monitoring, precision comes first. We explained the combined use of confidence thresholds, zone rules and temporal confirmation in the reducing false alarms article. The second mistake is not testing night footage and site-specific look-alike objects. The third is positioning the system as a disciplinary tool; when violations are counted anonymously and used as training data, employee resistance decreases.

Is a PPE monitoring system KVKK-compliant?

PPE monitoring does not use facial recognition and keeps no biometric records; the system knows not who someone is, but whether the hard hat is on their head. In the default setup, footage is processed on an edge server inside the facility and raw video is not sent to the cloud (processing can also be done in the cloud if preferred); event clips are automatically deleted at the end of the defined retention period. Obligations under Law No. 6698 on the Protection of Personal Data, KVKK (Türkiye's Personal Data Protection Law), such as the privacy notice, the purpose of processing and the retention period, remain with the employer as data controller. We summarised the general framework in the KVKK and video analytics article; this is not legal advice.

Where should you start with a PPE monitoring system?

The right starting point is a narrowly scoped pilot in one or a few zones where violations and risk are most concentrated. First, the suitability of existing cameras for PPE analysis is assessed in terms of angle, distance and lighting; CX Teknoloji carries out this assessment free of charge before installation. Installation is typically completed within a week, followed by a 30-day pilot. Acceptable accuracy and the false alarm level per shift are set in writing before the pilot; for a framework, see the designing the right pilot article.

Frequently asked questions

Does a PPE monitoring system work with our existing cameras?

In most cases, yes. IP cameras that provide an RTSP stream are sufficient; the recommended conditions are 1080p resolution and the person's body appearing at least 80 pixels tall in the frame. Entry points, views over the line and corridor views give good results; a steep top-down angle is not recommended. Before installation we assess your camera inventory free of charge and report which points are suitable and, if necessary, where cameras should be added.

Which types of PPE can be detected?

The ready-made model searches for hard hats, high-visibility vests, gloves and safety glasses as separate classes and associates each with the relevant person; additional equipment rules such as masks can be defined in chemical areas. Results are more reliable for large, high-contrast equipment such as hard hats and vests; for gloves and glasses, camera distance and angle are more critical. For site-specific equipment such as safety harnesses, a project-specific model is trained.

How accurate is PPE detection?

The hard hat detection accuracy measured on our solution page is 90%. Accuracy varies with the camera angle, the person's size in the frame, lighting and the type of equipment. A single frame does not trigger an alarm; the violation is confirmed over consecutive frames, and thresholds are calibrated on your site during the pilot. The accuracy measured in the pilot is reported by comparing it with manually labelled real events.

How long do installation and the pilot take?

PPE is a ready-made solution; because existing cameras are used, installation is typically completed within a week, followed by a 30-day pilot. The first two weeks of the pilot are the baseline measurement period: alerts are compared with the actual situation on site, and zone rules and thresholds are calibrated. In the following weeks, notifications are opened to the shift supervisor and the violation trend is tracked. The rollout decision is made based on the pilot's measurement table.

Are employees monitored individually, and what about KVKK?

The system does not use facial recognition and keeps no biometric records; violations are reported not by personal identity but by zone, shift and equipment. In the default setup, footage is processed on an edge server inside the facility and raw video is not sent to the cloud (processing can also be done in the cloud if preferred), and event clips are deleted at the end of the defined period. Obligations such as the privacy notice, the purpose of processing and the retention period lie with the employer as data controller; this information does not replace legal advice.

How quickly are notifications sent, and to which systems?

The time from violation to notification is under 2 seconds, and every notification includes an image of the event. Options include the web dashboard, email/SMS, WhatsApp, an IoT alarm relay and REST API; a turnstile warning or light signal can be triggered via the relay. Because events are published over REST API and MQTT, they connect to your existing VMS, MES or OHS software.