The sentence heard most often in PPE (personal protective equipment) analytics projects is: “We already have cameras; let’s turn on PPE detection.” The result is usually disappointment; hard hats are often caught, goggles almost never, and on the night shift both become uncertain. Analytics classifies whether the person on camera is wearing equipment such as a hard hat, high-visibility vest or goggles; but if it isn’t clear from the start which equipment is detected reliably under which conditions, the project turns into expectation management. Below we explain how detection works, the camera angle, distance, resolution and lighting requirements, the choice between alarms and reports, and integration with the OHS process.
How PPE detection works
PPE analytics usually works in two stages. First, the model finds people in the image and produces a bounding box for each. Then the region inside this box (head, torso, face) is passed to a separate classifier or to additional outputs of the same model, producing a “hard hat yes / no,” “vest yes / no” decision. Some systems make the decision not on a single frame but after tracking the person for several frames; this approach reduces momentary classification errors.
This structure has two consequences. First: if person detection fails, PPE detection never starts. If the person isn’t found, nobody asks about the hard hat. We covered the limits of person detection in the person detection with AI cameras article. Second: equipment occupies a very small area inside the person’s box. A hard hat is a small part of a person’s height; goggles are much smaller. Resolution sufficient to detect the person may not be sufficient to classify the equipment.
Detectable equipment table
Whether a piece of equipment can be distinguished visually depends on its size, how its color stands out from the background and where it is on the body. The table is general guidance; the definitive result comes from the classes the model supports and pilot measurement.
| Equipment | Visual detection | Decisive condition |
|---|---|---|
| Hard hat | Reliable under suitable conditions | Head visible from above or the side, color contrast |
| High-visibility vest | Reliable under suitable conditions | Torso visible, fluorescent color |
| Safety goggles | Limited | Close range, face turned toward the camera |
| Gloves | Limited | Hands visible, glove color different from skin |
| Safety shoes | Usually unreliable | Steel-toe safety shoes and ordinary shoes look visually similar |
| Ear protection (muffs) | Limited | Side view, enough pixels |
| Earplugs, mask filters | Practically undetectable | Very small or invisible |
The practical conclusion is clear: start the project with hard hats and vests. Goggles and gloves can be assessed only when the camera looks at the person up close and from a good angle (such as at a machine or a fixed workstation). Video analytics isn’t the right tool for safety shoes and earplugs; those are managed with entry control and site inspection. Keeping the scope narrow from the start saves the project.
Camera angle, distance and resolution
For PPE detection, camera geometry is planned differently from a general surveillance camera in enterprise camera systems. Assuming that existing cameras viewing a wide area will be enough is the most common mistake.
- Tilt angle: A camera looking almost straight down shows the top of the hard hat but hides the vest and face. A camera looking too horizontally hides people behind each other. To see the hard hat and vest together, a medium tilt, that is, an angle seeing the person from the front and above, is more suitable.
- Distance and pixel density: For equipment to be classified, the person must appear at sufficient height in the frame. The required value varies from model to model. Take the minimum person height (pixels) given by the manufacturer and read it as the lower limit for person detection, not for equipment; equipment classification needs more than this value.
- Analysis stream resolution: If analytics runs on the substream, the substream is what counts. A 4 MP main stream means nothing if analytics works with a substream 640 pixels wide. This is the first setting to check at installation.
- Frame design: Target bottlenecks rather than a wide hall: shift entrances, zone transitions, machine access points. Points where the person walks toward the camera and stays visible for a few seconds give the best results.
The relationship between model requirements and camera geometry is covered in more detail in the effect of camera angle and resolution on analytics accuracy guide.
Lighting and color factors
PPE classification relies heavily on color and contrast. Fluorescent yellow, orange-red and red vests stand out in daylight; the following conditions erase this advantage:
- IR night mode: When the camera switches to black-and-white, color information is lost. Reflective strips strongly reflect IR light and flare; this can make the vest visible but can also distort the person’s bounding box. In areas with a night shift, choose cameras with enough white light that keep a color image.
- Backlight: In areas such as doorways and docks, if the background is bright the person turns into a silhouette; if WDR is insufficient, equipment can’t be distinguished.
- Color similarity: A yellow vest or hard hat in front of yellow-painted machinery and railings may not stand out from the background. The organization’s PPE color policy (such as visitor and staff hard hat colors) should be consistent with the colors the model was trained on.
- Soiling and wear: Vests covered in dust, paint and oil lose their fluorescent properties over time. This isn’t an analytics error but a real visibility problem, and it should be treated as feedback to the OHS team.
High-visibility garment classes and colors are defined in the ISO 20471 standard. If your PPE policy runs on garments chosen according to this standard, use the same garments in model training and testing.
Alarm or report
The basic decision to make before installation is how the result will be used. The first idea is an instant alarm to the operator for every violation. In a busy production area, this method creates alarm fatigue within a few days; the operator starts closing notifications without reading them. A closed alarm is an alarm that doesn’t exist.
| Approach | When it fits | Risk |
|---|---|---|
| Instant alarm | High-risk, low-traffic point (e.g. entrance to dangerous machinery) | False alarm load |
| On-site warning (screen, light) | Bottleneck point, the moment the person can correct themselves | Maintenance needs of the warning device |
| Periodic report | Large area, trend tracking, training needs analysis | No instant intervention |
| Sampled review | Pilot period, model validation | Manual workload |
In most facilities, the sustainable model combines on-site instant warnings (such as a “put on your hard hat” screen at an entry point) with a report showing the distribution by area and time per shift or week. The report should show areas and time slots, not individuals; that’s how it becomes clear which zone needs training, signage or physical changes.
Integration with the occupational safety process
If the data analytics produces isn’t connected to the OHS process, it stays on the screen. Follow this order for integration:
- Define the scope with the OHS team: Which PPE is mandatory in which area is taken from the risk assessment, and the rule map is drawn accordingly.
- Define event classes: Events such as “no hard hat,” “no vest” and “visitor without PPE” are reported as separate classes.
- Plan evidence management: The short clip or frame of the event is deleted after the set retention period; only numerical data is kept for reporting.
- Set up transfer between systems: Events can be transferred from the VMS to OHS software, MES or a workflow tool via API or webhook.
- Close the feedback loop: Reports become input to the monthly OHS meeting, and the effect of measures taken is tracked in subsequent reports.
The current versions of Türkiye’s Regulation on the Use of Personal Protective Equipment in Workplaces and related regulations can be reviewed on mevzuat.gov.tr. Analytics doesn’t replace the employer’s obligation to provide PPE and supervise its use. It only supports supervision.
Pilot and validation
Performance on site can differ markedly from the manufacturer’s demo; demos are usually shot in good light, facing the camera, with a single person. That’s why you should plan a pilot of at least two to four weeks:
- Choose 2–4 representative cameras: one well-lit bottleneck, one wide area, one zone with a night shift.
- Run controlled tests: passes with and without hard hats, with and without vests, by staff of different heights and clothing; visitor hard hat color; bending, carrying loads and passing in groups.
- Take a sample of events produced during the passive period, review them manually, and count correct and false detections.
- Correct camera angles, zones and thresholds based on the results, and repeat the measurement.
The general framework for camera placement in a factory setting is in the factory camera system design guide. For the scope of PPE analytics and the pilot plan, see our PPE detection and occupational safety video analytics page, and for the general deployment approach, the AI video analytics deployment page.
Frequently asked questions
Are existing general surveillance cameras enough for PPE detection?
Often not. On cameras viewing a wide area, people appear at a small pixel size and equipment classification becomes unreliable. For PPE analytics, use cameras positioned at bottlenecks that see people up close and at a suitable angle.
Does PPE detection work on the night shift?
When the camera switches to IR mode and gives a black-and-white image, color information is lost and distinguishing vests and hard hats gets harder. Areas with a night shift need enough white light and cameras that can keep a color image; test performance separately in night conditions.
Which equipment can PPE analytics detect reliably?
At a suitable angle and in suitable lighting, hard hats and high-visibility vests are detected most reliably. Goggles, gloves and ear muffs can be assessed only at close range and with limited success; safety shoes and earplugs can’t be distinguished in practice by video.
Is it right to raise an alarm for every PPE violation?
In crowded areas, sending an alarm to the operator for every violation creates alarm fatigue. A more sustainable way is to combine instant on-site warnings at entry points with periodic reports by area and time.