AI for Occupational Safety: Use Cases and Limits

AI in occupational health and safety: what cameras see and where they fail in PPE, falls, forklift–pedestrian, restricted zones, fire and ergonomics, and how to deploy under Law 6331 and KVKK.

Translated from the Turkish original · Türkçe aslı

AI for occupational safety is an inspection layer that analyses the footage of existing cameras (by default on a server inside the facility), reports missing PPE, entry into restricted zones, forklift–pedestrian proximity, falls, risky postures and smoke the moment they occur, and turns these events into countable data. It does not replace the OHS specialist, the risk assessment or engineering controls; it covers the hours when the inspector is not on site.

The logic is the same in every application. In each frame the model finds the person, vehicle, equipment or flame; the rule engine compares this with the condition you define (hard hat mandatory in this zone, at least 3 m between pedestrian and forklift in this aisle). If the condition is confirmed over consecutive frames, a notification is sent with a short clip and a confidence score. Below we go through the application areas relevant to OHS one by one, with what each does well and where it struggles.

How does PPE detection work?

The model first finds the person, then classifies the hard hat, high-visibility vest, gloves and glasses in the head, torso, hand and face regions. In our PPE detection solution, hard hat detection accuracy measured on site is 90% and the time from violation to notification is under 2 seconds. The rule is tied to the zone: glasses are required in the filling area, a vest on the ramp, gloves on the sheet-metal cutting line; when the same person moves into the office corridor, the rule changes. Hard hats and vests are large, colourful objects and easy to detect. Gloves, glasses and ear protection are small; as the person moves away from the camera the pixel count drops and the confidence score falls. For small PPE, cameras are therefore positioned at points where people pass close by, such as entrance doors or workstations. Non-compliance is sent to the supervisor as a notification with footage; the report shows the zone, time and rule rather than the person's name.

Can unsafe behaviour and ergonomics be measured with a camera?

They can, but object detection is not enough for this; a skeleton is needed. Safe behaviour and pose analysis represents the body with 17 keypoints, runs at 25 fps and distinguishes five defined risky behaviour classes such as reaching into a machine or climbing a guardrail. Trunk, neck and arm angles are calculated from the same skeleton data; these angles are the input to REBA and RULA scoring, widely used in ergonomics. A manual REBA/RULA assessment scores a single moment; the camera measures the same workstation throughout the shift, counting how long a risky posture lasts and how many times it is repeated. Load weight and coupling quality cannot be read from the image and are entered manually per workstation. In pose analysis the camera angle is decisive: a side or oblique angle is needed, as a camera looking down from above cannot see forward bending.

How are falls and immobility distinguished?

There are two separate time periods here, and they should not be confused. Fall detection detects a sudden loss of height and transition to a horizontal position in under 5 seconds; this is a candidate event. Then the immobility window begins: if the person does not get up or move for the default 30 seconds, the event is escalated to emergency level and the location and live footage are sent to the response team. This way, an employee crouching to pick up a part or reaching under a workbench does not trigger an alarm. The window is adjusted to the site's risk; in confined spaces such as tanks, holds or pits, where the risk of fainting is high, it is kept shorter. In a confined space the real question often comes before the fall: how many people are inside, and has everyone who went in come out? This question is answered with entry–exit counting, which is set up together with fall detection. An additional advantage for lone workers is that the person does not need to carry a device; a panic button is useless if the person loses consciousness.

How are forklift–pedestrian collisions prevented?

The camera tracks forklifts, pallet trucks and pedestrians as separate classes and calculates the distance between them and the vehicle's speed in every frame. In the vehicle and equipment safety solution, vehicle and equipment safety accuracy is 96%; the safe distance can be defined as 3 m and changed to suit the site. When the threshold is exceeded, a light or audible warning is triggered on site and a notification with footage goes to the supervisor. A speed violation is written as a separate rule; blind corners, pedestrian crossings and ramp entrances are marked as separate zones. At the pilot site on our solution page, the number of near contacts fell by 55%; this is the result of a single site, and the effect in your own facility is seen through before–after measurement. Having the camera stop the forklift is a separate engineering job and does not replace the safety systems on the vehicle.

How are restricted zones and machine safety areas monitored?

Three types of rules are written for each polygon drawn on the camera image: no entry, staying longer than a set time, and more than a set number of people at once. Zone violation detection works with these rules at 93% accuracy. The rule can be tied to machine state; entering the safety area while the press is running is a violation, but in maintenance mode it is not. Article 4 of Law No. 6331 (Türkiye's Occupational Health and Safety Law) obliges the employer to take the necessary measures so that employees who have not been given adequate information and instructions do not enter places with vital and special hazards; the camera monitors whether this measure is applied on site. The limit also needs to be stated clearly: unless video analytics is designed and validated as a machine safety function, it does not replace a light curtain, safety interlock or guarded door. It works behind them as a second pair of eyes that makes a bypassed or temporarily disabled guard visible.

What does it do in fire, smoke and hazardous materials areas?

Fire and smoke detection looks for the visual signature of flame and smoke in the image; accuracy measured on site is 95%, and the time from detection to notification is under 10 seconds. In a high-ceilinged warehouse or an open area, smoke takes time to reach the detector; the camera shortens this delay. It does not replace the fire detection system installed as required by regulations and insurance; it is an early-warning layer with footage added in front of it. Welding light, steam and vehicle headlights are the most frequent sources of false alarms; known hot spots are monitored with a separate threshold.

In a hazardous materials store, the question comes before the flame: is the flammable material in its defined zone, has the stacking height been exceeded, is the escape route clear, is the person entering the zone wearing the required equipment? Warehouse and hazmat monitoring checks these four rule sets (area, height, route, equipment) 24/7. The camera does not know what is inside a drum; it sees its position and presence, and if the material class is needed, it is matched with the warehouse management system. Priorities change by sector: in shipyards hot work, confined spaces and working at height come to the fore, while in chemical plants PPE in the filling area, entry to the tank farm and ATEX zone rules take priority.

How does near-miss data turn into prediction?

Article 14 of Law No. 6331 obliges the employer to investigate and report events that, even without causing injury, have the potential to harm employees or equipment. In practice, near-miss records remain incomplete because they depend on employees reporting them: when a forklift speeds past a pedestrian, usually nobody fills in a form. The camera records these events automatically, with type, zone, time, confidence score and a short clip. When the records accumulated over months are combined with the shift plan, shipment intensity and staff turnover, a risk score can be calculated by zone and by day; we explain the method in our predictive safety article. For these records to be useful in the OHS committee, each event type must have an owner and a closure record; notifications that are never closed soon become a list nobody looks at.

Where does AI fit under Law 6331 and KVKK?

In addition to taking measures, Article 4 of Law No. 6331 gives the employer the duty to monitor and inspect whether the measures taken are complied with and to eliminate non-conformities. Video analytics supports this monitoring and inspection part; it does not take the measure itself. Under the same article, employees' obligations do not affect the employer's responsibility, which means violation records cannot be used as a tool to shift responsibility onto the employee. Article 5 of the Law gives collective protection measures priority over personal protection; PPE detection is therefore positioned as a tool that checks the last layer alongside engineering and organisational measures. Article 18 requires employees or their representatives to be consulted on the introduction of new technologies; talking with the OHS committee and the employee representative before installation is both a legal and a practical step.

On the KVKK (Türkiye's Personal Data Protection Law) side, the key balance is proportionality. Video processing is personal data processing; the purpose must be clearly defined as OHS, recording must be limited to that purpose and employees must be informed. Event-based clips instead of continuous recording, anonymous person detection instead of face recognition, and zone and shift reports instead of individual reports are design choices that make this balance easier to achieve. We cover the detailed framework in our workplace camera recording and KVKK article. The information here is general in nature and does not constitute legal advice.

Where does AI get it wrong?

Every model makes two kinds of error: reporting an event that did not happen (false alarm) and missing an event that did. The two are linked through the same threshold; lowering the threshold reduces misses and increases false alarms. The balance is set on site by manually marking real events and comparing them with the system's output. The second limit is the field of view: the camera only evaluates what it sees; the back of a rack, the shadow of a stack, a column or two workers overlapping create blind spots. The third is light. In poorly lit areas, infrared night footage loses colour information; a rule that relies on vest colour weakens under this condition, and smoke is harder to see in the dark. Dust, steam, dirt on the lens and racks that change position also reduce accuracy over time; periodic camera checks should therefore be part of the maintenance plan. Finally, the model does not know the context: it can only know whether a person in a restricted zone has a work permit if it is integrated with the permit system.

Where to start?

The starting point is the risk assessment, not the camera inventory. The pilot is set up in this order:

1. Two or three scenarios with the highest scores in the risk assessment are selected; for example, forklift–pedestrian proximity and the ramp in a warehouse-heavy facility, or the safety area and reaching into the machine on a press line.
2. For each scenario, the event definition, to whom and within what time the notification will go, and the success criterion are written down before the pilot starts.
3. A 30-day pilot is run on a few cameras that see these scenarios; the first weeks are for threshold tuning and the rest for measurement.
4. At the end of the pilot, a table of events caught, missed and marked as false alarms is produced for each scenario; the rollout decision is made on the basis of this table.

We describe the pilot set-up in detail in our a pilot you can decide on in 30 days article.

At CX Teknoloji, we analyse existing IP cameras on an edge server inside the facility by default; we run PPE, pose and ergonomics, falls, forklift–pedestrian, zone violation, fire and warehouse rules on the same infrastructure, and deliver events via the web dashboard, email, WhatsApp or an alarm relay. We do not use face recognition, and we keep event-based records. During the site survey we evaluate your camera inventory scenario by scenario and report which point can carry which analytics; for a pilot with your two highest-risk scenarios or a live demo, you can reach us via our contact page.

Frequently asked questions

Will AI replace the occupational safety specialist?

No. An AI-assisted system monitors throughout the shift whether the rules defined by the specialist are applied on site and records events. Carrying out the risk assessment, choosing the measure, training employees and interpreting events remain the specialist's job. Under Law No. 6331 these obligations rest with the employer; camera data helps these tasks be done through measurement rather than estimation.

Are existing security cameras sufficient for occupational safety analytics?

Usually they are; what matters is angle, resolution and lighting. ONVIF-compatible IP cameras that provide an RTSP stream generally handle PPE, zone violation and forklift–pedestrian scenarios. Pose and ergonomics analysis needs a side or oblique angle; for small PPE such as gloves and glasses, the person needs to pass close to the camera. During the site survey each camera is evaluated per scenario, and an angle change or an additional camera is recommended for unsuitable points.

Does an AI-assisted occupational safety system produce false alarms?

It does; no model works with zero false alarms, what matters is measuring and managing the rate. An event is reported only after being confirmed over consecutive frames, every notification comes with a confidence score, and the threshold is adjusted to the site. Throughout the pilot, real events are marked manually and compared with the system's output; the numbers of false alarms, missed events and correct detections are tabulated per scenario, and thresholds are updated based on this table.

Is monitoring employees with cameras compliant with KVKK?

It is possible when set up properly; what matters is purpose, proportionality and informing employees. Clearly defining the purpose as OHS, informing employees, keeping event-based records, limiting access by role and avoiding processing that generates biometric data, such as face recognition, make this balance easier to achieve. Each installation should be assessed under its own conditions; this information is general in nature and does not constitute legal advice.

Does the system work on night shifts and in dark areas?

It works depending on the lighting, but performance varies. Cameras with infrared night vision detect people and vehicles in the dark as well; because the image loses colour information, rules that rely on vest colour weaken and smoke is harder to make out. At critical points, additional lighting or a thermal camera is considered. Measuring night and day performance separately during the pilot allows thresholds to be adjusted per shift.

Can the camera stop a machine or a forklift?

Through relay or PLC integration it can generate a warning light, siren or stop signal; however, unless this signal is designed and validated as a machine safety function, it should not replace a light curtain, safety interlock or on-vehicle safety systems. In practice, the camera is an additional layer that works behind these guards and makes a bypassed guard visible. How the signal meets the existing safety system is designed separately and tested on site.

Which scenario should AI for occupational safety start with?

Start with two or three scenarios that score highest in the risk assessment and that the camera can clearly see. In warehouse-heavy facilities this is often forklift–pedestrian proximity and restricted zones; on production lines, the machine safety area and PPE; in areas with lone working, fall detection. The selected scenarios are measured in a 30-day pilot against criteria written in advance, and the rollout decision is based on this measurement.