AI & Safety · Technical explainer
Computer vision for PPE detection: what a pilot should prove
Camera-based detection of hard hats, vests or exclusion-zone breaches is easy to demonstrate and hard to validate. This is what to measure before you trust it.
Object-detection models. The YOLO family is a widely used example. Can be trained to recognise people, helmets, high-visibility clothing or vehicles in video frames. Vendors increasingly offer this for PPE compliance, exclusion zones and pedestrian–vehicle proximity. A demonstration on a sunny test video can look impressive. The safety question is different: how does the system perform on your site, in your conditions, and what happens when it is wrong?
How these systems work, briefly
A detection model outputs boxes, class labels and confidence scores for each frame. A rule layer then turns detections into events: “person without helmet in zone B for more than five seconds”. Both layers matter. A good detector with a poorly designed rule can still flood supervisors with alerts or miss the event that matters.
Measure the right things
- Missed detections (false negatives): how often does the system fail to flag a real breach? For a safety use, this is usually the most important number.
- False alarms (false positives): how often does it flag something that is not a breach? High rates lead to alert fatigue and ignored alarms.
- Conditions coverage: night shifts, rain, dust, steam, glare, occlusion by equipment, unusual PPE colours, different body sizes and postures.
- Latency and availability: how quickly does an alert reach someone who can act, and what happens when a camera or network fails?
A pilot reports “95% accuracy”. On inspection, the test set is mostly daytime footage where nearly everyone is wearing a helmet. A model that always answers “helmet present” would also score highly. Ask for missed-detection and false-alarm rates on footage that includes the conditions and breaches you actually care about.
Safety of the AI itself
The NIST AI Risk Management Framework’s Map and Measure functions are a useful structure: establish the context and intended use, then test and track risks with quantitative and qualitative methods [1]. NIOSH researchers highlight privacy, autonomy and transparency as workplace ethical principles, including workers' right to be informed about how their personal data is used [2].
- Is the camera system a supplement to supervision and physical controls, or is it quietly replacing them?
- Who sees the footage, how long is it kept, and is it used for anything other than the stated safety purpose?
- Were workers and their representatives consulted before deployment?
Computer vision is best treated as an additional sensor that helps people notice patterns. For example, where pedestrian–vehicle interactions cluster, rather than as a control that makes a hazard safe. The physical separation of people and plant still does the protective work.
This explainer describes general evaluation principles. It does not assess any specific product. Data protection and workplace monitoring laws vary by country; take advice before deploying cameras for monitoring.
General educational material, not legal advice, a site-specific risk assessment or a work authorisation. Confirm current requirements in your jurisdiction and follow site procedures and competent advice.
References
- NIST AIRC. AI RMF Core (Govern, Map, Measure, Manage)United States
- NIOSH Science Blog. Exploring Approaches to Keep an AI-Enabled Workplace Safe for WorkersUnited States
All sources opened and checked on 30 October 2026. Requirements and guidance change; always consult the current version.
Editorial status: source-checked technical explainer. No independent external technical review is claimed. Spotted an error? See corrections.
