01Choose the people and area
Enable the camera and create a hardhat rule with a zone where heads are clear. The rule requests person detection automatically; its processor needs a ready PPE model.
An ordinary YOLO person label does not tell you whether a helmet is worn. PPE performs an additional check on the detected person crop, so it uses additional processing.
02Understand the three possible outcomes
Compliant means the model detected a hardhat. Violation requires an explicit no-hardhat detection. Unknown includes a person that is too small, no reliable PPE detection, conflicting results or an unavailable model.
Unknown is not treated as missing a helmet. A blocked, dark or tiny head can therefore produce no violation even when a person is detected.
03Confirm over several frames
The rule uses a grace period and a rolling observation window per tracked person. Default settings are a 2-second grace period, a 5-second window, at least 5 valid frames, at least 3 violation frames and a 60% violation ratio.
The ratio uses valid PPE observations; unknown frames are not evidence of a violation. PPE confidence is a separate setting from person-detection confidence. Saved custom settings remain in effect.
04Enable delivery and test locally
Choose recipients in the rule and enable both the rule and notifications. Telegram is recommended for setup and testing. Check person boxes, PPE status and the resulting event.
Test with a safely positioned person, with and without a helmet where permitted. Improve camera framing and lighting before lowering thresholds. Do not ask anyone to remove required protection in a hazardous area.
Common questions
Why are people detected but no helmet alert appears?
Check the zone, PPE model status, head visibility, grace/window settings and explicit no-hardhat evidence. Unknown is deliberately not a violation.
Can I apply one helmet rule to multiple cameras?
Yes. Each camera needs person detection and its own zone. Tracking, evidence and cooldown remain independent.