A camera system that checks whether a hard hat is on a worker’s head, and not just somewhere in the frame, produced 29 false alarms on a benchmark where a standard detector produced 91. The paper describing it was published Oct. 5 in Scientific Reports by a team led from Chungbuk National University in South Korea, which reports an F1 score of 0.961.
That’s a real improvement on a real problem. It’s also a result from 305 test images, and the same model scored about 0.82 when it was pointed at two other helmet datasets.
Why helmet detection AI miscounts compliance
Most camera-based PPE monitoring runs an object detector that finds helmets and people separately. A helmet hanging from a belt or sitting on a tailgate can register as compliance. The authors add a second branch that estimates five head keypoints (nose, eyes and ears) and then checks whether the detected helmet sits where the head is. No match, no credit.
The model is built on a slimmed-down YOLO11 network so that it can run on the device at the camera. The team tested it on two NVIDIA Jetson edge computers and reports a processing speed of 110.7 frames per second. They also built a web dashboard that watches several video streams and logs violations automatically.
What the safety helmet compliance monitoring test covered
The benchmark is small. The dataset holds 932 images of construction scenes, pulled from public datasets and the web, with 2,747 labeled helmets and 3,428 labeled workers. The model trained on 627 of those images and was scored on the other 305. On that set it posted precision of 0.943 and recall of 0.980, and the authors put the drop in false compliance alarms at up to 67% against detection-only YOLO models.
Cross-dataset results are the ones to look at before a pilot. On two outside datasets the F1 score came in at 0.820 and 0.819. A system that’s right 96 times in 100 on familiar images and 82 in 100 on unfamiliar ones will behave more like the second number on a jobsite it has never seen, at least until it’s retrained on local footage.
And nothing here is a field trial. The paper measures whether software labels pictures correctly. It doesn’t measure head injuries, and it doesn’t report weeks of use on a live project with dust on the lens, low winter sun or a crew working with their backs to the camera. Occupational Health & Safety’s summary of the study describes it as a path toward scalable monitoring, which is fair. A path isn’t a deployment.
Jobsite questions an edge AI helmet monitor doesn’t answer
A method that keys on the nose, eyes and ears raises an obvious question about workers facing away, and a buyer should ask a vendor for those numbers specifically. Ask too what the system does with a helmet worn without its chin strap fastened. As head protection shifts toward Type II helmets, “on the head” and “worn correctly” stop being the same test.
Then there’s the crew. By one benchmark survey, contractors are piloting more automation than before, and continuous video scoring of individuals is the kind that gets noticed. A system that cuts false alarms by two-thirds still issues some, and every wrong flag lands on a named person.
The harder problem isn’t detection anyway. North Carolina State researchers are being funded to learn why crews resist better helmets in the first place. A camera can count who’s wearing one. It can’t keep the helmet on once the worker steps out of frame.