01
The challenge
Production safety monitoring required reliable labels, difficult present-versus-absent PPE classification, and efficient inference on constrained edge hardware.
Production computer vision · Edge AI
Real-Time PPE Compliance Detection
An edge-deployed vision system that detects people, safety equipment, and compliant or non-compliant PPE states across ten classes.
95% fewer false positives
Implementation details presented without public source code.
02
Engineering approach
I combined YOLO11, YOLO-World, OWLv2, Weighted Box Fusion, supervised CLIP classifiers, leakage-safe multi-GPU training, and a custom evaluation harness before optimizing deployment with INT8 TensorRT.
03
Selected outcomes
- Raised annotation label F1 from 0.39 to 0.81.
- Reduced false positives from 9,490 to 619.
- Reached 99%, 97%, and 95% accuracy for mask, hairnet, and glove compliance states.
- Deployed on NVIDIA Jetson Orin NX with tracking and temporal consistency.
Next project