Edge-Deployed AI Safety Compliance System with Real-Time Multi-Class PPE Detection, Face Recognition, and Automated Physical Access Control

Manual monitoring of personal protective equipment (PPE) compliance in industrial workplaces is labor-intensive, subject to supervision by humans and presents challenges for continuous safety enforcement. In this work, we propose a single embedded platform to integrate real-time multi-class personal protective equipment detection, worker identification, temporal compliance verification, and automated physical access control for an edge-deployed AI safety compliance system. It is based on Raspberry Pi 4 Model B (8 GB RAM) and does not need continuous cloud computing or external GPUs. YOLOv11-nano detector optimized with NCNN framework is used for the PPE detection in 12 classes, while YuNet-FaceNet pipeline implemented via ONNX Runtime is used for the worker identification based on face. The physical access-control interface is implemented with a GPIO-controlled relay, and the stability of compliance decisions before granting access is increased with a five-state finite state machine with a 15-frame majority-voting scheme. The detection model was trained on a combined dataset of >6,000 images of 12 PPE classes, including a subset collected specifically for earmuff and headset detection. The system achieved 0.843 mAP@0.5, 0.841 F1-score on the held-out test set, and field-tested response times of 312-321 ms, well below the 500 ms operational goal. The system can reach 2.3% false-alarm rate and 90% detection at 3 m under normal operating conditions. For adverse lighting, the performance drops to mAP@0.5 of 0.693. The results demonstrate that multi-class PPE detection, worker identification, temporal decision making, event logging and physical access control can be integrated on low-cost embedded hardware without the need for continuous internet connectivity or dedicated server infrastructure.