🦺 YOLOv8n PPE Detection (6 Classes)
A lightweight and fast Personal Protective Equipment (PPE) detection model trained on 6 safety classes, optimized for edge devices such as NVIDIA Jetson Orin, ONNX Runtime, and TensorRT deployment.
1. 📌 Model Details
- Model name: YOLOv8n PPE Detection – 6 Classes
- Architecture: Ultralytics YOLOv8n (Nano)
- Task: Object Detection (PPE / Safety Gear)
- Framework: PyTorch (Ultralytics), exported to ONNX
- Input size:
1 × 3 × 640 × 640 - Output shape (ONNX):
1 × 10 × 8400- 4 channels → bounding box (x, y, w, h)
- 6 channels → class logits
- Number of classes:
6
2. 🏷️ Classes & Dataset Distribution
Class List with Image Counts
| ID | Class Name | Image Count |
|---|---|---|
| 0 | Gloves | 2,693 |
| 1 | Vest | 4,418 |
| 2 | goggles | 1,431 |
| 3 | helmet | 2,703 |
| 4 | mask | 2,763 |
| 5 | safety_shoe | 2,006 |
3. 🎯 Intended Use
Primary use-cases
- Real-time PPE compliance monitoring
- Construction site safety analytics
- Industrial / warehouse safety automation
- Worker equipment detection on:
- Helmets
- Safety shoes
- Reflective vests
- Gloves
- Masks
- Goggles
Target Platforms
- NVIDIA Jetson (Orin Nano, Orin NX, Xavier NX)
- ONNX Runtime
- TensorRT (FP32 / FP16 / INT8)
- NVIDIA DeepStream SDK
- Python (Ultralytics inference)
Not intended for
- Legal enforcement without human review
- Medical or clinical decisions
- Surveillance without consent
- Environments restricted by privacy regulations
⚠️ Human supervision is mandatory for high-risk applications.
4. 📂 Training Data
- Dataset source: Custom PPE dataset (Roboflow format)
- Annotation format: YOLO TXT
- Data types: Indoor/outdoor industrial environments
- Split:
- Train: majority
- Val: held-out validation
- Test: separate set for unbiased evaluation
Dataset is slightly imbalanced:
- Vest is most common
- safety_shoe & goggles have fewer samples
5. ⚙️ Training Procedure
- Base model:
yolov8n.pt(COCO pretrained) - Framework: Ultralytics YOLOv8
- Hyperparameters:
epochs = 50imgsz = 640batch = 32optimizer = auto
- Augmentations:
- Mosaic
- HSV color shift
- Random flip
- Scaling
- Translation
6. 📊 Evaluation & Metrics
Overall Performance
| Metric | Value |
|---|---|
| mAP@50 | ~0.81 |
| mAP@50–95 | ~0.53 |
| Precision | ~0.80 |
| Recall | ~0.74 |
Per-Class mAP@50
| Class | mAP@50 |
|---|---|
| Gloves | ~0.69 |
| Vest | ~0.90 |
| goggles | ~0.90 |
| helmet | ~0.90 |
| mask | ~0.80 |
| safety_shoe | ~0.64 |
7. ⚠️ Limitations & Ethical Considerations
Technical
- Sensitive to low-light & motion blur
- Small distant objects can reduce accuracy
- PPE variations (colors, styles, regional differences) may affect performance
Bias & Data Risks
- Dataset may not represent all ethnicities, industries, or regions globally
- Model may fail on unseen uniform types or PPE color variations
Responsible AI Guidance
- Always include human review
- Follow local privacy laws (GDPR, workplace regulations)
- Use only for transparent, ethical monitoring
8. 🚀 How to Use
A) Inference with Ultralytics (.pt)
from ultralytics import YOLO
model = YOLO("best.pt")
results = model("image.jpg", imgsz=640)
results[0].show() # visualize