SafeGuard-CV 🦺

Real-Time PPE Compliance Monitoring System using YOLOv8 for Construction Site Safety

SafeGuard-CV is a real-time computer vision system that detects missing PPE (helmets) on construction sites using a fine-tuned YOLOv8s model, optimized for edge deployment via ONNX FP16 quantization.

---

1. Problem Statement

Construction sites are high-risk environments where workers often skip wearing safety helmets. SafeGuard-CV automatically detects PPE violations in real-time using a camera feed, enabling instant safety alerts.

---

2. Project Structure

SafeGuard-CV/
├── safeguard-cv(finetuning).ipynb   # Model training notebook (Kaggle)
├── safeguard-cv(validation).ipynb   # Model validation notebook (Kaggle)
├── live_inference.py                # Real-time inference script
└── README.md

---

3. Dataset

• Source: Hardhat Detection by Michael — Roboflow Universe
• Images: 19,800
• Classes: Helmet, No-Helmet, Person
• Split: Train / Validation / Test

---

4. Project Pipeline

⚡ Phase 1: Training (FP32 Baseline)

• Model: YOLOv8s
• Epochs: 25
• Batch Size: 16

Baseline Validation Results (FP32)

Class	mAP50	mAP50-95
All	0.572	0.289
Helmet	0.665	0.334
No-Helmet	0.582	0.266
Person	0.471	0.267

⚡ Phase 2: Edge Conversion & Quantization

• Format: ONNX
• Quantization: FP16
• Command:

yolo export model=best.pt format=onnx half=True

📊 Performance Benchmark Table

Metric	FP32 (Base)	FP16 ONNX (Edge)	Difference
Model Size	21.5 MB	21.4 MB	-0.1 MB
mAP50	0.572	0.570	-0.002
mAP50-95	0.289	0.289	0
Accuracy Drop	-	0.02%	Negligible loss
Avg FPS (RTX 3050)	24.5 FPS	36.7 FPS	+12 FPS

⚡ Phase 3: Live Inference Script

The live_inference.py script loads the base model and edge model, and runs real-time inference on a video with the following overlays:

- Bounding boxes with class labels and confidence scores
- Live Inference FPS (excluding rendering time)
- Preprocessing latency (ms)
- Postprocessing / NMS latency (ms)

---

5. Downloads & Demo

Model Weights

Model	Format	Size	Link
best_fp32.pt	PyTorch FP32	21.5 MB	Download
best_fp16.onnx	ONNX FP16	21.4 MB	Download

Demo Video

🔗 Watch Demo on YouTube

---

6. Tech Stack

• Python 3.11.9
• YOLOv8s (Ultralytics)
• ONNX Runtime
• OpenCV
• Roboflow
• Kaggle (Tesla T4 GPU)

---

7. How to Run

# Install dependencies
pip install ultralytics opencv-python numpy

# Run with edge model 
python live_inference.py

---

8. Project Highlights

• Fine-tuned YOLOv8s on 19,800 construction-site images.
• Evaluated FP32 and FP16 ONNX deployments for edge inference.
• Built a real-time PPE compliance monitoring pipeline using OpenCV and ONNX Runtime.

---