🪖 Helmet Detection System using Faster R-CNN

📌 Project Overview

This project is an AI-powered Helmet Detection System built using PyTorch (Faster R-CNN) and deployed via a Flask web application. It detects whether a person is wearing a helmet in:

• 📷 Uploaded images
• 🎥 Live webcam stream

The system identifies:

• Person
• Head
• Helmet

Then applies geometric logic to determine whether a helmet is properly worn.

---

🚀 Features

• ✅ Real-time helmet detection using webcam
• ✅ Image upload inference
• ✅ High-accuracy detection using Faster R-CNN
• ✅ Custom logic for helmet-head matching
• ✅ GPU acceleration (CUDA + AMP support)
• ✅ Adjustable thresholds via environment variables

---

🧠 Technologies Used

🔹 Backend

• Python 3.11
• Flask (Web server)
• OpenCV (Image processing)
• NumPy (Numerical operations)

🔹 Deep Learning

• PyTorch
• Torchvision (Faster R-CNN ResNet50 FPN)

🔹 Model

• Faster R-CNN (Object Detection)
• Custom-trained on helmet dataset

---

⚙️ Working Principle

Step 1: Input

• Image uploaded OR webcam frame captured

Step 2: Preprocessing

• Image is resized using letterbox technique
• Converted to tensor and normalized

Step 3: Model Inference

• Faster R-CNN predicts:
  • Bounding boxes
  • Class labels
  • Confidence scores

Step 4: Filtering

• Apply class-specific thresholds:
  • Person → low threshold
  • Head → medium threshold
  • Helmet → high threshold

Step 5: Helmet Detection Logic

For each detected head:

• Check overlap with helmet
• Ensure helmet is:
  • On top region of head
  • Proper size relative to head
  • Sufficient intersection area

Step 6: Output

• 🟢 Green box → Helmet worn
• 🔴 Red box → No helmet

---

🛠️ Installation Guide

1️⃣ Clone Repository

git clone https://github.com/your-username/helmet-detection.git
cd helmet-detection

---

2️⃣ Create Virtual Environment (Recommended)

python -m venv venv
venv\Scripts\activate   # Windows

---

3️⃣ Install Dependencies

pip install -r requirements.txt

---

⚠️ Important Note (PyTorch Installation)

• If you have a dedicated GPU (NVIDIA dGPU), install the CUDA-enabled version of PyTorch for much faster training and inference:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

• If you do not have a GPU, install the regular CPU version:

pip install torch torchvision

👉 You can verify GPU support using:

import torch
print(torch.cuda.is_available())

---

4️⃣ Verify GPU (Optional)

import torch
print(torch.cuda.is_available())

---

▶️ Running the Application

Step 1: Place Model

Put your trained model inside:

models/
 ├── best_model.pth
 OR
 ├── final_model.pth
 OR
 ├── fasterrcnn_scripted.pt

You can donload the model from Here

---

Step 2: Run Flask App

flask run

---

Step 3: Open in Browser

http://localhost:5000

---

🎥 Usage

📷 Upload Image

• Click upload button
• Select image
• View detection results

🎥 Webcam Detection

• Open live feed
• Real-time helmet detection

---

⚙️ Environment Variables (Optional)

You can customize behavior:

set THR_HEAD=0.3
set THR_HELMET=0.75
set IMG_SIZE_CAM=320
set IMG_SIZE_UPLOAD=640

---

🚀 Performance

• ⚡ ~6 minutes per training epoch (RTX 4060)
• 🎯 High accuracy with custom dataset (~5000 images)
• 🔥 Optimized with mixed precision (AMP)

---

🧩 Model Loading Priority

The app automatically loads:

1. best_model.pth
2. final_model.pth
3. fasterrcnn_scripted.pt

---

❗ Troubleshooting

❌ Model not loading

• Ensure file exists in /models

❌ Webcam not working

• Change camera index:

CAM_INDEX = 1

❌ CUDA not used

• Ensure PyTorch GPU version installed

---

🔮 Future Improvements

• 📱 Mobile app integration
• ☁️ Cloud deployment
• 📊 Dashboard analytics
• 🎯 YOLOv8 integration for faster inference

---

👨‍💻 Author

Developed as part of an AI/ML project for helmet safety detection.

---

⭐ If you like this project

Give it a ⭐ on GitHub!