🚦 Real-Time Traffic Signs Classification using CNN & OpenCV

This project implements a real-time traffic sign classification system using a Convolutional Neural Network (CNN) built with TensorFlow/Keras, and OpenCV for video capture and visualization. The model is trained on image data of traffic signs and deployed for real-time camera-based inference.

---

📌 Features

• ✅ Real-time traffic sign recognition from webcam video.
• ✅ Custom-trained CNN with 30 epochs for improved accuracy.
• ✅ Image preprocessing (grayscale + histogram equalization).
• ✅ Live predictions with confidence score overlay.
• ✅ Lightweight and fast — no object detection dependency.

---

🧠 Model Overview

• Input shape: 32x32x1 (grayscale)
• Architecture:
  • 4 Convolutional layers (with ReLU)
  • MaxPooling layers
  • Dropout layers to prevent overfitting
  • Dense layer (500 nodes)
  • Output layer with softmax for multi-class classification
• Loss: Categorical Crossentropy
• Optimizer: Adam
• Epochs: 30
• Batch size: 50

---

🧑‍💻 Development & Execution Environment

• 📌 Training was done on Google Colab using the notebook Training_code_epoch30.ipynb.
• 📦 After training, the model was saved as model_trained_epoch30.p using pickle and downloaded to the local machine.
• 🖥️ Testing / inference was performed locally using PyCharm, with the script Test Code.py running in real time via webcam.

---

📂 Project Structure

├── myData/                      # Training images organized in subfolders by class
├── labels.csv                   # CSV mapping class numbers to class names
├── Training_code_epoch30.ipynb  # Jupyter Notebook for training (Google Colab)
├── model_trained_epoch30.p      # Trained model (saved with pickle)
├── Test Code.py                 # Python script for real-time camera inference (run on laptop)
├── demo_ts.mp4                  # Demo video showing classification in action
└── README.md

---

🏋️‍♂️ Training the Model

Run the Jupyter notebook:

Training_code_epoch30.ipynb

Main steps:

1. Load and preprocess all images from myData/.
2. Normalize and equalize histogram.
3. One-hot encode the labels.
4. Augment data with ImageDataGenerator.
5. Train the CNN for 30 epochs on Google Colab.
6. Save the trained model with pickle as model_trained_epoch30.p.
7. Download the .p file to your laptop for testing.

---

🎥 Real-Time Inference

Run the test script on your laptop using PyCharm:

python "Test Code.py"

What it does:

• Opens webcam feed.
• Captures frame-by-frame.
• Resizes and preprocesses each frame.
• Predicts the traffic sign class.
• Overlays:
  • 🚫 CLASS: predicted class name
  • 📊 PROBABILITY: confidence percentage
• Press Q to quit.

---

---

📈 Model Performance

After training for 30 epochs on the traffic sign dataset:

• ✅ Training Accuracy: ~98.5%
• 🧪 Validation Accuracy: ~96.8%
• 🧾 Test Accuracy: ~96.2%

These results show the model is well-generalized and performs reliably on unseen data.

📊 Demo Results

Model running in real-time with sign classification overlay.

You can also watch the full demo in demo_ts.mp4.

---

🔧 Requirements

• Python 3.x
• TensorFlow / Keras
• OpenCV (cv2)
• NumPy, Pandas, Matplotlib
• Pickle

Install dependencies:

pip install tensorflow opencv-python numpy pandas matplotlib

---

📌 Notes

• This is a classification-only model (not detection). For object detection, you can integrate with YOLO or SSD.
• The model performs best under good lighting and with the camera pointed directly at signs.