🚦 Indian Traffic Sign Classification

A deep learning-based project to classify Indian traffic signs using a fine-tuned ResNet50 model. This project leverages transfer learning, data augmentation, and a two-phase training process to achieve robust performance across 85 traffic sign classes.

📌 Project Overview

With the increasing deployment of autonomous driving and intelligent traffic systems, the ability to accurately identify traffic signs is crucial for safety and automation. This project focuses on the classification of Indian traffic signs using a convolutional neural network (CNN) architecture based on ResNet50.

The dataset is sourced from Hugging Face: kannanwisen/Indian-Traffic-Sign-Classification

---

🔍 Key Features

• ✅ Fine-tuned ResNet50 architecture for traffic sign classification
• 📚 Support for 85 unique Indian traffic sign classes
• ♻️ Advanced data augmentation for better generalization
• 🧐 Two-phase training: classifier training + full fine-tuning
• ⛑️ Predicts traffic sign class with confidence scores
• 📀 Weights saved as best_traffic_sign_model_resnet50_finetuned.pth

---

🧑‍🧬 Technologies & Libraries

• 🐍 Python
• 🧱 PyTorch
• 🔍 Torchvision
• ♻️ Albumentations (data augmentation)
• 📊 Matplotlib (for visualization)
• 🧦 ResNet50 (pre-trained model)

---

⚧ Challenges & Observations

• Some classes with visually similar signs (e.g., Pedestrian Crossing vs School Ahead) led to misclassifications.
• Performance drops with low-resolution or poorly lit images.
• Improvement areas include:
  • Better class separation
  • Training with additional real-world noisy data
  • Adding localization (bounding box) for real-time detection use cases

---

🧪 Example Predictions

Image	Predicted Class	Confidence	Notes
	STOP	0.3815	Correct, low confidence
	SCHOOL_AHEAD	0.6646	Misclassified
	U_TURN_PROHIBITED	0.4256	Correct
	NO_PARKING	0.5688	Correct
	SIDE_ROAD_LEFT	0.2342	Misclassified

---

💻 How to Run

1. Clone the repository:

   git clone https://github.com/pranavdhawale/itsd.git
   cd itsd

2. Install dependencies:

   pip install -r requirements.txt

3. Train the model (optional):

   python model_training.py

4. Run prediction:

   python predict.py path_to_image.jpg

---

📁 Project Structure

├── images
│   ├── no_parking.png
│   ├── parking.png
│   ├── pedestrian_crossing2.png
│   ├── pedestrian_crossing.png
│   ├── speed_breaker.png
│   ├── speed_limit_30.png
│   ├── speed_limit_80.png
│   └── traffic_signal_ahead.png
├── models
│   └── best_traffic_sign_model_resnet50_finetuned.pth
├── model_training.py
├── predict.py

---

🙌 Acknowledgements

Thanks to @kannanwisen for providing the Indian Traffic Sign dataset. Also, gratitude to my mentors and peers for their constant feedback and support.

---

📬 Let's Connect!

Feel free to connect on LinkedIn or contribute by creating a pull request or issue.

Contributors