🚦 Traffic Sign Classification (Mini Project - Y3S1)

Computer Vision / Python / Jupyter Notebook
Dataset: Chinese Traffic Sign Recognition Database (TSRD)
Platform: Google Colab • TensorFlow • Keras • OpenCV
Demo Included ✅

📘 Project Overview

This project develops a real-time traffic sign classification system for autonomous vehicles and Advanced Driver Assistance Systems (ADAS).
By leveraging ensemble deep learning, we ensure both high classification accuracy and lightweight performance, making it ideal for deployment on embedded systems.

🧠 Model Architecture

We use an ensemble of the following pretrained CNN models:

• ⚡ EfficientNetB0
• 📱 MobileNetV3Small
• ❌ Xception

These are combined via Global Max Pooling, followed by:

• Dense Layers with LeakyReLU
• Batch Normalization
• Dropout regularization

Base layers are frozen (transfer learning) to reduce training time.

✅ Results Summary

Metric	Our Model	Prior Work (ResNet50 + DenseNet121 + VGG16)
Accuracy	94.60%	96.41%
Model Size	107.37 MB	622.49 MB
Total Parameters	28.15 M	163.18 M
F1-Score	93.93%	96.16%
Precision	95.08%	96.16%
Recall	94.60%	96.16%

📉 80% reduction in model size with just ~1.8% accuracy drop.

🧪 Project Highlights

• ✅ Lightweight ensemble model using transfer learning.
• 🔄 Moderate oversampling for rare traffic signs (RandomOverSampler).
• ⏱️ Grid search for hyperparameter tuning (batch size, learning rate, optimizers).
• 📉 Dropout + L2 regularization to reduce overfitting.
• 📊 Evaluation with confusion matrices, learning curves, and visual analysis.

📷 Demo Included **

👥 Contributors

Thanks to the amazing team who made this possible:

• @chewliy
• @angqiaoyi
• @teohmingxue
• @wongsumhui