TRAFFIC SIGN RECOGNITION AND CLASSIFICATION USING DEEP CONVOLUTIONAL NEURAL NETWORKS

Abstract

Traffic sign recognition is a critical component of intelligent transportation systems and
autonomous vehicle navigation, enabling vehicles to interpret regulatory, warning, and
informational signage in real-time. Accurate detection and classification of traffic signs under
varying environmental conditions — including illumination changes, partial occlusion, weather
variations, and motion blur — remains a challenging problem in computer vision and machine
learning. This project presents a comprehensive traffic sign recognition system built upon
Convolutional Neural Network (CNN) architectures, specifically leveraging deep learning
techniques to achieve high classification accuracy across diverse traffic sign categories. The
proposed system utilises the German Traffic Sign Recognition Benchmark (GTSRB) dataset,
which comprises over forty-three distinct classes of traffic signs captured under real-world
conditions. The dataset includes more than fifty thousand images exhibiting substantial variation
in scale, rotation, lighting, and image quality, thereby providing a robust foundation for training
and evaluating classification models. A custom CNN architecture is designed and implemented,
incorporating convolutional layers for feature extraction, max-pooling layers for spatial down
sampling, dropout regularisation to prevent overfitting, and fully connected layers culminating in
a softmax activation function for multi-class classification. The methodology encompasses
comprehensive data preprocessing including image resizing, histogram equalisation for contrast
enhancement, and data augmentation techniques such as rotation, translation, and brightness
adjustment to improve model generalisation.