Optimizing Traffic Sign Recognition for Malaysian Roads Using Deep Learning Models

This study evaluates models for traffic sign detection and classification. Initially, LeNet was trained from scratch on the GTSRB dataset, achieving 98.74% training accuracy and 86.54% test accuracy. In the second phase, pre-trained models like DenseNet201 and Xception were fine-tuned, with DenseNet201 achieving 100% test accuracy and Xception 95%. DenseNet201's performance was best with a batch size of 20, and the Adam optimizer yielded the highest performance. The fully trained DenseNet201 model was integrated with a laptop camera for real-time detection, classifying images within 2 seconds. The study highlights the potential of deep learning models for real-time traffic sign detection and addresses challenges like overfitting and dataset limitations.