Description of Invention
Traffic Sign Recognition (TSR) is essential for autonomous vehicles and intelligent transportation systems. This study compares deep learning-based TSR models trained on two datasets: Belgium Traffic Sign Classification (BTSC) with 62 classes, and Bangladesh Traffic Sign Recognition (BDTSR) with 48 classes. Using a DenseNet201 backbone pre-trained on ImageNet, we applied transfer learning with a custom classification head, data augmentation, and fine-tuning. The model achieved 95.34% validation accuracy on BTSC and 99.17% on BDTSR. The results demonstrate the effectiveness of transfer learning across regions, even with differing signage styles and environmental conditions. This highlights the importance of localized training data for building accurate TSR systems, helping improve road safety and autonomous navigation in diverse environments.