Abstract:
Bangladeshi road traffic sign detection is a must for enhancing road safety and aiding
autonomous driving systems by accurately identifying and interpreting local traffic signs.
This work focuses on the development of deep learning methods for Bangladeshi road
traffic sign identification and Support, so as to improve the safety of roads and better
organize traffic flow. The dataset of raw images selected is 2710 which has been carefully
augmented to 5420 images with specific traffic signs existing in Bangladesh and distributed
over 18 categories. The study aims to discuss and compare the results of four Transfer
Learning, namely Xception, VGG19, InceptionResNetV2, and MobileNetV2 with a novel
CNN architecture proposed for use. Qualitative results reveal that Proposed CNN has the
maximum accuracy of 99.54%, surpassing other architectures. Most of the features
developed will be on how to pre-process data such as normalization and data augmentation
to enhance the quality of the set and models. Evaluation parameters such as accuracy,
precision, recall, and F1-score show that deep learning approaches give a reliable technique
for the classification of traffic signs irrespective of the background context. The studies
support the idea of deep learning in playing a central role to re-define the intelligent
transportation systems through real-time sign recognition and improving the driver safety
and traffic management in both complex and simple urban and rural geography of
Bangladesh.
