Indian Traffic Sign Classification Using Transfer Learning and Fine Tuning

The need for prompt and accurate detection of traffic signs has increased rapidly with the ever-increasing development of autonomous vehicles, driver assistance technologies, and advanced traffic monitoring systems. This study aims to evaluate the efficacy of four convolutional neural network architectures (VGG16, VGG19, ResNet50, and MobileNet) in the context of Indian Traffic Sign Recognition. The models employed transfer learning and fine-tuning to assimilate the distinctive features of Indian traffic signs, thereby reducing training duration and improving their capacity to generalize under challenging settings. All models were tested on the dataset according to their accuracy, loss, recall, and F1-score for recognizing traffic signs under various conditions. In this paper, we introduce a novel methodology to address sign classification in near real-time, enabling autonomous systems and drivers to make informed decisions on the road. This paper addresses the issues posed by Indian traffic signs and elucidates the regional applicability of these models, which may facilitate the advancement of intelligent transportation systems. The experimental findings demonstrated that the accuracies for the four models (VGG16, VGG19, ResNet50, and MobileNet) were 90.73%, 93.33%, 87.90%, and 83.33%, respectively.