Implementing ViT Models for Traffic Sign Detection in Autonomous Driving Systems

The advent of autonomous vehicles necessitates sophisticated visual recognition systems capable of interpreting traffic signs in real-time. Despite significant advancements, accurately detecting and recognizing traffic signs in diverse environmental conditions remains a challenge. Conventional methods, predominantly based on Convolutional Neural Networks (CNNs), have been integral in this domain but often struggle with complex scenarios and real-time processing demands. In response, this work proposes the utilization of Vision Transformers (ViT) as a novel approach to enhance detection and recognition performance. We leverage three distinct datasets: GTSRB, BelgiumTS, and a custom dataset curated under various weather conditions, totaling over 80,000 images across different classes. Our experimental results demonstrate a marked improvement in accuracy and inference speed, with ViT models outperforming traditional CNN architectures across standard benchmarks. Specifically, the ViT-Base model achieved an accuracy of 98.5% on the GTSRB dataset, showcasing its efficacy in real-world applications. The study concludes that ViT models, with their superior capability to process complex visual data, represent a promising direction for advancing the perceptual systems of autonomous vehicles, offering a viable solution to the real-time detection and recognition of traffic signs in varied conditions.