Real-Time Traffic Sign Recognition Technology Based on Vision Transformer

Abstract

Real-time traffic sign recognition is an important component of intelligent transportation systems and autonomous vehicles, and it should be fast and reliable in adverse weather conditions. This paper introduces an improved visual transformer. This transformer addresses the limitations of local receptive fields in previous convolutional networks while handling complex urban traffic scenes. The framework introduces an adaptive spatial attention mechanism, two-stage decoding, and patch normalization for specific areas to improve the accuracy of recognition and classification. Improve the accuracy of recognizing and classifying similar traffic signs. In the experiment, the standard benchmark dataset collected over 50,000 images, covering various weather, lighting, and occlusion conditions. Under the same standard conditions, the proposed model achieved a Top-1 accuracy of 95.3% (±0.2%) and an average inference speed of 23 milliseconds per image. Better than well-known baseline models such as mixed architectures, EfficientNet-B3, and ResNet-50. In all major environmental categories, precision and recall are relatively stable, with an F1 score of approximately 0.951 ± 0.004. Ablation studies indicate that different parts of the aforementioned architecture have varying degrees of impact on them. In adverse weather conditions, adaptive attention and patch normalization also need to be used to perform well. This method is suitable for real-time use in intelligent road systems and high-end vehicles, with high recognition rates and wide applicability.