Real-time traffic sign detection based on multiscale attention and spatial information aggregator

Traffic sign detection, as an important part of intelligent driving, can effectively guide drivers to regulate driving and reduce the occurrence of traffic accidents. Currently, the deep learning-based detection methods have achieved very good performance. However, existing network models do not adequately consider the importance of lower-layer features for traffic sign detection. The lack of information on the lower-layer features is a major obstacle to the accurate detection of traffic signs. To solve the above problems, we propose a novel and efficient traffic sign detection method. First, we remove a prediction branch of the YOLOv3 network model to reduce the redundancy of the network model parameters and improve the real-time performance of detection. After that, we propose a multiscale attention feature module. This module fuses the feature information from different layers and refines the features to enhance the Feature Pyramid Network. In addition, we introduce a spatial information aggregator. This enables the spatial information of the lower-layer feature maps to be fused into the higher-layer feature maps. The robustness of our proposed method is further demonstrated by experiments on GTSDB, CCTSDB2021 and TT100k datasets. Specifically, the average execution time on CCTSDB2021 demonstrates the excellent real-time performance of our method. The experimental results show that the method has better accuracy than the original YOLOv3 and YOLOv5 network models.