Learning multi-layer interactive residual feature fusion network for real-time traffic sign detection with stage routing attention

Traffic sign detection is an important research content of Autonomous Driving Systems, which can effectively guide vehicles or driver to make correct decisions and reduce traffic accidents. The existing real-time traffic sign detectors have low detection accuracy for small objects. Therefore, we propose a novel real-time traffic sign detector based on YOLOv5 for the accurate detection of small objects. Specifically, we propose a new Multi-layer Interactive Residual Feature Fusion Network (MIRFFN) in the neck, which can effectively combine the position information of the low-layer feature maps with the semantic information of the high-layer feature maps, and refine the features by fusing different layers of feature maps. Then, we design a Residual Information Fusion (RIF) module for MIRFFN to fuse feature maps from different layers. The RIF module is composed of three residual blocks to refine spatial position information. Inspired by Bi-level Routing Attention effectively extracting small objects, we design a Stage Routing Attention (SRA) module in the backbone. The SRA modules can search the most relevant regions and enhance attention to small traffic signs in high-layer feature maps. We conduct experiments on GTSDB, TT100K, and CCTSDB2021, and achieve mAP of 96.2%, 72.7%, and 90.8%, respectively. Our method achieves 48.91 FPS on the GTSDB dataset. The experimental results show that our method can accurately perform real-time traffic sign detection.