Research on traffic sign detection based on SA-YOLOv5

In the object detection algorithm of autonomous driving, it is easy to miss detection and misdetection of small targets when recognizing distant traffic signs, which affects the judgment of on-board equipment on road conditions. Therefore, there are strict requirements for the accuracy of target detection algorithms. Aiming at the problem of low detection accuracy or even missed detection of small traffic signs, a new traffic sign detection method that Shuffle-Attention-YOLOv5 (SA-YOLOv5) based on Shuffle Attention Module (SA), Convolutional Block Attention Module (CBAM) and target detection algorithm YOLOv5 is proposed. This method integrates the SA module into the Backbone network of YOLOv5 to form a network for pixel-level information feature extraction, accurately extracts all relevant input features, and integrates CBAM into the Neck part of the network to better utilize the extracted data in Backbone. In the up-sampling and down-sampling, the attention area in the scene of small objects is found, so as to strengthen the feature extraction and recognition ability of small traffic signs in the distance. Training and model comparison experiments were conducted on the CSUST China Traffic Sign Detection Benchmark (CCTSDB), and deploy the model in an embedded system to verify the actual detection performance. Through screening, the mAP value when detecting tiny targets reaches 98.10%, compared with other The improved attention mechanism of YOLOv5, the mAP value is increased by 1.70%, which verifies that SA-YOLOv5 can effectively focus on the region of interest in the image, and has good performance in detecting small target scenes such as traffic signs.