Research On Road Traffic Sign Recognition Algorithm Based On YOLO

In recent years,traffic sign recognition technology has played a crucial role in the fields of assisted driving and unmanned driving.Traffic signs are usually located in outdoor scenes,and due to natural factors such as light,rain,and wind,it is difficult for drivers or driverless systems to instantly distinguish categories.Obtaining traffic sign information in advance can not only remind drivers of the conditions of the road ahead,but also provide advance decision-making and judgment for unmanned vehicles,thereby reducing driver operational errors and erroneous judgments of intelligent vehicles,and reducing the probability of traffic accidents.In order to solve and effectively improve the low accuracy of traffic sign recognition in complex environments,multi-dimensional improvements and experiments are needed for traffic sign detection.At the same time,the recognition of traffic signs must take into account high-precision detection and real-time feedback.The main research work of this article is as follows:（1）Based on the Chinese traffic sign detection dataset CCTSDB,various improvements were made and the original sample dataset was expanded.Summarize the four types of traffic signs with the most categories from the CCTSDB dataset,including speed limits,prohibitions,instructions,and pedestrians;And collect representative traffic signs through the network,including traffic signs in complex scenes,such as rainy days,nights,strong lights,etc;The collected traffic sign data set is then subjected to image enhancement,mainly including image denoising,rotation,compression,and other methods to expand the data set.The final sample set is named the reproduced CTSDB traffic sign data set.（2）In view of the fact that most traffic signs are located in small,dense,and complex environments,resulting in the low accuracy of traffic sign recognition for small target detection.An improved S-YOLO traffic sign recognition algorithm based on YOLOv5 is proposed:a binary K-means clustering algorithm is used to cluster the reproduced dataset,generating a prior frame suitable for traffic sign targets,and improving the detection rate of the model on the prior frame;Embed the CBAM module into the YOLOv5 network,conduct end-to-end training simultaneously with the CNN network,pay more attention to the important parts of the feature map,and improve the recognition efficiency of small targets;Introducing a Bi-FPN module,using cross connections to eliminate points in PANet that have a low feature contribution,and adding jump connections to input and output nodes at the same scale to fuse more feature information,allowing the model to obtain better feature fusion capabilities;Using EIoU as a regression loss function optimizes the convergence speed of the model and improves the positioning accuracy of the target.（3）The above experiments were completed based on a deep learning approach.The proposed S-YOLO algorithm was trained and tested on a reproduced CTSDB traffic sign dataset,and data changes during the training process were recorded.The above four improved point combinations were subjected to ablation experiments.At the same time,the accuracy of the four types of traffic sign recognition was verified,and the S-YOLO model was compared with other algorithms,Finally,verify the actual detection effect of S-YOLO and visually compare it with the original YOLOv5.Experimental results show that the S-YOLO algorithm proposed in this paper has good detection performance on the reproduced CTSDB dataset,with an average recognition accuracy of 97.2%and a frame count of 72.3f·s^-1 per second.S-YOLO algorithm not only has high recognition accuracy,but also meets the requirements of real-time target detection.