Traffic Sign Recognition Based On Improved YOLOv5

The widespread use of deep learning in image recognition provides an effective way for autonomous vehicles to obtain the instructions,warnings and prohibitions contained in traffic signs through image recognition algorithms,thus guiding vehicles to drive correctly,improving accessibility and reducing traffic accidents.Therefore,aiming at the problems of small size and poor real-time performance of traffic signs in practical scenes,this paper proposes a traffic sign recognition algorithm based on CBAM attention module,and improves the regression loss function and postprocessing mode of the network.The specific contents are as follows:(1)In order to enhance the robustness of the network,traffic sign images are obtained in the following three ways to establish the exclusive data set of this paper.①Some images in CCTSDB and TT100K data sets;②Real traffic sign images obtained by web crawlers;③Traffic sign pictures taken on site.Then the existing filtering denoising and image enhancement algorithms are studied and analyzed,and finally the improved median filtering and histogram equalization image enhancement algorithms are selected to preprocess the images.At the same time,mask mask algorithm is used to mask part of the images in the training data set to improve the sensitivity of the network to the blocked traffic signs,so as to improve the recognition ability of the network to the blocked traffic signs in the images.The final experiment shows that the data set constructed in this paper accords with the complexity of the actual scene.(2)An improved YOLOv5 algorithm is proposed to solve the problem that most of the traffic signs in the actual scene are small and dense,which leads to the low accuracy of identifying small and medium targets in the traffic sign recognition task.CBAM module is embedded in Backbone and Head of YOLOv5 network at the same time to improve the accuracy of small target recognition;In terms of regression Loss function,DIoU Loss is used to replace GIoU Loss by introducing diagonal distance and Euclidean distance to accelerate the model convergence speed.In terms of post-processing,Weighted Cluster NMS algorithm is used to replace NMS algorithm to improve model recognition speed.Finally,experiments are carried out on self-built data sets,and 96.40%mAP is obtained,which is 6.9%higher than the original YOLOv5 algorithm.The results show that the improved YOLOv5 algorithm can recognize small traffic signs more accurately,thus improving the overall recognition accuracy.(3)As traffic sign recognition technology will eventually be applied to driverless cars,the algorithm in this paper applies the improved algorithm in this paper to Jetson TX2 platform to recognize traffic signs in live video,verify whether it can run smoothly,and compare with other classical algorithms.Experiments show that the proposed algorithm can run smoothly in embedded system and obtain the optimal recognition accuracy.