Research On Traffic Sign Recognition Algorithm Under Bad Weather Conditions Based On Deep Learning

In the field of intelligent transportation and self-driving,the environment is perceived through the camera around the driving vehicle,the traffic signs are recognized,and the vehicle responds to the recognition results in time.However,in the process of detecting and recognizing traffic signs in bad weather,the traditional target detection algorithm needs to use manual method of feature extraction to obtain the target candidate area.The amount of calculation data in the detection process is large and those methods are difficult to meet the requirements of high accuracy and real-time recognition during vehicle driving.This paper,which is based on the study of traffic sign detection and recognition technology at home and abroad,proposes a traffic sign detection and recognition method based on YOLOv4 network in bad weather conditions.The Chinese traffic sign data set(CCTSDB)of Changsha University of Technology is expanded by image flipping and random rotation to balance the difference in the number of samples in different categories.Imgaug library is used to enhance the image of the data set,and generate sample data under four adverse weather conditions: fog weather,rain and snow weather,strong light weather,dark light weather.The improved YOLOv4 network can correctly mark the location information of the bounding box,significantly improve the model detection speed,improve the recognition accuracy of remote small traffic signs under adverse weather conditions,and reduce irrelevant background interference.The main research work of this paper is as follows:1.Aiming at the problem of inaccurate boundary box position marking in the process of traffic sign recognition under adverse weather conditions,an improved K-medoids algorithm is proposed to re-cluster the traffic sign targets in CCTSDB data set.Nine new priori frames are obtained to replace the nine priori frames clustered by K-means algorithm in the original YOLOv4 algorithm.Then,the preset boundary box is more accessible to the size of traffic signs and detection rate of bounding box value is more higher.2.Aiming at the problem of low speed of traffic sign recognition under bad weather conditions,a deep separable convolution method is proposed to replace the standard convolution method in the YOLOv4 original trunk feature extraction network.This method can reduce the workload of network parameter calculation,reduce the calculation cost,and improve the speed of network detection.The traditional convolution operation process is divided into two different processes.Firstly,the characteristic map is divided into a corresponding number of single channel characteristic maps based on the number of channels.And these characteristic maps are convoluted respectively.Secondly,the convolution kernel of 1×1 is convoluted again to complete the further extraction of features.3.Aiming at the problem of low accuracy of remote small traffic sign target recognition under adverse weather conditions,an improved multi-scale feature fusion method is designed.The size of 26×26 of feature layer expands to four times of the original size by up sampling,and the size becomes 104×104.The feature layer has smaller receptive field,which is more suitable for detecting remote small traffic signs under adverse weather conditions.4.Aiming at the problem of background interference similar to the shape and size of remote small traffic signs under adverse weather conditions,this paper improves the ability of the network to distinguish correct semantic information.CBAM attention mechanism module is introduced after the three feature layers of the size of 13×13,26×26,104×104 to enhance the context interaction information of traffic signs and reduce the interference of irrelevant information to detection.5.According to the above research results,a traffic sign real-time detection system is designed.Pictures or video data containing traffic signs are uploaded,and the system can locate the location information of traffic signs in real time.The system can display the confidence of traffic sign categories and improve the efficiency of traffic sign detection.