Research On Traffic Sign Recognition Algorithm In Rain And Fog Based On Convolutional Neural Network

Traffic sign recognition is a key component of autonomous driving systems and driver assistance systems,and has long been a hot topic in the field of machine vision.The accuracy of traffic sign recognition directly affects the safe driving of vehicles.However,road traffic sign images captured in rainy or foggy weather often lack bright colors,have a grayish tone,and low clarity,which cause the loss of image details and the blurry or hazy appearances of traffic signs.Subsequently,it is difficult to recognize the exact traffic signs.In addition,visual-based road detection tasks for autonomous vehicles have high computational demands,and the real-time performance of most current algorithms still needs to be improved.Therefore,this thesis focuses on the accuracy and real-time performance of traffic sign recognition algorithms,and researches the use of image enhancement algorithms and convolutional neural networks for traffic sign recognition in rainy and foggy weather conditions.The main contributions of this thesis are:Firstly,in the field of image enhancement,an analysis is conducted on the reasons for the degradation of rainy and foggy day images with the attenuation model of incident light and the atmospheric imaging model.A DCP-MSR image enhancement algorithm is proposed to address the problems of edge blurring and darkening of foggy day images caused by the dark channel prior algorithm.The rainy lines are denoised by K-means clustering and median filtering,and then the transmission rate is optimized and the image edges are refined by using guided filtering.Finally,the preliminary restored image is converted to the HSV color space and enhanced with a multi-scale Retinex algorithm to improve image contrast and brightness.The algorithm’s image enhancement effect is evaluated through experiments.Then,the convolutional neural network and object detection methods are studied.Starting from the structure of the convolutional neural network,the advantages and disadvantages of current two-stage and one-stage object detection algorithms are analyzed.To address the recognition of traffic signs in high-definition and complex street view images,an improved YOLOX traffic sign recognition algorithm GC-YOLOX is proposed.In the feature extraction stage,the Ghost Net lightweight network model is introduced to speed inference.Then,in the feature enhancement layer,multi-scale feature fusion and the CA attention mechanism are used to accelerate the network’s perception ability for small targets of traffic signs.The algorithm’s inference speed and accuracy are validated.Finally,a traffic sign recognition system is designed.A hardware platform is built and a traffic sign recognition interface is designed with Py Qt under the Pycharm development environment.First,the simulation experiment is carried out indoors to verify the feasibility of the traffic sign recognition system.Then,traffic signs are recognized in outdoor scenes to demonstrate the effectiveness of the system.