A Convolutional Neural Network Method For Traffic Sign Detection And Recognition

In the field of intelligent traffic system,traffic sign detection and recognition are two important research topics,which can assist the driver or unmanned driving system to master the road condition information and make clear the driving behavior rules,so as to effectively ensure the driving safety and reduce traffic accidents.However,the traditional detection methods based on image processing technologies have big limitations.It is difficult for them to adapt to the complex and changeable traffic scenes,and they cannot meet the actual needs.With the rise of deep learning,the traffic sign detection methods based on convolutional neural network have made great breakthroughs and become the mainstream research direction.However,it still faces many challenges.This paper analyzes and summarizes the difficulties in traffic sign detection and the shortcomings of the existing methods.In view of the problems of small size,multi-scale,similar appearance and performance-efficiency-tradeoff in traffic sign detection,a convolutional neural network detection method is designed in this paper to realize accurate and fast detection and recognition of traffic signs in large and complex traffic scenes.The main work of the paper is as follows(1)Since it is difficult to make a good tradeoff on performance and efficiency for a network that needs to complete complex tasks.Thus,this paper divides traffic sign detection into two more specific tasks: localization and classification,and uses two independent networks,localization network and classification network,to accomplish the two sub-tasks respectively.Experimental results show that although two separate networks are not end-to-end,they are more practical because they cost less time for training,and achieve faster inference and better performance.(2)In order to quickly locate traffic signs in high-resolution images,this paper adopts fire-modules to build a lightweight and efficient localization network to reduce the complexity of the network;A detection framework based on center point estimation is used to simplify the process of candidate region generation;In order to reduce the input scale of the network,the input image of the network is properly down-sampled.In addition,the structure of the fire-module has been improved in this paper,and multiscale information is extracted by fusing feature maps of different levels,so that the localization network can achieve high recall value and location accuracy for traffic signs of small size and multi-scale while improving efficiency.(3)Traffic signs of the same subclass are very similar in appearance.Thus,it is difficult to classify them accurately.However,through observation,the pattern of the central regions of the traffic signs are often different.Therefore,this paper adopted global pooling on the center regions of the traffic signs to extract local features with richer identification information,so as to improve the accuracy of the classification network.The experimental results show that the local features can effectively improve the classification performance of the network,and the global pooling can improve the classification accuracy better than the conventional convolutionFinally,this paper selected the large-scale traffic sign data set TT100 K to test and evaluate the proposed method.The experimental results show that the method proposed in this paper is superior to many existing state-of-the-art detection methods in performance and efficiency,and can significantly improve the detection performance of small-sized traffic signs.