Research On Traffic Sign Detection And Recognition Under Complex Background

The detection and recognition of traffic signs has been a hot and difficult issue in the field of intelligent driving.It provides an important guidance for the intelligent driving system,which can better assist the driver to understand the road condition information ahead and guide the driver driving safely and reduce the occurrence of traffic accidents.However,the real road condition is so complex,a series of factors such as rain,snow,fog and haze weather,light intensity and shade of trees have brought difficulties to the detection and identification of traffic signs.Facing above disadvantages,this paper adopts machine-learning and convolutional neural network algorithms to detect and identify traffic signs in a complex road.This paper is divided into two stages: traffic sign detection and recognition.Detection refers to the area containing traffic signs detected from the original picture,and recognition is to identify the class of detected traffic signs.The main work of this paper is as follows:(1)In traffic sign coarse detection phase,in order to get a better robustness to the light,adopting the color of probability model which based on YCb Cr space,and then propose a method based on improved color enhancement of Maximum Stable Extremal Region(MSER)algorithm to detect the candidate area.Then combine two algorithms,reducing the interference areas.Finally,the Region Of Interest(ROI)was obtained by filtering geometric feature.Experiments show that the method in this paper is still reliable under complex conditions,and compared with other algorithms,the improved method has more advantages in reducing interference areas,increasing the accuracy of classification,and reducing the workload for subsequent feature extraction and classification.(2)As for ROI's feature extraction,this paper adopts the PHOG feature which has strong spatial expression ability and combined with the Support Vector Machine(SVM)classifier for classifying.In order to find the optimal parameter pair(C,g)of SVM based on Radial Basis Function(RBF),using the v-fold cross-validation method which based on the optimization of grid parameters.Then the red,yellow and blue signs and the background samples were trained,and the trained model was used to detect and classify the test images in the Chinese Traffic Sign Dataset(CTSD),obtaining a good classification accuracy.Compared with the mainstream algorithms,experimental results validate the advantages of proposed method in number of ROI.(3)In the stage of traffic sign classification,to improve the speed and accuracy of identification,this paper chooses the relatively simple Lenet-5 network and improves it.In order to obtain a large dataset,the Chinese Traffic Sign Dataset(CTSD)was amplified by 20 times.Compare on amplified dataset,the experimental results shows that the improved network model far exceeded the traditional Lenet-5 network in terms of convergence speed and accuracy.To verify the reliability of model,this paper compare with other mainstream algorithms on representative German Traffic Sign Dataset(GTSDB),validating the competitiveness of proposed method on time and accuracy.