Traffic Sign Detection And Recognition Algorithm Research Based On Convolutional Networks

Nowadays,with the rapid development of our country’s economy,there are more and more cars.However,with the increase of the cars,our country is facing the huge traffic pressure and the traffic accident is frequent.The problem of safe driving is more and more serious.So it is necessary to develop intelligent transportation system.Traffic sign detection and recognition is an important part of the intelligent transportation system,and has great practical significance in traffic safety.It can reduce the driver’s fatigue and better ensure the safety of travel.However,the detection and recognition of traffic signs is a challenging task.In real environment,traffic signs are often in a complex background,and it may be obscured,severe aging,vague,strong light leading to reflection.All of these factors will make the system more difficult in the process of detection and recognition.Because the intelligent transportation system is installed in the car,the system must be accurate to help the driver in detection and recognition.On the basis of the former research,I mainly use the deep learning technology to detect and identify the traffic signs.The main contributions of this paper are as follows:First,traffic signs are small target,which occupies very small in the whole image.But in the existing algorithm,the purpose of applying deep learning is to detect the big target.The proportion of the target is relatively large in the whole image.Applying convolutional neural network to process image is the mainstream method,but it is very difficult to find the detail information in the last feature map after a few layers of convolution,so the existing methods are not very good for small target detection.In order to solve this problem,the algorithm is improved on the basis of previous work,not only using the last layer convolution feature map but also using the low level feature map.Because the low level feature map still retains the feature information of the small target,so we use the previous methods to operate on the last feature map and the second feature map to improve the accuracy of detection.At the same time the algorithm is universal,the traditional methods are designed for the specific characteristics of the country’s traffic signs,the system can only detect the specific countries’ s traffic signs,but applying the system to detect other country’s traffic signs is not good.And our algorithm apply deep learning to learn feature automaticly,which can be used in other country in traffic sign detection,so our algorithm is universal.Second,in order to solve the problem that traffic signs are complex in real environment and the speed is slow when training the network,this paper proposes a traffic sign recognition algorithm based on the optimized convolution network structure.In this method,the structure of convolutional neural network is optimized.In order to solve the problem that the network structure is slow in training,we add a batch-normalization layer to the network.When training the convolution neural network structure,the first layer’s training will inevitably lead to the input data distribution change of the second layer,so that the second layer adapt to the data distribution again.Joining the batch normalized layer make the output data of each layer normalize to zero mean value and the variance 1.And the convolution layer does not have to adapt to the data distribution after each training,so that the training speed is greatly accelerated.At the same time,the support vector machine classifier is used in the final classification.Support vector machine is only concerned about the error sample of classification,no longer concerned about the correct samples.This reduces the training samples and accelerates the training speed.In order to solve the problem that the recognition rate is not good due to the complexity of the real environment,the greedy layer-wise pretrain method is used.The method train data layer by layer,and first training the first layer to retain the parameters,and then training the second layer to retain the parameters,and so on.Based on the above methods,the convolution network structure is optimized,which not only improves the training speed,but also improves the recognition rate of traffic signs in real environment.