Research On Optimization Of Traffic Sign Detection And Recognition Based On Deep Learning

In recent years,in the context of the rapid development of my country’s social economy and the rapid development of road transportation networks,the number of vehicles driving on roads in the natural environment has also seen a blowout growth.People are paying more and more attention to driving safety in road traffic.In the context of intelligent traffic,the traffic sign recognition system is extremely important,and has been widely used in various fields,such as the more popular automatic assisted driving in recent years.Autonomous driving on roads in real environments means that the adopted traffic sign recognition system must be able to quickly and accurately capture traffic signs.However,the environmental factors of actual roads are very complicated.Systems with high accuracy and good real-time performance for traffic sign recognition are under constant research.After reading a large amount of academic literature,this article mainly conducts research on motion detection,recognition,recognition and the driving safety of motor vehicles based on the convolutional neural network in deep machine learning technology.Specific work includes:(1)An improved recognition method for Lenet-5 network model is proposed.The traditional model requires a large amount of parameter calculation during training,which leads to low real-time performance.Besides,there are many kinds of traffic signs,and the recognition accuracy of traffic signs will not be very high in the trained network.Based on the traditional Lenet-5 network,this paper firstly uses the method of HOG feature extraction combined with SVM classifier to classify the traffic signs.After that,the structure of extra layers in the network model is optimized,and multiple fully connected layers are changed into one fully connected layer.In consideration of the real-time performance of the algorithm,the calculation amount for parameters of each layer in the network is reduced.To achieve the purpose of accelerated convergence of network model.(2)An improved traffic sign recognition network model is created.Combined with the characteristics of Inception and ResNet residual network model,the Inception-ResNet network model is finally constructed.On the one hand,the Inception module in the Inception network model can be used to extract more detailed traffic sign features,and all Inception modules used in the network have linearly activated 1 × 1 convolution layer to achieve the number of expanded feature extraction channels.On the other hand,due to the particularity of the propagation mechanism of ResNet residual network,the phenomenon of gradient disappearance can be avoided in the training of the network.Two convolution kernels of different sizes were selected for the experiment on GTSRB data set.The results show that the recognition accuracy of this method is 99.24% on the network model with the convolution kernel size of 3 × 3.It is a fast and effective traffic sign detection and recognition network model.