Traffic Sign Recognition Algorithm Research Based On Convolutional Neural Network

Traffic sign detection and recognition is an important part of intelligent transportation system.It can provide necessary visual extension for drivers,real and effective data for traffic information system,and necessary decision-making basis for unmanned drive system,which can effectively guarantee the driver’s safety,improve the intelligence level of the vehicle.There is not only important academic value,but also great social value and economic value for this research.However,the research involving a wide range of field,still faces many difficulties because of complexity of the traffic signs detection and recognition environment.In this thesis,based on full research and summarization of the existing methods,trying to take a full use of the traffic signs of color information and shape information,we propose a quick traffic signs detection algorithm.We also proposes a traffic signs classification network which is fast and accurate.The main work of this thesis is as follows:(1)Multiple channel color enhanced map gained by enhancing specific traffic signs color information,which improves the performance of the maximally stable extremal region al compare with traditional single channel image.In addition,limit value range of DtB feature descriptor to several discrete values,making it more robust.Then,construct a small shape data set,on which a Bayesian classifier based on DtB feature is trained.Combined with the above work,a fast traffic sign detection algorithm is completed.(2)In order to improve the accuracy and fastness of the convolution neural network used in traffic sign recognition,two improvements are proposed.Firstly,a small convolution neural network is added to the classification network,which is called spatial transformation module.A spatial transformation parameter(usually an affine transformation matrix)is obtained,and the feature map of the deformation is corrected and then sent to the classification for identification,thereby improving the accuracy of the classification network.Secondly,the multi-scale parallel convolution structure improves network’s local presentation,which can provide the performance of classification network on similar traffic signs.What’s more,the multi-scale parallel convolutional structure reduces the number of parameters of the classification network,which can not only improves the efficiency of network training and use,but also reduces the risk of the network being over-fitted.