Traffic Sign Recognition Based On Deep Learning

Intelligent driving is a technology that helps the driver to recognize the road condition and make corresponding judgment and response in the process of vehicle driving.It USES on-board sensors to obtain environmental information,and obtains the perception result through the comprehensive calculation of the acquired information.Traffic sign recognition is an important part of this technology.As the system is to be applied to the moving vehicle,the wrong recognition result will bring huge security risks,so the recognition accuracy of the algorithm is required to be high,and the real-time performance of the algorithm determines whether the system can be converted into a product with practical application.In this paper,the deep learning-based traffic sign recognition algorithm is studied to improve the recognition accuracy and running rate of the deep learning model in the aspects of data amplification,model compression optimization and quantification.Basically has the following three aspects:(1)the designated cities traffic signs on the application of the system data sets collected,through experiment contrast choose VGGNet network on the model of training data set for the idea of SamplePairing multiplies the size of training data,at the same time for training results increased the regularization penalty term,clearly the boundary between the categories and improve the model accuracy.(2)with the idea of channel pruning,the channel selection and feature map reconstruction are carried out based on LASSO regression and least square method,and the VGG model is compressed and optimized while the accuracy is basically unchanged.(3)the training model is quantified according to the idea of k-l divergence optimization,the parameter distribution of the model is fitted,and the convolution operation of float32 bit is converted into the convolution operation of int8,so as to reduce the running time of the compression model.In this paper,the training model and related optimization algorithm are tested and the results are analyzed.The accuracy of the model was improved by 0.9% after data amplification for the self-collected data set.Through the two-step optimization experiment of model channel pruning and quantization,the VGG model was compressed from 537 M to127M,and the running speed was increased by 53.3%.The interface of traffic sign recognitionsystem based on QT graphical interface library and caffe framework is designed.