Research On Traffic Sign Recognition Method Based On The DNN Technology

As the important part of the intelligent driving system, traffic sign recognition is the typical application of image processing, pattern recognition and machine vision crossing multidisciplinary research, but also it is the difficult real graphics recognition, which has not been solved in the intelligent driving system field.In this thesis, it firstly introduces the development status and theoretical research results of traffic sign recognition algorithms at home and abroad, such as the nearest neighbor method, similarity coefficient method, clustering analysis method, decision tree and matching projection method based on statistical, based on pixel level or feature level template matching method, syntactic classification and integration classification algorithm, and it also introduces in detail the deep neural network(DNN) algorithm of machine learning, and applies on traffic sign recognition after carefully studying. The main work in this thesis is as follows:(1)Traffic sign recognition based on DNN technology. In this part, it is studying on the traffic sign recognition using pretreatment method and optimal DNN structure. Six different pretreatment processes are carried out on the same sample data set, which is the different pretreatment training set. At the same time, it adjusts the neural network structure of convolution kernel size, sampling window size, bulk sample number and number of figure maps, and then identifies the optimal DNN structure of each training set by experiments. At last, it analyzes the corresponding relations between the network structure and different pretreatment methods.(2) Traffic sign recognition based on the multi columns deep neural network(MCDNN) and optimized stategy. It applies the different pretreatment on the same sample set, and consisting a MCDNN by selecting the best effect DNN on each pretreatment training set. And fuzzy mathematics is applied to the comprehensive evaluation of the each DNN. There are two decision ways of each DNN output, which are comprehensive membership decision and discrete membership decision. The final classification of the input image is according to the principle of maximum membership degree, and it compares the recognition results with single DNN.The pre processed training sets as input of MCDNN, and the best recognition effect of DNN on each training set constitute MCDNN, and it applies fuzzy math to the output of each column. It carries on the comprehensive membership decision and discrete membership decision on the output of each column, according to the principle of maximum degree of membership on each training image classification, and then comparing the results of single DNN recognition.The experiment shows that the deep neural network can be used on traffic sign recognition and has a good effect on traffic sign recognition without the tedious adjustment and modification. When the pretreatment method is different, the optimal network structure is also different, which shows that there is a link between the network structure and the pretreatment method. And the recognition effect is not ideal on the training set that integrates more kinds of preprocessing methods. At the same time, it builds multi columns deep neural network on the same data set after different pretreatment, and fuzzy judgment is used on the output multiple columns. The experiment result shows that MCDNN is better than single DNN on recognition rate.