Research On Traffic Sign Recognition Technology Of Autonomous Vehicle Based On Dense Network

As we know,the detecting and recognizing of traffic signs are the important parts of the environment perception technology for autonomous car.It points out the direction for the research of traffic signs detection and recognition that the continuous improvement of computer computing abilityand the further research of machine learning and deeping learning generated by artificial intelligence.In this paper,we aimed at studying the high real-time and accurate automatic traffic signs detection and recognition system.By studying and analyzing the deep convolution neural networks with excellent performance in various aspects,this paper designed the traffic signs detection and recognition model with the method of deeping learning.The mainly research work is carried out as follows:1)A review of deep convolution neural network.Firstly,the basic components and methods of convolution neural network are introduced.These components and methods play the important roles in the efficiency of network feature extraction,reducing computation and accelerating network convergence.Then,the structure and development of various deep convolution neural networks are introduced.The result is that the performance of the deep convolution neural networks is constantly improved.This paper focuses on the analysis of Res Net(residual network)and Dense Net(dense network).By expanding Res Net,it has been found that when the actual path is longer than the effective path,the weights of forward propagation process can not be updated.It is the significance of Dense Net which can ensure updating weights in the process of backpropagation.2)Designing the dense network of traffic signs detection and recognition model.In this paper,the dense network model is composed of three parts,namely,preprocessing network,feature extraction network and classified network.The preprocessing network refers to conversing color space of the dataset roughly,and extracting the image features,which is the preprocessing method for the subsequent feature extraction layer.The feature extraction network adopted the dense block structure of Dense Net,but the designing idea of wide and shallow dense network is proposed in this paper,which reduced the training time and memory occupation.The classification network adopted average-pooling(global mean pool),and each feature graph corresponded to an output class feature.3)The dataset argumentation process.This paper adopted the German Traffic Sign Dataset to train and validate the model.The number of training set was expanded by the flip operation and the data enhancement operation.Taking into account the unbalanced number of images in various types,the training set was resamplied,and the training samples were randomly selected through morphological transformation and illumination transformation.After the expansion of the data,the number of training sets increased to 860000.4)Training and testing the dense network model.The super parameters were set up in the training process.The loss function included Soft Max classification cross entropy loss and L2 regular loss.In this paper,the loss function optimization parser adopted SGD(Stochastic Gradient Descent).In this paper,the dynamic data expansion strategy was used in iterative training to make the network adaptive to the change of training dataset.In the detection and recognition experiment,this paper detected and recognized the traffic signs in the general environment and the special environment.