Research On Vehicle Recognition Technology Based On Deep Convolutional Neural Network

In recent years,the rapid increase in the number of automobiles has provided convenience for people’s travel.It has also caused a series of traffic problems such as traffic jams and parking difficulties.With the rapid development of the intelligent transportation system,the state encourages the development of traffic science and technology projects such as the "three-dimensional social security prevention and control system","smart city" and "safe city".The vehicle identification technology is the core of the traffic science and technology strong police project.Therefore,it is significant to study an efficient model for vehicle classification and classification.The traditional vehicle identification can only identify the general model of the vehicle.The research on the identification of the fine model is still relatively small,and the accuracy is not ideal.Therefore,the vehicle identification technology based on the deep convolutional neural network has become the current research hotspot.This technology uses the self-learning vehicle characteristics through the use of deep convolutional neural network to calculate the traffic flow information of the bayonet,and realizes the accurate recognition of vehicle models.The accurate identification of the models plays an important role in the fight against counterfeit deck vehicles and illegal vehicle investigations,and is of great significance to road traffic safety.This article adopts the vehicle identification scheme based on the deep convolutional neural network model to realize vehicle identification.First,introduce the basic theory of vehicle identification technology and the construction of database:introduce the basic principles of vehicle identification technology,analyze the structure of convolutional neural network in detail,introduce the characteristics and construction process of caffe deep learning framework,and complete the construction of caffe environment;select Qingdao The image of vehicles in the bayonet city has established a database of 2000 types of detailed models.Each category contains 1000 vehicle images,which are named after the vehicle brand-sub-brand-year format（eg jeep_free passenger₂0122014,Audi_S7₂0132014,BMW_X5₂014,etc.）Using the sample comparison screening method to complete the filling of the fine model database,realizing any feature combination of vehicle brands,sub-brands,and annual models,making the vehicle identification technology application platform search elements more diverse and targeted.More accurate.Secondly,the vehicle identification algorithm based on feature extraction is introduced.The optimal parameters of the algorithm are found on the database of refined models established in this paper:SIFI feature dimension is 20,LBP feature dimension is 30,HOG feature bin is 5,and block is 8.Using SIFI,LBP and HOG features,combined with SVM classifier for vehicle identification testing,the performance comparison and test results of the three algorithms were analyzed.The results show that:HOG features better recognition of vehicle models.Through comparison of alternatives,a sophisticated vehicle identification technology solution based on deep convolutional neural network was established.The deep convolutional neural network model was optimized from four aspects of network layers,iterations,learning effi ciency,and activation function to determine the optimal model.:The number of convolutional neural networks is 5,the maximum number of iterations is set to 30000,the learning rate is set to 0.001,and the activation functions used in the convolutional layer and downsampling are ReLu functions and sigmoid functions,respectively,enabling the recognition accuracy of fine models to reach 89.23%.A new classification method is proposed:classification is based on Softmax full-connected layer,and the recognition effect of this method is proven.Finally,the results of the recognition test of the deep convolutional neural network model are analyzed.The results of the network model test are compared with SIFI-based,LBP-based,HOG-based,and CNN+SVM-based algorithms.The results show that:The deep convolutional neural network model has the best recognition effect.The recognition accuracy of the refined model reaches 89.23%.