Research On Vehicle Model Recognition Method Based On Convolutional Neural Networ

Vehicle identification is not only an important research area in the field of intelligent transportation,but also the road traffic law enforcement departments to trace the "gray license"vehicles,narrow the scope of the vehicle to determine the scope of the urgent needs.The traditional model recognition method is limited by the vehicle scene and the bottom feature extraction,and has not been able to apply the recognition effect.Convolution neural network can directly use the image data for network model training,and then used for vehicle identification,to overcome the different differences in vehicle scenarios and other differences,and to avoid the difficulties of the underlying characteristics of access.The main work of this paper is summarized as follows:1.Establishment of Picture Data Set for Expressway High-Definition Vehicles.Highdefinition vehicle picture data collected in the expressway of the various bayonet actual monitoring of high-definition equipment captured pictures,and the various bayonet of high-definition vehicle pictures manually marked the car,micro-face,truck,bus four categories of models.Respectively,including different scales,lighting,angle and other different circumstances of the vehicle pictures,according to a certain percentage of selected training set and validation set.The data set is the basis of vehicle identification research.2.Comparative Study on the Recognition Methods of Commonly Used Machine Learning.The three kinds of neural network methods,which are commonly used in machine learning,are used to experiment with the vehicle size characteristics in the model standard feature model library.The Kohonen network is used to analyze the number of samples,the number of classified samples and the recognition rate.The classification of the vehicle is better and the average recognition rate is 91%.3.Research on Vehicle Recognition Method of Convolution Neural Network.Depth learning platform framework in the Ubuntu system to build,the system open source,easy multi-user remote access.Utilize the CUDA driver to make the depth learning framework call GPU acceleration calculations to solve the problem of a large number of image datasets running slowly.The training speed of deep convolution neural network is accelerated by CUDNN,which provides an experimental environment better than MATLAB simulation.The data set is experimented by using the convolution neural network model of three different network layer depths integrated in Caffe framework.Then,the convolution neural network model of three different network layers integrated in the Torch framework is used to experiment the data set,and the experimental results are compared and analyzed.Finally,two kinds of network models with high accuracy of frame recognition are compared and analyzed.4.Analysis of Experimental Results of Traditional Machine Learning and Convolution Neural Network.In this paper,for small sample data and a large number of vehicle image data sets for vehicle identification experiments.The results show that the vehicle recognition method based on convolution neural network has a high recognition rate of 95%compared with the traditional vehicle identification method.In the data processing stage with less manual engineering,and vehicle identification efficiency for different scales,lighting,camera angle and other factors have a strong resistance,but need to consume more resources and time.Through the comparative analysis and applicability of the experiment,the deep convolution neural network is used to improve the adaptability of vehicle models in large data and complex environment.The conclusion of this paper can provide a reference for the further study of vehicle identification technology.