Research On Vehicle Identification Algorithm Based On Convolution Neural Network

Intelligent transportation system is an information service and management system for transportation based on modern electronic information technology.One of its outstanding technical features is that it takes the real-time collection,processing,release,exchange,analysis and utilization of information as the main line of services,providing a variety of information services for the participants and the public of intelligent transportation.Identification of vehicle detection and vehicle identification technology is a new generation of intelligent transportation system such as the important technical foundation and the core,in the system execution on deck of the vehicle recognition check,high-speed fee,vehicle fail to stop the traffic behavior,such as vehicle identification test,as well as the specific models of the stolen vehicle recognition,specific stolen vehicle tracking when performing specific tasks,such as vehicle identification technology in the intelligent transportation system has excellent performance.Traditional methods monitor the information of motor vehicles manually.So it not only fails to achieve the purpose of accurate tracking,but also easily hurts people’s eyesight.With the development of modern computer science,it is inevitable to use computer vision,data analysis and intelligent information processing methods to deal with cumbersome and mechanical motor vehicle related problems.The image field is the most mature field of deep learning.The image processing system constructed by convolutional neural network can better fit the training data and recognize pictures faster and more accurately after fusing GPU acceleration technology.Aiming at the above problems,this paper studies a more refined vehicle identification method based on deep convolutional neural network,which can enrich the vehicle target types in the Internet of vehicles and improve the accuracy of vehicle identification in the intelligent transportation system.The research content of this paper is as follows:(1)The network data set used for vehicle recognition and semantic segmentation was analyzed and introduced in detail,and the data in local overpass,street and other complex scenes were taken as the data set of vehicle recognition and semantic segmentation,that is,self-built vehicle data set,and the experimental data set was expanded.(2)Vehicle type identification.In view of the shortcomings of the basic SSD vehicle recognition algorithm in the detection accuracy and speed,as well as the excessive operation memory consumption due to multiple parameters,this paper proposes an improved SSD vehicle recognition algorithm based on the deep separable convolution,which is used as a lightweight algorithm model.Firstly,data expansion is carried out through spatial geometric transformation and pixel color transformation to solve the phenomenon of "under-learning" caused by the large gap in the number of samples of different car models.Secondly,a inverted residuals module is introduced before deep convolution to solve the problem of reduced accuracy due to fewer channels and feature compression.Thirdly,based on the rigid body characteristics of the vehicle,the K-means clustering algorithm is used to reconstruct the aspect ratio of the regional candidate frame to reduce the calculation amount of the model parameters and improve the recognition accuracy and speed of the vehicle model.Finally,the pre-trained model on the large classification data set was selected for network parameter fine-tuning,and the convergence of the model was accelerated by weight sharing to obtain better recognition effect.Moreover,the generalization ability of the model was tested on different data sets.The comparison experiment results show that the improved SSD proposed improves the frame rate of vehicle model recognition,verifies the effectiveness of the algorithm,reduces the network parameters under the condition that the recognition accuracy is guaranteed,accelerates the convergence speed of the vehicle target type recognition model training phase and the recognition speed at the detection stage,meets the real-time recognition requirements of vehicle target types,and can be applied to the detection of vehicles at traffic toll gates,stolen vehicle tracking,and other aspects.(3)The vehicle target is segmented.Aiming at the problem of the lack of edge information extraction and insufficient recognition ability of target detection algorithms in complex scenes,a U-Net network based on residual network unit is proposed to semantically segment vehicle targets to improve the model’s robustness in vehicle segmentation in complex scenes.The experiment increases the field shooting data and improves the generalization ability of the model to more conditional factors.The U-Net network is used for vehicle target segmentation,which combines shallow texture edge information with deep layer high semantic information to solve the vehicle target segmentation problem in complex environments.At the same time,drawing on the residual idea of Res Net(Residual Network),a residual network unit is added to the U-Net network to construct a residual U-Net network vehicle target semantic segmentation network to deal with the problem of network gradient disappearance for deeper network training,improve the training accuracy of segmentation.(4)Practical application of vehicle recognition algorithm.Based on the recognition algorithm of light SSD and the semantic segmentation algorithm of residual U-Net,a detection algorithm of vehicles in violation of traffic lane is proposed.By marking the offending vehicles and assisting the staff to judge the offending behaviors,this paper provides a reference for the application of the vehicle recognition algorithm based on the convolutional neural network.