Research And Implementation Of Vehicle Identification Based On Vehicle Full Face Features

With the development of urban intelligent transportation and the rapid growth of urban vehicle ownership,public security,traffic management and policy formulation have placed urgent demands on the intelligent identification and rapid information recording of vehicles.At present,the identification of vehicles mainly stays in the intelligent extraction of single information,that is,the external characteristics of the vehicle's license plate,model,color,etc.,the lack of overall information can not achieve the modified car,the deck car,the shelter car,and the same type identification of the vehicle.Most of the research of car face'recognition is to use some areas of the car face or roughly extract the texture features of the entire face of the vehicle,make the extraction of the car face features not perfect.In accordance with the development of intelligent transportation system,this thesis proposes a set of vehicle identification methods based on vehicle full face features,which does not depend on license plate number to identify vehicle identity information.Through the stepwise step-by-step extraction of the features of the vehicle face,the vehicle detection,the classification of the vehicle and the identification of the vehicle identity information have been completed.This process is related to the computer visual direction from detection classification to fine classification,and the stepwise deepening of instance recognition is corresponding.The whole process uses different deep learning techniques to design and achieve specific application goals.The main work of this thesis is as follows:1,.The front end of the vehicle is called the car face.The image data collection of the car face mainly uses the vehicle data set like PKU Vehicle ID,BIT Vehicle,Vehicle-1M,and uses the web crawler to crawl the picture containing the vehicle to enrich the dataset.On this basis,according to the needs of subsequent tasks,the corresponding sub-data sets of vehicle detection,component category,vehicle classification,image segmentation,and identification test are produced.2.Based on the Mobile Net-SSD network,an efficient vehicle detection is realized.At the same time,the feature pyramid model is introduced to design an important part extraction network for vehicle face based on SSD network.The network makes better use of local low-level texture features and global high-level abstract features,which improves the detection accuracy by 4.9%.Design and implement the vehicle classification network, separately input the extracted components and the full face image into the improved Alex Net and shallow volumes.After the feature combined,the full connection layer and the softmax layer are used to complete the feature fusion and classification,and improves the classification accuracy by 1.8% compares to the depth residual network.3.Using a single multi-tasking network to simultaneously complete the target detection,image segmentation,and feature extraction at the front windshield of the car face.A feature vector extraction method is designed.The segmented front windshield and the characteristics of different categories of interest regions are filtered by the segmented semantic information.After splicing integration and regularization process it is used to represent the identity information of the vehicle to realize the overall and fine identification of vehicle identity information.Based on the above research on the complete extraction and utilization of vehicle full face features and vehicle detection,classification and identification,compared with the last two re-identification work of vehicles this thesis has improved the recognition accuracy of vehicle identity by 6%.,and realized a multi-tasking vehicle identification system based on Vue.js and flask open source framework using Caffe and Tensorflow deep learning platform under linux system.