Research On Vehicle Re-Identification Methods Based On UAV Aerial Images

As an important part of the urban transportation network,vehicles have become a key research object in the current intelligent transportation system.Vehicle re-identification aims to identify the identities of vehicles.It can search for vehicle images with the same identity as the specified vehicle in a large-scale surveillance network.It plays an important role in applications such as traffic security personnel searching for specific vehicles,cross-view vehicle tracking,and vehicle behavior analysis.The current vehicle re-identification researches are mainly based on road surveillance cameras,which have certain limitations in terms of geographical restrictions and flexibility.In comparison,because the camera platform based on a UAV has high degree of freedom.It has important research value in the field of vehicle re-identification.Therefore,this dissertation conducts research on vehicle re-identification based on UAV aerial images.Currently,vehicle re-identification is faced with challenges such as different viewpoints of vehicle images,insufficient specificity of vehicle feature expression,and the difficulty in identifying hard samples whose intra-class distance is greater than the inter-class distance.In response to the above challenges,this dissertation studies the use of vehicle viewpoint informations,the accurate expression of vehicle visual features,and the targeted learning of difficult samples.And then proposed a series of methods and models.Through a large number of experiments on the vehicle re-identification dataset based on UAV aerial images,the effectiveness and advancement of the proposed methods are verified.The main research contents and innovations of this dissertation are as follows:(1)A large-scale vehicle re-identification dataset based on UAV aerial named VeRi-UAV.Currently,there are few researches on vehicle re-identification based on UAV aerial images,and there is a lack of directly obtainable aerial vehicle re-identification datasets.Therefore,this dissertation constructs a UAV based vehicle re-identification dataset containing 17,516 vehicle samples with rich annotations to provide a data basis for aerial vehicle re-identification researches.(2)View-Decision Based Compound Match Learning Network(VD-CML).Aiming at the difficulty of the re-identification of the aerial vehicle samples in different viewpoints,this dissertation proposes a View-Decision Based Compound Match Learning Network to conduct the vehicle re-identification.First,the viewpoint decision model is designed to generate sample pairs with the speci fic connections,and then a multi-branch Siamese network is constructed.The multi-branch separable Siamese network learns the specific features of the composite sample pair associated with a specific view.Then the network parameters are updated by the multivariate compound loss function to obtain the final vehicle re-identification model,which effectively improves the accuracy of the vehicle re-identification in the cross-view.(3)Posture Calibration Based Hard Samples Perceptive Metric Learning(PC-HSPML).Aiming at the variable posture of vehicles and the low recognition rate of difficult samples in aerial images,a novel metric learning method based on posture calibration and perception of hard samples is proposed.This method implements a complete framework from vehicle detection,segmentation to re-identification.It mainly includes three parts:Firstly,a vehicle segmentation network based on a guided anchoring region proposal net(GA-RPN)module and an atrous spatial pyramid pooling(ASPP)module is proposed for segmenting UAV-captured vehicles under different heights and directions;secondly,a posture calibration model is designed to uniform the vehicle postures based on the segmentation results,with the purpose of reducing the influence of different postures;thirdly,the novel Hard&Cross-view Perceptive Metric Learning(HCPML)method is proposed to train a ReID network with random cross-view training constraint and hard perceptive principle,which improves the low ReID accuracy brought by cross-view or hard samples.This dissertation conducts a large number of verification experiments on the proposed methods in accordance with the evaluation standards in current advanced vehicle re-identification research.And then compares our methods with a number of advanced methods in recent years,which fully proves the advancement and effectiveness of methods in this paper.