Research On Aerial Vehicle Re-Identification Based On Multi-View Mask Segmentation And Counterfactual Attention

With the development of urban traffic network,the dominant position of vehicles in the city is becoming more and more prominent,which has become a key research object in the current urban intelligent transportation system.Vehicle re-identification technology is a hot field in urban intelligent transportation systems.This technology can determine whether the vehicle image in a specific scene and the target vehicle belong to the same vehicle.It is essentially a retrieval problem.Vehicle re-identification technology is widely used in the field of urban transportation,and it is involved in vehicle search and tracking,automatic toll collection,parking lot access,access control and other fields.Most of the current vehicle reidentification technologies are researched based on the datasets captured by road monitoring.These data sets have a fixed perspective and a single scene,and the effect of vehicle reidentification technology is limited.The vehicle dataset based on aerial photography of UAV has rich perspectives and changing scenes,which has important research value in the field of vehicle re-identification.Therefore,this paper studies vehicle re-identification based on UAV aerial images.The current vehicle re-identification technology mainly faces challenges such as intraclass difference,inter-class similarity,insufficient detail of vehicle feature expression,and intraclass distance is greater than inter-class distance.In response to the above problems,this paper conducts research from three perspectives:vehicle perspective information,vehicle detail features,and metric learning methods.The main research contents and innovations of this paper are as follows:(1)A multi-view mask vehicle re-identification dataset based on UAV aerial named VeRiUAV-MVM,The current vehicle image datasets are relatively lacking in the annotations of the various perspective parts of the vehicle,especially on the vehicle datasets based on drone aerial photography.However,the perspective information in the vehicle dataset of UAV aerial photography is relatively complex,so the extraction of vehicle perspective information from aerial images is particularly important.This paper constructs a VeRi-UAV-MVM dataset containing 3,000 mask annotation images of different perspective parts of vehicles based on UAV aerial photography,which provides a data basis for the research on vehicle reidentification methods from the perspective parts.(2)Multi View Mask Feature Based Hard Sample Learning(MVMF-HSL).Aiming at the problem of intra-class difference and inter-class similarity caused by viewing angles,this paper proposes Multi View Mask Feature Based Hard Sample Learning Method.First,a mask segmentation network is designed to segment the vehicle image into four different regions,front,back,side and roof,and the mask features of these four different viewing angle regions are extracted respectively.A Common Visible Attention Module is then designed to assign weights to the four mask features.Finally,a metric learning method based on difficult samples is proposed to train the vehicle re-identification network.The utilization of information from different perspectives and the learning of difficult samples effectively improve the recognition accuracy of the vehicle re-identification network for vehicle images from different perspectives.(3)Fine Grained Features Based Binomial Joint Promotion Learning(FGF-BJPL).Aiming at the problem that different vehicles with the same model and color are difficult to distinguish,this paper designs a Counterfactual Attention Learning Network.The network extracts finegrained features in vehicle images by introducing a counterfactual attention mechanism,which effectively solves the problem of indistinguishability between similar vehicles;Aiming at the occlusion problem in UAV aerial images,this paper designs a Counterfactual Attention Data Enhancement Module,which uses the graph attention learned by the Counterfactual Attention Network to selectively crop and erase the original vehicle image,In this way,the diversity and complexity of sample data can be enhanced,and the network’s recognition rate of occluded images can be improved;Finally,aiming at the problem that the intra-class distance of vehicles is greater than the inter-class distance,Hard Samples Binomial Joint Promotion Loss is proposed,which comprehensively considers the relative and absolute distances between positive and negative samples,and the separate training for difficult samples improves the recognition accuracy of the network for difficult samples,thereby improving the generalization ability of the network.In this paper,the vehicle re-identification evaluation indicators(Top-k and mAP)widely used by scholars today are used to experimentally verify the proposed method.Compared with the vehicle re-identification methods in recent years,and verified on two challenging aerial vehicle datasets,VeRi-UAV and UAV-VeID,it fu lly proves the advancement and effectiveness of the method in this paper.