| Vehicle re-identification belongs to the category of urban intelligent transportation,and has received widespread attention due to its ability of identifying vehicles only through the appearance of the vehicle.Vehicle re-identification methods not only can save manpower and material resources,but also can efficiently,quickly,and accurately identify vehicles,winning the necessary opportunities for handling emergencies.Vehicle re-identification faces enormous challenges,mainly includes:(1)How to more efficiently utilize global attributes to identify vehicles;(2)How to use some subtle information(such as stickers,decorations,etc.)to identify vehicles;(3)Vehicle view differences seriously affect re-identification performance,how to more effectively utilize vehicle view differences for vehicle re-identification.In face of these challenges,this paper has conducted research on vehicle re-identification from the following three aspects: First,how to efficiently utilize global attributes such as color,vehicle type,and contour for vehicle re-identification,and effectively integrate attributes such as color,vehicle type,and contour from a global perspective,and improve the re-identification performance from a macro perspective;Second,how to more effectively extract fine-grained vehicle features based on the detail information richness of vehicle components,and more effectively use the vehicle details to improve re-identification performance;Third,because vehicle view differences can produce re-identification differences,how to use view information for testing without affecting training progress to improve re-identification performance during the testing phase.The first two points are both from a training perspective,hoping to obtain better model parameters.The third point is a post processing method that is independent of the training progress and improves performance by re-ranking the re-identification results.In face of these challenges,this paper has made the following contributions:(1)In order to better reduce the distance between features of the same vehicle and increase the distance between features of different vehicles,making it easier to distinguish vehicles in different images,this paper proposes a dual distance center loss function(DDCL)for the field of vehicle re-identification.The traditional center loss method uses Euclidean distance as a tool to calculate the distance between vehicle features.Based on this,the proposed dual distance center loss method also adds a Pearson distance.The combined effect of Euclidean distance and Pearson distance makes the trained reidentification model more stable.By setting a center isolation threshold,the proposed dual distance center loss can independently train the vehicle re-ientification model without cross entropy loss.(2)Different parts of the vehicle,especially the window part and the below-window part,have different detail information.In order to make full use of this difference,this paper proposes a vehicle re-identification method based on image segmentation to improve the accuracy of vehicle re-identification.According to the richness of vehicle detail information,the method uses the orientation extraction network to extract the vehicle view label,uses the keypoint detection network to detect the keypoint coordinates,combines the orientation label and keypoint indexes to select the corresponding visible keypoint coordinates,and divides the vehicle into window parts and below-window part,uses three re-identification branches to extract the window image features with relatively little detail information,the below-window image features with relatively rich detail information and the global features contained in the original unsegmented image,and fuse these three parts features for vehicle re-identification.(3)This paper proposes a view-aware-based post-processing method(VABPP)for vehicle re-identification.This method integrates view information into the trained vehicle re-identification model to improve the re-identification performance of the model.During testing,the VABPP method reduces the distance between the candidate image that has a larger view difference from the query image and the query image is reduced to a greater extent,and their ranking moves forward significantly to reduce matching errors caused by view differences.Since VABPP is only used during the testing phase,the corresponding vehicle re-identification method has good overall performance.For a query image,VABPP optimizes the final re-identification result by overall scaling the distance between the features of all gallery set images with the same view and the query image feature using the same scaling factor during testing.Because the view-pair vehicle identity feature distance scaling factor is obtained statistically through training set,and then directly used in the test set,and due to the different feature distance distributions between the training set and the test set,this paper also designs a scheme to unify the two feature distance distributions,making the feature distance distribution of the test set approximate that of the training set,which makes view-pair vehicle identity feature distance scaling coefficient matrix calculated in the training set can be directly used in the test set.In summary,this paper mainly conducts research on global feature based vehicle re-identification,local feature based vehicle-identification,and view-aware based postprocessing vehicle re-identification.The effectiveness of the proposed method has been verified through experiments on multiple vehicle re-identification datasets. |
