Research On Vehicle Re-identification Method Based On Pose Constraints

Vehicle re-identification(Vehicle Re-ID)refers to the task of searching the same ID(Identity)vehicle in the cross-camera vehicle dataset given a query vehicle image.It has great application values in public safety and intelligent transport.There are two main challenges in vehicle re-id: due to the same vehicle model,the vehicles with different IDs may have high appearance similarity(high inter-class similarity),and due to the different poses,the vehicles with the same ID may have low appearance similarity(low intra-class similarity).Between these two challenges,the different poses will influence the performance of vehicle re-id model dramatically.Therefore,it is important to learn the pose-independent feature for vehicle re-id.In this paper,the main contributions are as follows:(1)The existing methods only consider the ID but ignore the impact of the same model and different poses.To address this problem,we proposed a multi-cluster center loss based method,to reduce inter-class similarity and increase the intra-class similarity by adding both model and pose constraints.(2)To address the problem that the different poses make it difficult for the deep model learn the discriminative features,we proposed a multi-pose generation adversarial network based on vehicle keypoints,which is called pose-GAN in this paper.This model uses ID and pose constrains to make the feature representation more robust and insensitive to different poses.The experimental results on Vehicle ID and Ve Ri-776 dataset show that the proposed methods in this paper are superior to the existing vehicle re-identification methods on rank-1 and m AP.