Local Feature Based Vehicle Re-Identification Method

Vehicle re-identification(Re-ID)is the basis of security,surveillance video analysis and un-derstanding.It is a key method of building Skynet Project and Safe City.According to whether the training data contains manually labels,the vehicle Re-ID problem can be divided into su-pervised vehicle Re-ID problem and unsupervised vehicle Re-ID problem.The biggest problem of supervised vehicle Re-ID is the near similarity challenge.In order to address the problem,a novel two-branch stripe-based and attribute-aware deep convolutional neural network(SAN)is de-signed and implemented in this paper.At the same time,unsupervised vehicle Re-ID is facing two problems:the dataset bias and impossibility of training without labels.To overcome them,this paper proposes a local feature based and self-similarity aware deep convolutional neural net-work(LFSSN).Experimental results on two of the most common vehicle datasets show that the proposed SAN and LFSSN methods are both superior to the existing supervised and unsupervised vehicle Re-ID methods,respectively.The following are the detailed achievements of this paper:(1)Convolution neural network,cross-entropy loss and triplet loss with hard mining are adopted in this paper to implement baseline networks for supervised and unsupervised vehicle Re-ID tasks,respectively.And in the train phase,the smooth convergence loss and increasing accuracy of the verification set prove the rationality of the baseline networks.(2)In order to address the near similarity challenge of supervised vehicle Re-ID task,SAN is proposed by this paper.SAN integrates the stripe-based branch and attribute-aware branch into a unified architecture for end-to-end training.The stripe-based branch can extract local clues of vehicle image while the attribute-aware branch can utilize attribute labels to get more discrimi-native global feature.The extensive experiments on both VehicleID and VeRi datasets show that the proposed SAN method can discovery outline and subtle but discriminative local features of vehicle identity.As a result,a supervised vehicle Re-ID model with leading robustness and high accuracy is achieved by the proposed SAN method.(3)To overcome the dataset bias and impossibility of training without labels problem,an unsupervised vehicle Re-ID method is proposed in this paper,which makes use of both global and local features for clustering,and utilizes the pseudo labels generated by clustering for re-training the unsupervised vehicle Re-ID network.Experimental results on public datasets show that the proposed unsupervised vehicle Re-ID method can significantly improve the re-identification ability and generalization ability of the Re-ID model.