Research On Cross-view Matching In Vehicle Re-identification

Vehicle re-identification(Re-ID),also known as cross-camera vehicle tracking,aims to retrieve all the samples of the particular target vehicle from a large gallery set where all the vehicles are captured from non-overlapping road surveillance cameras.Vehicle Re-ID needs to extract the discriminative features from the visual appearance of vehicles directly,but the vehicle visual appearance changes dramatically due to the extreme camera viewpoint variation.The drastic variation in vehicle appearance under different viewpoints greatly affects the performance of vehicle Re-ID.Therefore,cross-view matching brings two main challenges for vehicle Re-ID task:(1)Dramatic intra-class variability: when observed from different viewpoints,the same vehicle may have a completely different visual appearance and(2)Small inter-class variability: for vehicles which have the same color and type,the visual appearance is similar when observed from the same viewpoint.Hence the key issue in cross-view vehicle Re-ID is learning an effective feature representation that is robust to both dramatic intra-class variability and small inter-class variability.To achieve this goal,the following two cross-view vehicle Re-ID approaches have been proposed in this paper.(1)The cross-view vehicle Re-ID algorithm based on view-aware sphere learningFirstly,the Sphere Embedding Feature Baseline(SEFB)is designed to solve small inter-class variability problem.The SEFB use L2-Normalization to project the feature to a hypersphere and the Sphere Cross Entropy Loss(SCE)and the Sphere Similarity Triplet Loss(SST)are proposed to supervise the training process of SFEB.Furthermore,an effective sample selection strategy for ranking loss is designed to obtain discriminative feature.Secondly,a semantic vehicle viewpoint classifier is trained to obtain viewpoint information.Due to the lack of labeled semantic viewpoint data,a dataset is collected for semantic viewpoint classification in this paper.Thirdly,by combining the predicted semantic viewpoint probability and the sphere feature embedding algorithm,the Viewaware Sphere Learning(VSL)method is proposed to extract view-aware feature.The Cross Viewpoint Orthogonal Regularization(CVO)is designed to reduce the relevance of features from different viewpoint.The Attention Global Pooling(AGP)is designed to help the Re-ID model extract effective features.(2)The cross-view vehicle Re-ID algorithm based on multi-center metric learningModeling vehicle viewpoint with supervised learning methods require large-scale manually annotated viewpoint labels.Due to the high cost of data annotation,an unsupervised learning vehicle viewpoint modeling method is proposed in this paper.This method model latent viewpoint from vehicle visual appearance directly without any extra labels except ID.Firstly,several latent viewpoint clusters are defined for a vehicle to model latent multiple viewpoints and each view cluster has a learnable center.Then the intra-class ranking loss(IRL)with cross-view center constraint and the cross-class ranking loss(CRL)with cross-vehicle center constraint are proposed to address crossview matching and cross-target matching,respectively.Thirdly,by combining all the items in this model,the multi-center ranking loss(MCRL)is proposed to learn intra-class and inter-class information jointly.With the optimization of MCRL,the multi-center metric learning algorithm can obtain a feature representation that is robust to both dramatic intra-class variability and small inter-class variability for cross-view vehicle ReID.In this paper,two algorithms are proposed to handle the cross-view matching problem in vehicle Re-ID.The two algorithms model vehicle viewpoint information based on supervised and unsupervised learning,respectively.The experimental results show that the proposed algorithms are practical and effective,improve the cross-view matching performance of vehicle Re-ID model.