Person Re-identification Based On Metric Learning

With the development of computer vision and pattern recognition,person reidentification has been a useful tool for preventing potential violent situations.Person re-identification is a process of matching person images of same identity across nonoverlapping camera views.It is still a challenging task because person images captured from different camera views often undergo significant variations in resolution,illumination,pose,background and occlusions.In order to tackle these problems,the existing research on person re-identification has two directions: extracting a robust person representation and learning a suitable distance metric.In this dissertation,I focused on the latter.For person re-identification,the main works in this dissertation are as follows:1.The overview of person re-identification was summarized.Firstly,I briefly introduced the background,significance and development history of person reidentification.Then,the existing person re-identification methods were introduced from the aspects of feature extraction and metric learning according to the research emphasis of person re-identification.2.The kernel metric learning based person re-identification methods were introduced.The main advantage of kernel methods is to improve the classification performance by mapping the original data into a high-dimensional space without the knowledge of nonlinear mapping explicitly.I introduced three kernel metric learning methods based on Large Margin Nearest Neighbor(LMNN),Local Fisher Discriminant Analysis(LFDA)and Null Foley Sammon Transform(NFST).The experimental results and analyses on three challenging person re-identification datasets verified the effectiveness of the three kernel metric learning methods.3.Asymmetric geometrical metric learning(AGML)method was proposed for person re-identification.It learns camera-specific projection transformations from the perspective of geometrical relationship to improve the performance of symmetric metric learning models for person re-identification.The existing symmetric metric learning methods are all focusing on learning a unitary projection transformation for all camera views,which ignores the person appearance discrepancy problem across nonoverlapping camera views.AGML can successfully solve the problem of the symmetric metric learning model on person re-identification.The performance on three person reidentification datasets demonstrated the effectiveness of AGML for person reidentification.