Person re-identification is the problem of judging whether person images captured from different cameras belong to the same person.In large-scale distributed multi-camera systems,We need to distinguish persons at different locations and in different time.The system is widely used in long-term multi-camera tracking and criminal investigation.Due to the influence of factors such as changes in illuminations,views,background,occlusion and low resolution on person images,the research on person re-identification is challenging.At present,the research on person re-identification is mainly focused on two aspects:(1)Designing a robust feature representation model that can distinguish different persons and overcome the changes in illuminations,views,background,etc.(2)Designing a discriminative similarity metric learning algorithm that can make the distance of intra-class is small than inter-class.In the step of feature extract,we propose a feature representation model.Firstly,we apply the Retinex algorithm to preprocess person images.The Retinex algorithm aims at producing person images that is consistent to human observation of the scene.Then,foreground regions of person images are segmented by co-segmentation algorithm.Finally,we propose a feature representation model which based on foreground segmentation.In the step of similarity metric learning,we propose a discriminative similarity metric learning algorithm.Due to the feature eigenvectors of Person images captured by different cameras are generally inconsistent,the feature eigenvectors are projected into the subspace through the projection matrix.Then,the distance of a pair of person images are measured by combining the commonness and difference.Experiments on three challenging person re-identification databases,VIPeR,GRID,CUHK01,show that the proposed similarity metric algorithm and feature representation model performs better than others. |