Key Technology Of Person Re-Identification

Public security has become one of hot issues. A large amount of surveillance cameras are set in crowded public place, such as school, shopping malls, subway stations, sports stadiums and parks etc. Video surveillance plays an important role in public security. It not only helps to prevent, react and record incidents of public, but also is helpful in backtracking, identification and search. Person re-identification is a challenging problem in multi-camera surveillance system. It aims at recognizing and searching persons from the images taken by spatially disjointed cameras. We proposed two novel approaches for person re-identification in this paper. The main contributions are the following:First of all, we introduce bagging into large margin nearest neighbor (LMNN) method, and propose a bagging-based LMNN method for person re-identification. Metric learning is widely used to model the transformation between cameras recently. However, traditional metric learning based methods only learn one metric for the whole feature space, which cannot model different kinds of appearance variations well. In this paper, multiple LMNN predictors are generated on sub-regions of the feature space and leveraged to obtain an aggregated predictor for performance improvement. Two bagging strategies, sample-bagging and feature-bagging, are proposed and compared. Extensive experiments on three benchmarks demonstrate the superiority of proposed approach over state-of-the-art methods.Secondly, we combine trace-norm regularization with LMNN and propose large margin nearest neighbor with trace-norm regularization (LMNN-T) method. Existing methods always are liable to be over-fitted when model the appearance variations of pedestrians between different cameras. In order to control the capacity of over-fitting, we propose LMNN-T, which combines trace-norm regularization with LMNN. The trace-norm regularization encourages the learned feature projection matrixes of low rank and thus controls the capacity of over-fitting. In addition, feature (attribute) bagging strategy is introduced to avoid dimension reduction, which may cause the loss of subtle feature information, and maintain the discriminative ability of image feature. Extensive experiments on two benchmarks demonstrate the superiority of proposed approach over state-of-the-art methods.