Multi-view Learning Based Person Re-identification

With the further development of major projects such as "safe city" and "sapiential city",the awareness of public safety sinks deeper and deeper into people's hearts.The rapid development of intelligent video surveillance technology makes it easier to monitor social public safety in real time.The problem of intelligent surveillance video retrieval for specific pedestrians has become a research hotspot in the field of computer vision,and it's known as person re-identification,which refers to retrieval for specific pedestrians under non-overlapping monitoring equipment.However,person re-identification is facing a huge challenge,which is due to the change of light,view,posture and other factors.As a consequence,the appearance characteristics of the same person have undergone significant changes.Distance metric learning.and sparse representation are widely applied to person re-identification.This paper combines these two methods and multi-view learning method,and the differences between the features of person images and the different influence on matching is fully taken account of.The main research content of this paper is as follows:1.A multi-view kernelized feature based metric learning algorithm is put forward.Firstly,the limitation of the existing impostor-based distance metric learning method is sufficiency analyzed,i.e.,the constraints of the negative matching pair and the positive matching pair is not perfect.On this basis,a distance metric learning of symmetric impostor constraint is proposed.Then,combined with kernel method and multi-view learning,the original linear inseparable feature is mapped to a non-linear feature space with better separability,and different distance metric learning of symmetric impostor constraint is learned for global feature view and local feature view.Finally,we consider the different effects of different feature views on recognition matching,and learn different weights of recognition matching for different feature views,matching is carried out by weighted distance method.2.A multi-view feature projection and semi-coupled projective dictionary learning algorithm is put forward.The original feature spaces of the person images taken by different cameras have inconsistent distributions,and the original feature spaces also contain a certain amount of noise information.In order to deal with the difference between the different features and the difference between the feature space of different cameras,the different feature projection matrix and semi-coupled projective dictionary are learning for the original feature spaces of two cameras corresponding to each feature view,the feature projection and the semi-coupled projective dictionary ensure that the differences between the cameras and the noise information of the samples are reduced in the new feature space.Finally,different weights are assigned to different feature views,and multi-view feature conversion and weighted distance are used in matching.The experimental results on three databases,such as CMC curve and training time,show that the methods proposed in this paper can obtain better recognition accuracy than other well-known methods while keeping lower time complexity,and it is suitable for person re-identification.