Research On Pedestrian Re-identification Method Based On Cross-perspective Discriminative Dictionary Learning

Person re-identification refers to match the person's images captured from disjoint camera views.Due to the complexity of camera surveillance scenes,Person reidentification is still a great challenge.(1)Under different camera perspectives,the same pedestrian images have posture differences illumination changes,background clutter,and occlusions,it have led to significant dissimilarities in the appearance of features of the same persons in complex surveillance scenarios.(2)Most existing methods pay attention to the fact that different pedestrians often show great similarities,resulting in lower recognition rates of current method.The main contents of this paper are as follows:(1)We introduce the idea of metric learning into the dictionary learning framework so as to project the original features of the pedestrian image into a discriminative feature subspace,and then a joint learning model integrating projection transformation and dictionary learning is designed.The model can simultaneously realize the discriminative projection of original features and the learning of discriminative dictionary,thus achieving the goal of integrating the advantages of metric learning and dictionary learning.(2)To effectively reduce the ambiguity between the coding coefficients of different pedestrians,a stretch regularization is designed and utilized to regularize the solution space of the coding coefficient of pedestrian images.As a result,the distance between the coding coefficients of different pedestrian stay far away from each other.(3)In order to fully exploit the label information of labeled samples,we develop label consistency constraint term in the dictionary learning model,and construct a joint learning model of identity discriminator,projection transformation and discriminative dictionary,and design a pedestrian similarity measurement scheme by integrating the identity labels and the coding coefficient.Experimental results on four popular person re-identification benchmarks VIPeR,PRID450 s,PRID2011,GMUL-GRID indicate that the approach developed in this thesis has higher identification performance than traditional method.