Person re-identification is usually affected by background debris or external occlusion.Therefore,it is difficult to quickly and accurately distinguish different pedestrians between disjoint camera views.This paper presents a hybrid scheme of scalable deep feature learning architecture,which regards pedestrian re-recognition as a learning di stance measurement problem,and applies symmetric regularization to help train deep neural networks with relative distances.First ,combined with the latest development of Convolutional Neural Networks (CNN),its main framework triple model is used to extract a robu st representation,which aims to solve the problem of single image extraction descriptors.Secondly,the triple loss and the center los s are combined,and the gradient weighting algorithm is used to update and optimize the network weights and parameters to overcome sample imbala nce.In addition,the introduction of a symmetric regular term modification optimizes the back-propagation of the asymmetric gradient der ived from the relative distance metric,which minimizes the intra-class distance and maximizes theinter-class distance in the triple unit, Inorder to overcome the "angle deviation" problem in pedestrian re-identificati on.The results prove that in the same environment,the matching effect of pedestr ian re-recognition on cameras that do not intersect in space issignificantly im proved.