Research On Person Re-identification Based On Deep Learning

In public places,dense crowds are prone to public safety incidents.Therefore,a large number of surveillance cameras are installed and applied in various places of city.In traditional way,identification of pedestrians by human observers from surveillance images is a time-consuming task and cannot be performed on large scale.Person re-identification is a key component technology in urban monitoring,which is used in public places to identify pedestrian image.Person re-identification is the method of querying target pedestrian image in video or images taken by different cameras.Most current person re-identification methods are mainly researched in two areas: supervised and unsupervised.For most current supervised person re-identification methods,some inconspicuous details are easily ignored when extracting global or local features of pedestrian images.The neglect of some detailed features can affect the result of person re-identification.Most of supervised person re-identification methods show their excellent performance,but using labeled datasets is very expensive which limits its application in practical scenarios.This paper proposes corresponding methods for supervised and unsupervised person re-identification,respectively:(1)This paper proposes a Multi-level Feature Fusion model,which uses deep learning networks to extract the global and local features of pedestrian images and combines global and local features to generate more discriminative pedestrian descriptors.In Multi-level Feature Fusion model,Part-Based Multi-level Net is used to extract local features of different depths of network and combine local features extracted from shallow to deep layers of the network Global-Local Branches extract the local and global features at the highest level of the network to identify pedestrians.(2)This paper proposes an unsupervised Erased Feature Discrimination Learning framework which can learn discriminative features from the whole picture and local details of the entire image.We develop Erased Feature Discriminative Learning method,which learns pedestrian discriminative features by forcing EFDL network to learn similar local details of pictures.At the same time,in order to solve the problem of gap between pictures with same identities under different cameras,we use generated migration pictures to provide effective guidance for the network.In addition,we design unsupervised discriminative feature learning loss functions to learn discriminative global and local features on unlabeled datasets.A large number of experiments show the superiority of the proposed methods in supervised and unsupervised person re-identification.