Research On Person Re-identification Methods Under Different Supervisions

Deep learning approaches have demonstrated their capability in resolving multiple machine learning tasks,such as image processing and natural language processing,and achieved state-of-the-art performances,which draws intense attention from both academia and industry.Deep learning studies a large range of pattern recognition systems with multi-layer nonlinear transformations.These systems abstract data from low to high levels,and are more suitable for highly nonlinear problems in the real world than the shallow approaches with less layers.With the increasing attention to social public safety and the development of video capture technology and large-scale data storage technology,there is an increasing demand for automated and intelligent analysis of video content under large-scale surveillance systems.Person re-identification is an important research topic in the field of intelligent monitoring.It mainly studies the identification of a specific person in the surveillance video who has already appeared in the monitoring network.The research of this problem can provide basic technical support and feasibility guarantee for many applications in the field of intelligent monitoring,for example,pedestrian tracking.This dissertation focuses on the issue of person re-identification,works under different supervision constraints,from full supervision to semi-supervised and then unsupervised,and carries out different levels of research and exploration.Multiple machine learning techniques such as convolutional neural networks(CNN),model regularization technique and clustering have been studied and researched in this work.The specific research content and innovations of this dissertation mainly include the following aspects:(1)Under the setting of full supervision,this dissertation proposes a complementary discriminative feature extraction algorithm to deal with the shortcoming of losing local features by the single-branch network.By designing a Siamese-like feature extraction network architecture,the feature mask module is used to connect the main branch and the other branch,thus features from the main branch can be used as a part of input to the other branch's for complementary feature mining.On the basis of this,an inter-branch pairwise ranking loss function is added on top of the network for training,which promotes the diversity and complementary characteristics between two branches.Experimental numerical results and visual representations show that the proposed method significantly improves the accuracy of person re-identification.(2)Under the setting of semi-supervision,in order to combat the problem of “overfitting”in the person re-identification task whose datasets only contain limited numbers of annotations while deep learning network are data-hungry.This dissertation proposes a method for generating pseudo-labels for pseudo-data to perform model regularization.A pseudo-label is assigned to an unlabeled sample according to its similarity to the labeled person representations in the feature space.The pseudo-label generation criterion based on the feature representation can provide two schemes of label encoding under the unified framework,i.e.,the one-hot scheme and the distributed scheme.Experiments were conducted on public datasets to demonstrate its competitive performance to its counterparts.(3)Under the setting of no supervision,the clustering and merging criterion in the existing person re-identification clustering methods can not achieve satisfactory clustering results.This dissertation proposes to use dispersion as a cluster selection and merging criterion.By simultaneously measuring the inter-cluster dispersion and intra-cluster dispersion,the results of clustering can be optimized to boost person re-identification performances.The criterion based on dispersion not only can automatically raise the clustering priority for isolated points,but also prevents the formation of bad clusters.In addition,the method can play a mutually reciprocal promotion role with the convolutional neural network,leading to faster convergence and better stability.State-of-the-art person re-identification performance has been achieved on both images and video based datasets.