| In recent years,with the rapid development of the construction of "safe city" and"smart city",person re-identification technology has become increasingly significant in the field of public safety.However,traditional person re-identification technology rely on a large amount of manual annotation data,which brings great challenges for practical implementation.In order to overcome the limitations brought by traditional supervised learning,this dissertation conducts a series of researches on unsupervised person reidentification based on deep learning methods.The main work and contributions are summarized as follows:1.Aiming at the issue with the updating error of the exemplar memory in the unsupervised person re-ID,a novel unsupervised re-ID method is proposed based on central feature learning.First,an error memory module is designed to reduce the static error causedby sample noise in training.Second,in order to improve the stability of convergence,an error prediction module is incorporated on the basis of the error memory module,and the performance of the algorithm is further ameliorated by the principle of camera invariance.The experimental results on Market-1501 and DukeMTMC-relD demonstrate that the Rank-1 accuracy of the proposed method gets improved respectively by 11.9%and 2.1%compared to the traditional update method.2.Aiming at the problem with the representation learning enhancement in the unsupervised person re-ID,a novel unsupervised re-ID method based on momentum contrastive learning is proposed.The proposed method develops a learning framework consisting of a teacher encoder and a momentum-updated student encoder,then utilizes joint tuning of individual loss and clustering loss to obtain the better performance,and achieves a one-stage end-to-end training.Experimental results show that the method achieves Rank-1 accuracy of 86.5%,74.5%and 36.6%on Market-1501,DukeMTMC-relD and MSMT-17 datasets,respectively.3.Aiming at the occlusion problem in unsupervised person rc-identification,a pose-guided occlusion-based unsupervised person re-identification method is proposed.First,the backbone network is used to extract global features,and then a pose-guided branch is introduced to extract keypoint information of pedestrians.Further,the global features of pedestrian images extracted by the backbone network and the pose feature information extracted by the pose guidance branch are fused to realize the focus on nonoccluded areas.Experimental results show that the proposed method achieves Rank-1 and mAP of 53.2%and 54.2%on Occluded-DukeMTMC,respectively.4.Aiming at the adversarial robustness of the unsupervised person re-ID model,the anti-interference performance of the existing models is analyzed and a robustness enhancement method for unsupervised person re-ID is proposed.First,the existing model is attacked by using multiple adversarial noise samples.Then,adversarial training strategy and feature denoising module for unsupervised person re-ID are proposed to overcome the impacts of adversarial attack samples on the model.Experimental results show that the Rank-1 accuracy of the proposed robustness-enhanced model achieves a significant improvement of 10%～30%. |
