Research On Person Re-identification Technology Based On Data Augmentation And Domain Adaptation

With the rapid development of deep learning,person re-identification(Re ID)technology has made great progress,which aims to retrieve the identity of interested person from multiple non-overlapping cameras.However,on the one hand,as it is very difficult to obtain and label large person re-identification datasets,the further improvement of recognition performance is limited;on the other hand,due to problems such as inconsistent distribution among datasets,the performance of the model trained in the source domain is greatly degraded when applied to other domains.Therefore,it is of great significance to study how to realize person data augmentation and solve the model domain adaptive problem.This thesis is a research on person re-identification technology based on data augmentation and domain adaptation.The main innovative work is as follows:(1)In order to solve the problem that Cycle GAN produces a lot of redundant noise when generating cross-camera style images,we propose an improved Cycle GAN network,which can generate rich enhanced pedestrian images.While transferring the image style,it adds pedestrian posture constraints,which not only increases the diversity of pedestrian posture,but also reduces the style differences between different cameras.In addition,multi-pseudo regularized label(Mp RL)is used to dynamically assign virtual labels to the generated images.Simulation experiments are carried out on the widely used person re-identification datasets,and the experimental results verify the superiority of the proposed method.(2)In order to solve the problem of distribution difference of images between source domain and target domain,a two-branch dynamic auxiliary contrastive learning(DACL)framework is proposed.By dynamically reducing the local maximum mean discrepancy(LMMD)between the source domain and the target domain,the framework can effectively learn the invariant features of the domain,so that the trained network can identify the person identity of the target domain more accurately.At the same time,in order that the proposed network can adaptively aggregate the important features of the image,the generalized mean(Ge M)pooling is used to aggregate features after feature extraction.Compared with other advanced unsupervised domain adaptation models for person re-identification,the proposed two-branch dynamic auxiliary contrastive learning method has achieved the best results than many other Re ID methods.(3)In order to solve the problem that there are different domain gaps among multiple source domains in multi-source domain unsupervised domain adaptation and the improvement caused by jointly training of simply combining multiple datasets is limited,a multi-domain contrastive learning(MDCL)method based on exact feature distribution matching and multi-domain information fusion is proposed.First of all,in order to retain more information and enhance cross-distribution features,our proposed method achieves accurate feature distribution matching through accurate histogram matching based on sorting algorithm,so as to obtain more diversified feature enhancement.Then,the MDCL method constructs the knowledge graph based on the extracted features,and then fuses the multi-domain features through the two-layer residual graph convolution network.At the same time,the MDCL method also uses hybrid memory to update target domain clustering labels,and uses the self-paced comparative learning for iterative optimization.Simulation experiments are carried out,and the experimental results show that the proposed method has got the best performance on the widely used person re-identification datasets.