Research On Person Re-identification Method Based On Unsupervised Domain-invariant Feature Learnin

Person re-identification is a very popular research topic in the field of computer vision,and it plays an important role in video surveillance,intelligent security and other fields.Due to the difficulty of pedestrian dataset labelling,currently only unsupervised person re-identification can be applied to large-scale real world applications.However,the distributions of images under different cameras are inconsistent due to the different imaging factors such as shooting angles,person poses,and lighting in different scenes or perspectives.As a result,the performance of the unsupervised model is not ideal.We have improved the traditional unsupervised person re-identification model based on dictionary learning.In addition to learning an over complete visual semantic dictionary,we also introduced asymmetric projection to eliminate the distribution deviation between different camera data and improve the discriminative power and robustness of the overall model.In order to take full advantage of the non-linear transformation of the input space by deep features,and the person discrimination information carried by the labeled data in the source domain,we based on the convolutional neural network and the latest domain transfer scheme to learn an embedded encoding space that shared by all cameras and jointly learning global features and local features in it.The specific research contents are as follows:(1)Most traditional unsupervised person re-identification methods do not consider the negative impact of camera distribution divergence on the model.This thesis combines projection subspace and dictionary learning,and simultaneously introduces multi-view features for performance expansion to provide a robust ensemble model.First,the data under different cameras is mapped into a common subspace through asymmetric projections,then the error reconstruction term of dictionary learning is introduced,and the learned dictionary encoding is constrained by a visual similarity weight matrix,which is called Laplacian regularization term,and the final recognition effect of the model largely depends on the accuracy of the similarity matrix.The visual similarity matrix of a single view is not very reliable,so a variety of view features are introduced,and a simple weak model is learned for different views to obtain the final ensemble matching result through score fusion.(2)Existing deep models fail to effectively solve the impact of data domain bias and target domain pseudo-label noise on the model.This thesis combines domain-specific batch normalization with camera-based batch normalization to force all camera images from each domain projecting into the same subspace,which can effectively align the data distribution between all cameras in the two domains.Then in this space,the target domain is subjected to contrast learning based on clustering pseudo-label generation,and further a contrast reference space is proposed,where source domain person category features act as experts to determine whether target domain person images have some specific visual-semantic features,thereby correcting identity labeling errors caused by target domain clustering to reduce noisy labels.In addition,since the coarse-grained holistic appearance of pedestrians is not conducive to the mining of positive and negative sample pairs,i.e.,identifying different identities with similar appearances and the same pedestrian with different appearances.We utilize offline pose detectors to crop different regions from pedestrian images,and then perform nearest neighbor based similarity learning respectively according to them to achieve fine-grained feature learning,thus establishing our multi-granularity unsupervised re-identification framework.