Research On Person Re-Identification Based On Feature Distribution Consistency And Credible Sample Mining

Person re-identification(re-ID)aims to retrieve person images with specified identities from images captured by many non-overlapping cameras,and to identify and track person images.At present,person re-identification technology has been widely used in the fields of intelligent monitoring and social security.Benefiting from the rapid development of deep learning algorithm and relying on large-scale labeled data to train the model,supervised person reidentification research has achieved excellent performance.However,in the actual scene,most images often lack labels as supervise information,and labeling large-scale images manually is often time-consuming and laborious.At the same time,due to the style differences of datasets in different domains,a model that performs well in a single-domain dataset will inevitably have a sharp decline in model performance when it is transferred to other domain datasets,which also increases the difficulty of person re-identification in the new data domain.To this end,unsupervised person re-identification methods that are more in line with the requirements of practical applications have emerged.Their purpose is to solve the person reidentification task in unlabeled scenarios,which has attracted more attention from researchers in recent years.Current unsupervised person re-identification research is mainly divided into unsupervised domain adaptation and fully unsupervised learning methods.The unsupervised domain adaptation method uses the labeled source domain to pre-train the model to give the model a certain feature learning ability,and then fine-tunes model through the unlabeled target domain to make model better adapt to target domain samples.Fully unsupervised learning methods only rely on unlabeled data to perform clustering algorithms or K-Nearest Neighbor search to explore the potential relationship between samples in the feature space.In this paper,the above two person re-identification methods are studied,and the main research contents are as follows:(1)This paper proposed an unsupervised domain adaptation person re-identification method based on median center and credible sample mining.This method selects the median center point by calculating the overall sample distance in the cluster,uses the median center as cluster center,and reduces the influence of outliers on cluster results.At the same time,the K-Nearest Neighbor samples of the anchor are analyzed to judge the confidence of anchor,and the samples with high confidence are selected for the training of model to avoid further amplification of pseudo-label noise in the training process.Finally,a new inter-cluster distance metric is proposed,which not only makes full use of features of all samples,but also pays attention to feature learning of credible samples,which increases the contribution of credible samples to the calculation of intercluster distance.Moreover,the introduction of the regularization term of sample distribution in each cluster controls the number of samples in each cluster to be balanced.The experimental results show that the proposed method can reduce the impact of false pseudo-labels on model training,and the overall method has better robustness.(2)This paper proposed a fully unsupervised person re-identification method based on multi-granularity features and distribution consistency.This method no longer relies on the labeled data of the source domain to pre-train the model,but only applies the unlabeled dataset to complete the model training.Firstly,local feature clustering is used to enhance the model’s ability to learn different features from a fine-grained perspective.The clustering of two different granularity features can improve the model’s ability to discriminate difficult samples.Secondly,two partial feature spaces and a global feature space were used for mutual credible sample mining to select credible samples in their respective feature spaces for model updating,reducing the mislead of false pseudo-labels on model training.Finally,the inter-cluster proxy contrast loss was introduced to alleviate the problem that the samples in the same cluster were more closely distributed when they came from the same camera.The loss function was used to reduce the differences in camera styles in the same cluster and alleviate the impact of camera differences on clustering caused by different camera fields of view.Experimental results demonstrate the effectiveness of the proposed method.