An Unsupervised Person Re-identification Method Based On Multilevel Balanced Clustering

With the development of social science and technology,people pay more and more attention to public security.A large number of public places need to be monitored for security protection.As a key technology in public monitoring and security application,person re-identification is of great significance to improve public security.The main task of person re recognition technology is to utilize computer vision technology to retrieve pedestrian images from different cameras through the given pedestrian images.However,due to the privacy issues and the difficulty of data annotation,it is difficult for the person re-identification method based on supervised learning to obtain enough training data.Therefore,in recent years,unsupervised person re-identification algorithm has attracted the attention of researchers.Unsupervised person re-identification algorithm can spontaneously learn the common characteristics of the same person in different camera perspectives from unlabeled datasets,which can get rid of the requirements of a large number of labeled data.Most of the current unsupervised person re-identification algorithms are based on the domain adaptive method,which requires additional datasets,and the results are still far behind those of the supervised algorithm.Therefore,in this paper,we propose an unsupervised person re-identification method based on multilevel balanced clustering,which can fully mine the potential identity information in the unlabeled dataset without any additional samples,and improve the accuracy of re-recognition.The main contributions of this paper are as follows:Firstly,in this paper,we analyze the distribution of the data in the person re-identification task,and propose an unsupervised training algorithm based on the balanced criterion clustering according to its multi-centered distribution.In this method,the features extracted by the deep neural network are clustered to generate pseudo labels,and these pseudo labels are used to further train the network.For the clustering stage,we design a balanced clustering criterion.Under this criterion,the similarity between clusters are measured by the combination of average distance and centroid distance,so that the final clustering results have both good integrity and homogeneity.This method improves the accuracy of subsequent generated pseudo labels and reduce the error caused by the pseudo labels on the network training.Secondly,in view of the irreversibility and accumulation of misleading caused by false labels in the whole learning process,this paper proposes a multilevel label guided network learning strategy.In this strategy,the features extracted from images of different spatial levels are clustered to generate pseudo labels of different levels,and the pseudo labels of different levels are used to guide the network training.In this paper,we compare the results of multi-label and single label,it is expounded that multi-level label can reflect the multiple relationships of data samples,and provide complete supervision information for learning,which is helpful to improve the fault tolerance of the system.In summary,the proposed unsupervised person re-identification method based on multilevel balanced clustering can achieve accurate recognition results in cross camera pedestrian retrieval task,which has certain theoretical research value and practical application value.