Research On Attribute-identity Embedding And Discriminative Dictionary Learning For Cross-dataset Person Re-identification

Person Re-IDentification(PRID),as one of the key tasks of intelligent video surveillance analysis,can automatically match pedestrian images from multiple camera perspectives.Due to its importance in intelligent monitoring,pedestrian tracking,pedestrian image retrieval and other aspects,it has received extensive research attention in recent years and has made remarkable progress.However,the training and testing of most existing methods are usually supervised and usually carried out on a single labeled dataset.Due to the differences between datasets,if we directly deploy the model trained on one dataset to another dataset,this may lead to a significant decrease in recognition accuracy.This means that the supervised single-domain PRID method has limited scalability in practical applications,and often lacks data-specific labels in practical applications.By reading a great deal of literature,summarizing the existing methods,we find that the scalability of the existing methods is poor mainly due to the following two problems: 1)Due to the large distribution differences between different datasets,the model trained on one or more datasets will be directly applied to the test dataset,which will cause serious domain shift problem;2)Due to the lack of pedestrian label information in the test dataset,the recognition performance of this kind of method is much lower than that of supervised learning method.Therefore,this paper makes correlation research on the above two problems and obtains some research results as follows:(1)To solve the problem of scalability,we propose a self-supervised learning algorithm based on attribute-identity embedding,which can incrementally optimize the model by selecting unlabeled samples from target domain.Thus the gap between source domain and target domain is bridged.Specifically,we first develop an attribute-identity joint prediction dictionary learning model for simultaneously learning a latent attribute dictionary,a semantic attribute dictionary and an identifier.In our method,the predicted attribute from latent attribute space is used as a bridge to establish a preliminary link between different domains so as to predict the label of the target data sample.Second,to exploit the latent label contained in the predicted samples,we propose a predictiontraining cycle self-supervised learning to tune the model variables to make them more adaptive in the target domain.Finally,the similarity measurement of pedestrians is achieved by combining the attribute space with latent identity space.(2)In order to eliminate the influence of domain information and pedestrian pose information,we propose a dictionary learning algorithm based on matrix factorization and hypergraph structure alignment.Specifically,Our model mainly includes two novel components:(a)Based on the idea of matrix factorization,we decompose the original visual features into pose-invariant components,domain information components and interference information components,in order to extract visual components that are not affected by domain information and pedestrian pose information,and further eliminate the influence of interference components between pedestrians on recognition;(b)In order to further improve the adaptability of the algorithm,considering the domain invariance of semantic attributes,we introduce hypergraph structure alignment constraints to establish the relationship between pose-invariant features and semantic attributes,so as to accurately predict the pedestrian attributes of the target dataset in the later period,and finally combine the pose-invariant features and semantic attributes of pedestrians to measure pedestrian similarity.