Research On Cross-domain Person Re-identification Methods Based On Deep Mutual Learning

Person re-identification(re-ID)has been a popular research topic in computer vision and artificial intelligence.It is mainly used to determine whether pedestrians taken at various times with different cameras are the same person,which is a sub-problem of image retrieval.With the development of digital cities and smart cities,person re-identification has become widely employed in video surveillance,criminal investigative security,intelligent security,and other industries.Object occlusion,illumination change,posture change,and shooting angle change are the common challenges in person re-identification.With the widespread use of deep learning,person re-identification methods based on deep learning have provided new ideas to overcome these difficulties and have made significant progress.In particular,the person re-identification methods based on unsupervised domain adaptation can achieve better performance in the real scene.These methods are evaluated using unlabeled data in the target domain after training in the source domain and fine-tuning in the target domain.Although the accuracy of recent research methods is progressively increasing,the performance of person re-identification methods might be affected by the noisy pseudo-labels during training and testing.In order to solve the above problems,the main works of this thesis contain the following two aspects:(1)This thesis proposes a cross-domain person re-identification method based on three-branch mutual learning.This method uses a three-branch network structure,through mutual learning and collaborative training among the three branches,so as to evaluate the accuracy of the prediction result of the input sample,and select high-confidence samples to update the model.Specifically,we pre-train the initial model in the source domain to obtain a pre-trained model firstly,and then use a three-branch network structure to train in the target domain.Each branch contains a pre-trained model and an average network model to extract features and predict pseudo-labels for the unlabeled samples in the target domain.Finally,the features extracted from the average network model in the three branches are calculated by using the cosine similarity in pairs,and the high-confidence samples,whose cosine similarities are lower than the threshold,are selected to update the network in the third branch.Three different cross-domain tasks are carried out on three large-scale datasets commonly used for person re-identification.Compared with state-of-the-art person re-identification methods,the proposed method has better performance in several evaluation indicators.(2)In order to further reduce the impact of noisy pseudo-labels on cross-domain person re-identification,this thesis proposes a cross-domain person re-identification method based on self-paced collaborative learning.The difficult sample filtering classifier is used firstly to cluster the unlabeled samples in the target domain through a density clustering algorithm,and the difficulty degrees of the samples are decided according to the distance between the samples and the cluster center in the cluster space,so that the samples are input into the person re-identification model from easy to difficult.It is beneficial to alleviate the interference of noisy pseudo-labels.In addition,by combining the two methods proposed in this thesis,a three-branch mutual learning method fusing the difficult sample filtering classifier can be obtained,so that the model can effectively utilize the unlabeled samples in the target domain,reduce the influence of noisy pseudo-labels,update the model with more accurate pseudo-labels,and improve the performance of the person re-identification model.The proposed method is tested on several challenging datasets,and the experimental results verify that the proposed method is effective and robust.