Research On Cross Domain Person Re-identification Based On Unsupervised Domain Adaptation

As a sub-problem of image retrieval,person re-identification aims to retrieve interested persons from cameras with non-overlapping views.Person re-identification technology has important research and application value in the fields of security,criminal investigation and so on.With the increasing number of surveillance cameras in public places,there is a massive growth in video surveillance data.It is too expensive to label these massive images across cameras,so supervised person re-identification will be limited in practical application.At the same time,because the person re-identification data sets are collected from different environments,there will be domain differences in the data distribution between the data sets collected in different scenes.If the model trained on one data set is directly applied to another scene,the model recognition performance will be greatly reduced.The method based on label estimation to solve the domain adaptation problem has better performance than other methods,but this method only considers the global features of the target domain image,ignores the local features with fine-grained information,and the generated labels have noise.To solve the above two problems,this thesis proposes two models:(1)A cross domain person re-identification model based on global and local feature fusion is proposed.Firstly,IBN module is integrated into Res Net50 network to improve the adaptability of the model to image appearance changes.On this basis,the person feature map is evenly divided into two parts,the local features of the person image are extracted,and fused with the global features of the person image to improve the discrimination ability of person features.Then,the re-Ranking technology based on k-reciprocal neighbor is used to optimize the sorting of the samples in the target domain,and the pseudo labels are generated accordingly.Finally,the pseudo labels are used as the supervision information,and the multi-loss function strategy is used to optimize the model,iterating until the model converges.Experimental results on two public data sets show the effectiveness of person global and local feature fusion.(2)Aiming at the problem that the hard label generated in the label estimation method has noise label,a cross domain person re-identification model based on DBSCAN clustering and collaborative learning is proposed.Using the idea of deep mutual learning network and average teacher network for reference,the predicted value generated by the teacher network is used as a soft label to supervise the learning of each other’s student network,so as to ensure the output independence of the two student networks,so as to avoid the amplification of errors in the training process of the network model.Then,the loss functions corresponding to soft label and hard label are used to optimize the network model at the same time to reduce the interference of noise label on model training.Compared with other experimental methods,it is verified that this model can improve the generalization ability of the target domain model.