Cross-domain Person Re-identification With Unsupervised Domain Adaptation

Social security has always been one of the focuses of public attention.With the improvement of urban monitoring system,surveillance video has become an important means of maintain-ing public security.Person re-identification(re-ID)is a cross-camera image retrieval task that aims to find specific persons of a given query from the images datasets or video sequences,which can be widely applied in intelligent video surveillance and other fields.At present,most of the work focus on supervised learning in single domain.Although the impressive achievement of supervised learning in person re-ID,the models trained on one domain often fail to generalize well to another,which is the main obstacle to the practical application of person re-ID.In order to better apply person re-ID to practice,the problem of cross-domain person re-ID needs to be solved.A common approach for this problem is unsupervised domain adaptation(UDA).This the-sis studies cross-domain person re-ID based on unsupervised domain adaptation and deep learning.The main contributions of this thesis are summarized as follows:Frist,to address the problem that the domain styles of source domain and target domain are very different,we propose a domain style transfer model named c SPGAN,which is based on Cycle GAN network and adds a classifier to preserve the self-similarity during the transfer process.Using c SPGAN,we are able to generate images which not only possess the style of target domain but also preserve their underlying ID information.The experimental result demonstrates that c SPGAN can generate good translated images and compared with the ex-isting domain style model,the cross domain performance is better.Second,for the problem that the existing data enhancement methods need to train multiple models to achieve style translation of multiple cameras in dataset,in this thesis,the Star GAN method in the field of face recognition is applied to the person re-ID task.Based on Star-GAN,we only train one model to achieve style translation among multiple cameras,which can effectively expand the existing data set while improving the efficiency.Third,to address the problem that the target domain dataset has no labels and the model trained on source domain cannot adapt to the target domain well.We fully analyzes the data distribution on the target domain dataset and discover three underlying characteristics of the target domain data.Sample discrimination aims to push the different identities people away.Cam Style images consistency aims to pull the same person under different camera angles.Neighbors similarity aims to pull the sample and its nearest neighbor.Based on the above three characteristics,We propose a domain adaptation person re-ID framework named Unsu-pervised Joint Multiple Loss function network(UJML),which uses both the translated source dataset and target dataset with Cam Style images and adds the feature memory component to train the three characteristics of the target domain data and fuse them.In order to evaluate the performance of the proposed model,we make experiments on Market-1501,Duke MTMC-re ID and MSMT17.Under the cross-domain setting,the experimental result demonstrates that the proposed model has significantly higher rank-1 accuracy and m AP on the three datasets and outperforms the existing cross-domain person re-ID algo-rithm.