| Unsupervised cross-domain person re-identification aims to use the source domain training model to achieve the matching of specific person under the test set of the target domain.It is one of the challenging tasks in computer vision.Due to the large differences in the scenes of the source domain and the target domain,the light intensity of the shooting point,and the pedestrian dressing style,etc.,that is,the domain shifts,person re-identification method fitted to the source domain directly applied to the target domain introduces significant performance degradation.To solve this problem,depending on whether to introduce unlabeled target domain data to participate in training,this thesis investigates unsupervised cross-domain person re-identification from two perspectives,domain adaptation and domain generalization,respectively,in order to obtain better domain migration performance.The main work is as follows:(1)Aiming at the problem that domain adaptation methods based on clustering and fine-tuning only use the global features of person for clustering and ignore the person’s local features,this thesis proposes a learning feature fusion methods for unsupervised domain adaptive person re-identification to obtain better features for clustering.This method adopts the teacher-student network framework,and for the input target domain data,the output features of the teacher network with higher reliability are used for clustering,which alleviates the noise pseudo-label problem caused by the single network structure.At the same time,by embedding a feature fusion module with learning capabilities between the teacher and student networks to obtain fusion features that can not only pay attention to the global information of person but also highlight local fine-grained information.It solves the ambiguity learning problem of using different features to cluster independently.(2)For the problem that no target domain data are involved in the training,this thesis proposes a domain generalization person re-identification method based on implicit distribution alignment.This thesis firstly designs a domain normalization layer,which removes domain style information and extracts discriminative identity information by combining the distribution information of different normalization layers,effectively reducing domain shifts.Then,an implicit distribution space is constructed using multiple source domains and the distribution of each source domain is constrained to map into this space.In the training process,each source domain is regarded as the ”simulated target domain” of the remaining source domains,and the distribution of the ”simulated target domain” is also mapped into the implicit distribution space through an anti-normalization layer proposed by this method.In this way,the model learns to align the distribution of the source domain and the target domain which solves the problem that the target domain data cannot be used for training.(3)The two unsupervised cross-domain person re-identification methods proposed in this thesis adopt a variety of cross-domain experimental settings on multiple person re-identification datasets,and their performance has achieved the state-of-the-art performance.Taking the Duke MTMC-Re ID and Market1501 datasets as the target domain for domain adaptation experiments,the proposed method achieves 83.2% and 73.5% in m AP metrics respectively,which is 2.2% and 2.7% improvement compared to the current stateof-the-art methods,verifying the excellent cross-domain capability.In addition,experiments on a person re-identification dataset collected in campus prove the effectiveness of the methods in practical scenarios. |
