Research On Cross-domain Adaptive Person Re-Identification Algorithm Based On Deep Learning

Person re-identification aims to locate specific pedestrian from non-overlapping camera areas.It is an assistive technology for police handling cases and finding lost populations.With the development of deep learning technology and the introduction of large public datasets,the performance of supervised person re-identification methods based on deep learning has been greatly improved.The success of the supervised person re-identification method greatly relies on the real labels in large datasets.However,in reality,it is extremely time-consuming and labor-intensive to manually label a large number of person samples,which greatly hinders the development of the person re-identification technology.Therefore,the academia proposes an unsupervised person re-identification method,among which the domain adaptive person reidentification method has achieved the most advanced performance.Among the domain adaptive person re-identification methods,the pseudo-label predicting method had achieved the best performance.Such methods generally use the basic model which trained on the source domain and extract the person features of the target domain,then leverage the clustering algorithm to predicting the pseudo-labels for the unlabeled target domain dataset,and finally fine-tune the basic model based on the predicted pseudo labels until the model is stable.However,this kind of method will produce much outliers in the pseudo label predicting stage,and most of the work choose to discard the outliers directly.This will hinder the performance of the person reidentification model.Therefore,this paper designed a domain adaptive person reidentification method based on soft labels,which combines soft labels and pseudo labels to process outliers,and proposes a hardest sample selection strategy based on soft labels.Second,in order to solve the problem of limited multi-granularity features of the target samples extracted by existing models,this paper proposed a domainadaptive person re-identification method based on multi-granularity features and softlabels.In summary,the contributions of this paper as follow:(1).In order to mine the information of the outliers and improve the generalization ability of the person re-identification model,this paper proposed a domain adaptive person re-identification method based on soft labels.First,a new soft label generation strategy is proposed,which deals with the outliers through combine the soft labels and pseudo labels.Then a hardest sample selection strategy is proposed on the basis of soft labels.Finally,the global and local features of person are combined to improve the model’s ability to learning discriminative features of person images in target domain.(2)In order to mine the multi granularity features of the target domain samples and improve the generalization ability of the model,this paper proposed a domainadaptive person re-identification method based on multi-granularity features and softlabels.Based on a new sample augmentation method,multi-granularity samples are generated and combined with soft label strategy to provide more correct positive samples for model learning.Finally,train the basic network model by combining triple loss and Cross-Entropy loss.This paper experimentally verifies the two methods proposed in this paper on two large public datasets Market1501 and Duke MTMC-re ID.The experimental results show that the method in this paper achieves better results on both Rank-1 and m AP evaluation indicators.The method proposed in this paper can improve the generalization ability of the person re-identification model.