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

Person re-identification(ReID)is designed to retrieve images of specified pedestrians in a large database given a query object.In recent years,with the emergence of large-scale datasets,and the continuous improvement of feature extraction and metric learning methods,the researches of person ReID under the single-domain setting has achieved significant improvement.However,the performances of these single-domain trained models will be greatly reduced when these models are directly deployed to the real-word environment with a large-scale camera networks.To address this problem,this paper conducts a study of cross-domain person reidentification methods based on deep learning.The content mainly considers the adaptive ability,the model size and the accuracy of cross-domain person ReID by using deep learning technology and related algorithms.First,a cross-domain person re-identification method for enhancing domain adaptability is proposed.It combines the changes of source and target domains to enhance the domain adaptability.Secondly,a lightweight and accurate unsupervised method is proposed.This can improve the accuracy of person re-identification and reduce the size of models.The contributions of this article in academic and practical applications are as follows:1.To solve the problem of poor adaptability of the single-domain training model,a deep Re-ID model based on region alignment and "softening" processing is proposed for crossdomain tasks.Firstly,to collect part features,the feature map is partitioned into P regions horizontally and uniformly by the region alignment strategy.Secondly,the P regions are "softened",and the outliers generated in each region are re-zoned to neighboring regions to improve the consistency within the region.The alignment of the regions after "softening" plays an important role in cross-domain person re-identification.By strengthening the alignment of the model,on the one hand,the generalization performance of the model is improved,and the performance of the model directly across-dataset is improved.On the other hand,the region alignment model can easily utilize the unlabeled data in target domains to achieve the adaptation from source domains to target domains,so that the model can adapt to target domains.2.To further enhance the adaptability of the cross-domain person ReID model,an algorithm combining the changes of inter-domain and intra-domain is proposed.This could reduce the impact of cross-domain from two aspects: inter-domain and intra-domain.Firstly,the deep Re-ID model based on region alignment and "softening" processing is trained by using labeled source domains data.Then,for the inter-domain changes from source domains to target domains,pose-invariance between the domains is proposed.New person images are generated when person poses in source domains are learned by person in target domains.These images are added in train to reduce the gap between sources and target domains.After that,for the changes in the target domain,the pose-invariance in a domain,sampleinvariance,camera-invariance,and neighborhood-invariance are used.The pose-invariance in a domain learns the pose of different person in target domains to reduce the pose gap.The exemplar-invariance enforces each exemplar away from each other,and to enlarge the distance between exemplars from different identities.The neighborhood-invariance encourages each exemplar and its neighbors to be close to each other.It is beneficial to reduce the distance between exemplars of the same identity.The camera-invariance has the similar effect as the exemplar-invariance and also leads the exemplar and its camera-style transferred samples to share the same representation.Finally,the exemplar memory is introduced to store features of the target domain and accommodate the invariance properties.Experiments demonstrates that region alignment and five invariance properties effectively enhance the domain adaptability in cross-domain person ReID.At the same time,its recognition accuracy is effectively improved compared to focusing only on inter-domain changes or intra-domain changes.3.To solve the problem of low accuracy of cross-domain person ReID and large training model while enhancing domain adaptation,a lightweight and accurate unsupervised method is proposed.Firstly,a lightweight image classifier(LWC)is proposed.It combines Dense Net’s convolutional layer and batch normalization layer to simplify its model structure.Then the spatio-temporal information of datasets is added to form a powerful classifier.Secondly,an optimization learning algorithm is proposed.A feeble classifier(LWB)is implemented by constructing a binary network.For the training of the network,the ranking information of a powerful classifier is used to enhance the recognition capability of the LWB.Furthermore,spatio-temporal information of datasets is added again to construct LWA classifier.The results show that the model size of LWB is about 80% smaller than that of the Siamese network model on the CUHK01 and VIPe R datasets.At the same time,the recognition accuracy of the LWA classifier is slightly improved on the Market1501 and Duke MTMC-re ID target datasets.In summary,this paper researches the cross-domain person ReID task by deep learning.Related experiments and analysis are conducted from models,inter-domain information,and intra-domain information.