Cross-Domain Person Re-identification Based On Neighbourhood Discovery And Self-Supervised Learning

With people's increasing attention to public safety and the continuous improvement of so-cial informatization,thousands of surveillance cameras are installed in various public places.The traditional method of mass video surveillance only relying on manpower has huge cost and efficiency problems,so the research and deployment of intelligent surveillance sys-tem is urgent.As one of the key technologies in intelligent surveillance system,person re-identification(re-id)is a technology that utilizes computer vision technology to automat-ically judge whether there is a target pedestrian in an image.It also has a broad application prospect in unmanned supermarket,urban order management and other fields.In recent years,thanks to the development of deep learning,the performance of supervised person re-id has been greatly improved.However,in practical applications,if the dataset is ob-tained manually for each scene,it will consume huge resources.And directly deploying the model trained under a scenario to another scene,the performance will decrease significantly.Therefore,this paper focuses on unsupervised cross-domain person re-id.The main contents are as follows:1.In the cross-domain person re-id task,the images' labels are not available in the target domain.In order to mine discriminant information directly from the visual data,this paper proposes a cross-domain person re-id method based on neighborhood discovery.Firstly,the model is supervised by the source domain,so that the model can learn the general structural information of person images.Meanwhile,from the perspective of neighborhood discovery of local samples in the target domain,the neighbor points of each sample are mined accord-ing to the similarity,and the samples corresponding to the neighbor points are regarded as the same class to supervise the model updating.As the model is updated,the mining scope gradually expands from local to global,and the number of neighbor samples also gradual-ly increase.This is a strategy of divide and conquer,gradually increasing the number of intra-class samples from bottom to top.Experiments on Market1501 and DukeMTMC-reID demonstrate the effectiveness of this method.2.In order to mine more reliable category relation between images in the target domain,this paper proposes a cross-domain person re-id method based on progressive representation enhancement.The similarities of global and local features of images from the target domain are used to predict pseudo labels based on clustering method,and the samples with more reliable category relationship are selected for model training,so as to learn discriminant feature representation.Simultaneously,in order to mitigate the impact of different styles of images,captured by different cameras,the view-invariant representation learning is designed to reduce the distribution differences between sub-domains where each sub-domain are cap-tured from one camera.The above self-supervised learning process gradually enhances the discriminant of representation and improves the performance of the model.Experiments on three datasets Market1501,DukeMTMC-reID and MSMT17 prove the effectiveness of the method.