Research On Unsupervised Pedestrian Recognition Based On Generative Adversarial Model

With the rapid development of social security technology,intelligent video surveillance networks are widely used in various public areas,such as roads,airports,train stations,shopping malls,etc.Among the technologies related to intelligent video surveillance technology,pedestrian re-identification is one of the important key technologies.The goal of pedestrian re-identification is to retrieve pedestrians of interest under multiple non-overlapping cameras.With the development of deep neural networks and the growing demand for intelligent video surveillance,pedestrian re-identification has received widespread attention and development in the field of computer vision.However,the pedestrian re-identification still needs difficulties and has not been resolved.For example,the cross-domain adaptation problem of pedestrian re-identification,the goal of the solution is how to deploy a model trained in a data domain to an unlabeled data domain.This thesis mainly studies the unsupervised cross-data domain adaptation of pedestrian re-identification tasks.Based on the above research background,this thesis conducts research on pedestrian re-identification.By analyzing the research status of pedestrian re-identification technology at home and abroad,a new cross-domain adaptive pedestrian re-identification algorithm is proposed for the difficulty of cross-domain adaptation.Outcome:(1)This thesis proposes a pedestrian re-identification method with many-to-many sub-domain adaptation,to solve the problem of cross-domain adaptation in pedestrian re-identification tasks.In order to reduce the data distribution difference between the source data domain and the target data domain,A generative adversarial model is used to transform pictures of different sub-domains of the source data domain to different subdomains of the target data domain,then with supervised training the transformed source domain data,the adaptation of the source data domain model to the target data domain is completed.Based on Market-1501,DukeMTMC-reID's experiments have achieved results that compete with current cross-domain methods.(2)A cross-domain adaptive local feature learning method is proposed in this thesis,to solve the problem of local feature extraction in pedestrian re-identification tasks.In order to be able to extract local features of pedestrians,a method of random local feature learning is proposed.While improving the performance of the model,the amount of parameters of the model is reduced,and the difficulty of training the model is reduced.Experimental results on Market-1501,DukeMTMC-reID and CUHK03 show the effectiveness of our method.