Research On Unsupervised Domain Adaptation In Person Re-identification

Security systems based on video cameras play important roles in modern social security.Most of the existing security systems are face recognition systems,which can only be used in several limited scenes.However,most of the cameras in the city are open-air cameras.Due to the obstacles of person's walking route,posture,shooting angle of camera and other factors,it is very difficult to obtain complete face information.Generally,only the contour information of pedestrians can be obtained,so person re-identification technology arises as the supplement of face recognition.With the birth and wide application of convolutional neural network,social security requirement and "smart city" construction,researches on person re-identification develop rapidly.This thesis focuses on the unsupervised domain adaptation method of person re-identification technology,and its application to the existing person re-identification datasets and self-recorded video.Firstly,the feature extraction network in the model is improved.This thesis investigate the Multilabel Reference Learning(MAR)network whose feature extractor,Res Net-50,is replaced by Batch Drop Block(BDB)feature extractor applied in supervised method to improve the recognition rate of the original network.Meanwhile,in order to make BDB feature extractor fit the unsupervised domain adaptation method,this thesis improves the BDB feature extractor.Specifically,the features extracted by the two branches are concatenated into one feature after the fully connected layer,and the concatenated feature is used to calculate the loss function.Secondly,With the fact MAR network is restricted severely by hardware devices,Dissimilarity-based Maximum Mean Discrepency(D-MMD)network is also investigated in this thesis.On the basis of replacing D-MMD feature extractor Res Net-50 with BDB,the generalizing ability of the network is improved to force the trained network model to take into account the information of target person images.At the same time,this thesis also improves the loss function in metric learning by overall transformation,so that the recognition rate of the model learned by network in the target person image set can be further improved.Finally,the improved network and loss function are applied to large datasets(MSMT17,Market-1501)of person re-identification problem.The performance of each improved model is evaluated and compared with each other.In the mean time,this thesis also connects the process of person re-identification with the process of pedestrian detection,and realizes a more complete process of person re-identification from a recorded video.