Research Of Unsupervised Person Re-identification

Nowadays,with the continuous improvement of people's public safety awareness,a large number of intelligent video surveillance systems have become an indispensable part of life.However,because of the complexity of the actual monitoring scene,the change of walking posture and the difference of camera angle and light intensity,person re-identification technology faces many challenges.Besides,most existing supervised models require tremendous manual labeling.It is quite expensive and not applicable to the real-world applications.Consequently,we focus on unsupervised learning.Our main innovations are listed as follows:1.Aiming at the large visual difference in pedestrian appearance caused by the changes of camera angle and person posture for unsupervised person re-identification,this thesis elaborates a exemplar-level and patch-level feature learning framework.Firstly,global features at the exemplar level are extracted,and then pedestrian images are divided into three patches,four patches and six patches,respectively.Finally,fusing the global features and local features can greatly enhance the discrimination of feature expressiveness.Thus,the recognition accuracy can be improved.Specifically,the feature memory mechanism is introduced to realize the unsupervised task.2.For the problem of unsupervised domain adaptation,this thesis firstly apply the attention mechanism in this field.First of all,the non-local attention module is embedded in the shallow and deep network to strengthen features at space level,and then the channel attention module is further integrated in the deep network to obtain interdependent relationship on the level of channels.Thus,the attention guided network can effectively guide the neural network to learn more discriminative domain invariant features which can narrow the difference of distributions between source domain and target domain and improve the generalization significantly.