Person Re-Identification Technology Based On Relation Mining

With the development of technology and the promotion of projects aimed at maintaining social stability,such as "Safe City" and "Safe Campus",the penetration rate of monitoring equipment has been greatly improved.This phenomenon can lead to a lot of monitoring data each day.In the event of an incident,police officers need to manually view a large number of monitoring,which results in a huge waste of resources.For such problems,the Person Re-Identification(ReID)technology came into being.In recent years,with the rise of Deep Learning,represented by Convolutional Neural Networks(CNNs),more and more researchers have begun to apply CNNs to ReID tasks and made great results.The existing ReID methods based on deep learning are mostly divided into two steps.One is to use CNN to extract features from pedestrian images or video sequences,and the other is to measure the similarity of these features.After investigation,the existing methods did not fully exploit the various relationships in pedestrian images and video sequences,but only "violently" extracted and compared features.In order to dig out the various relationships of pedestrian data and improve the retrieval accuracy of the ReID system,this thesis has done the following work(1)In the process of extracting pedestrian features using a convolutional neural network,in order to fully mine and represent the deep-level structural relationships in the convolutional neural network,a group structure model is proposed,and according to the nature of this structure,a method for mining its relationships is proposed.Specifically,this article comes up with an approach to fully mine the symmetric relationship information contained in it.From the results we can know that the structural relationship mining method proposed in this paper makes the person re-identification system more accurate,and to a certain extent makes up for the shortcomings of using the convolutional neural network to extract features directly.(2)In view of the problem that existing RelD methods don't fully exploit the semantic relationships,this thesis proposes three semantic models and separately explores the semantic relationships they contain.The three semantic models are spatial semantic model,chanel semantic model and inverse semantic model,respectively.In the thesis,these three semantic models are put into the existing ReID systems,and then a large number of contrast experiments are carried out on different monitoring datasets.The results prove that the semantic relationship mining methods proposed in this thesis are very useful.(3)In the video sequence-based ReID tasks,since the adjacent frames of the video are connected,it is necessary to model and mine the relationships between these frames to improve such ReID tasks.Therefore,a ReID method based on mining timing relationships is proposed.Specifically,in the thesis,input monitoring videos are modeled in the way of time series,and then the relationships contained in the time series model are mined.From the results we can know that the method in thesis is more effective than the other methods.