Research And Application On Person Re-identification In The Weakly Supervised Setting

With the progress of society,the public security system with social safety as the core is attracting more and more attention,and the video surveillance system is an important foundation of the system.At the same time,with the continuous improvement of the monitoring system,the efficient analysis of large amounts of monitoring data has a very important significance for the security management,mainly for criminal investigation.The core goal of the Person Re-identification task is to correlate targets under the viewpoint of non-intersecting cameras.The task is very challenging in real scenarios due to differences in perspective,light,resolution,etc.In recent years,with the development of artificial intelligence,The Person Re-identification algorithms with deep learning have made a great breakthrough,but compared to the face recognition and object detection and other classic computer vision task,Person Re-identification is far away from the actual application.One of the most important reasons that the high cost of data labeling.There are also very significant gaps in weakly-supervised learning algorithms that utilize less labeled information compared to supervised learning algorithms.Therefore,this paper mainly investigates the weakly-supervised method of Person Re-identification,proposes one unsupervised and one semi-supervised learning algorithm respectively which using the global-based auxiliary memory.By a simple analysis of the relevant systems in recent years,we propose a cross-camera person retrieval system based on scene adaptation.It integrates model training and application which realizes a combination of theoretical algorithms and practical applications.The main contributions of this paper can be summarized as follows.1.The benchmark method ECN based on auxiliary feature memory is analyzed and defects are proposed: high-cost of training,the high noise in the begining of train-ing,unsuitable neighborhood selection method.Based on ECN,this paper proposes a multi-stage unsupervised Person Re-identification learning algorithm based on loop consistency matching.First the connections between samples are constructed step by step through the sequence of sample learning,intra-camera learning,and intercamera learning.Next,this paper proposes the loop consistency-based neighbor selection methods One Shot and IOU based on sample association tables.By these methods,we reduce the dependence on GAN and the cost of training compared with the benchmark,and we also balance the precision and recall during neighborhood selection.Finally,we prove the effectiveness of this algorithm by experience.2.By analyzing the Person Re-identification algorithms based on different supervision types,we prove the advantages of the semi-supervised method.By add a few low-cost labels,A progressive soft label semi-supervised Person Re-identification method based on neighbor selection is proposed,which extended from the unsupervised method above.It achieves huge performance gains.The approach constructs a semisupervised learning framework using the intra-camera label.Then determines weights through a neighbor selection method based on the affinity matrix and introduces the concept of progressive soft label.Finally proposes the classification and discriminating losses based on progressive soft label.Experiments have demonstrated that the method can achieve significant performance gains on public datasets.3.Based on the above methods,a cross-camera person retrieval system based on scene adaptation is proposed.Firstly,an analysis of the current existing Person Reidentification system is carried out,and we introduce the advantages of this system.On the one hand,to address the problem of low generalization performance in different scenarios,the system integrates the classical and excellent methods of recent years and provides a unified algorithm interface,which can flexibly train the model for different scenarios;on the other hand,all modules of the system are implemented on the Web,so it is convenient and fast to use the services such as retrieval and training through simple interaction.