The Auxiliary Information Of The Object For Person Re-Identification

With the ever-increasing requirement of social security and the development of video monitoring technology,more and more surveillance monitors are deployed in various public places.The traditional video monitoring systems rely on manually browsing videos to find the object.However,with the continuous expansion of the monitoring system scale,video monitoring data is growing at an exponential level.As a result,manually searching is not able to meet the demand of massive data.Person reidentification is a computer vision technology aiming to find the target pedestrian in a series of images or videos,which has broad applications in intelligent monitoring,unmanned supermarket,urban order management,and many other fields.While in the actual monitoring scenario,the appearance characteristics of the same pedestrian can vary greatly due to the influence of pedestrian view,posture change,and background interference.This paper focuses on two key difficulties,i.e.,'iew change' and'background interference'.The main contents are as follows:1.To solve the problem that the appearance characteristics of pedestrian images changing greatly due to different views,this paper proposed a view-aware metric learning method.It is found that there are significant differences in the correlation characteristics of pedestrian images in similar and dissimilar views.When the views of pedestrian image pairs are the same,some detailed features can be used to judge whether these two images contain the same pedestrian.When the pedestrian view changes significantly,some detailed features may be missed,so it is necessary to learn more robust features for view changes.If the image pairs in similar views and different views are extracted with unified features,it is difficult to achieve a balance between different correlation features.In this paper,pedestrian image pairs are divided into view-similar and view-dissimilar image pairs according to the view information,and then the image pairs in similar views and dissimilar views are projected into different feature subspaces by two fully connected layers with unshared weights respectively.Finally,multiple metric losses are used to supervise the learning of the network-Experiments on three common person re-identification datasets CUHK01,CUHK03 and PRID2011 demonstrate the effectiveness of our method.2.To solve the problem that similar background would cause interference to pedestrian identity matching,this paper proposed a method based on semantic segmentation.Since the existing datasets are only composed of pedestrian images captured by a few cameras,equal processing of all pixels in the image will generate a high degree of similarity between images with similar backgrounds.This paper constructs a joint framework of semantic segmentation subnetwork and person re,identification subnetwork.The pedestrian mask captured by the semantic segmentation network is used to guide the online learning of person re-identification network,so that the network focuses on the pedestrian area to learn more discriminative pedestrian appearance characteristics.At the same time,aiming at the problem of missing key parts and unclear contour of pedestrians in the semantic segmentation model,we combine the semantic segmentation network with the middle-level features of the person re-identification network to assist the segmentation task.Experiments on CUHK01,CUHK03 and a large dataset Market-1501 demonstrate the effectiveness of the proposed method.