Research On Metric Learning And Data Balance In Person Re-Identification

Due to the rapid development of video surveillance technology,the explosive growth of surveillance video has brought a lot of video information to public safety management,but at the same time,it has also greatly increased the difficulty of video processing.Extracting useful information in videos often requires manual screening,which not only consumes a lot of manpower,but also takes a lot of time.In this context,intelligent processing of large-scale surveillance video has become an important research content in the field of computer vision.Person re-identification aims to identify specific pedestrians in large-scale camera surveillance networks,and establish the identity matching of target pedestrians in different camera fields,in order to quickly locate pedestrians and achieve target tracking across cameras.Person re-identification technology has attracted a lot of research interests of researchers with its broad application prospects.However,due to the complex and diverse shooting environment and the variability of pedestrian appearance,the images taken by the same pedestrian in different surveillance areas will appear significantly different This brings great challenges to the research of pedestrian re-identification.Based on a comprehensive analysis of existing person re-identification methods at home and abroad,this paper mainly conducts research from the perspective of metric learning.In view of the current challenges in the field of person re-identification: the problem of extreme imbalance between positive and negative samples,a supervised and semi-supervised method was proposed to solve them.The specific research is as follows:(1)Most of the existing person re-identification metric learning methods treat all samples equally,which results in that the role of positive samples is not fully reflected,and ultimately the matching rate cannot be further improved.In response to this problem,we initially proposed the use of weighted Euclidean distance to measure pedestrian similarity,and assign different weights to the target positive and negative samples,so that the measurement model can learn more effective discrimination information from different types of samples.Thereby improving the discriminativeness of the model.(2)At present,most researchers of person re-identification use the publicly available data sets for research.In these data sets,the number of negative target samples is very large compared to the positive samples,which has resulted in model learning.Very large interference,and the impact of negative samples is also different.In(1),we use weighted Euclidean distance to measure the samples.In fact,different weights are given to positive samples and negative samples,which is equal to Treated all negative samples,and therefore ignored the different discrimination information provided by different negative samples.In response to this problem,we propose to use adaptive metric learning to classify negative samples and assign different weights to different types of negative samples.The new method was named Enhanced Metric Learning(EML).Experimental results show that this method can obtain higher accuracy of pedestrian re-identification.At present,it takes a lot of human and financial resources to obtain the information of the descendants of the surveillance video.The imbalance of data has become an urgent problem in pedestrian re-identification.We proceed from the perspective of the data itself and use the picture style conversion technology to shoot pedestrian pictures Style conversion to achieve the purpose of data enhancement.Since the sample images generated by the style conversion do not have pedestrian labels,we propose to use self-paced learning graph technology to learn the pseudo-paired relationship between the samples.Through the pseudo-paired relationship,labeled samples and unlabeled samples can be combined so that the metric model can learn Information about pedestrians in different shooting styles.Experimental results show that the model has outstanding performance for pedestrian target recognition problems.