基于二次相似度函数学习的行人再识别

发布时间：2018-02-13 17:55

本文关键词： 视频监控行人再识别行人分析度量学习二次相似度函数学习机器学习　出处：《计算机学报》2016年08期 　论文类型：期刊论文

【摘要】：行人再识别是一个有着非常重要现实意义的研究问题,它可以应用于刑事侦查、在公共场所中寻找丢失的小孩、个人相册管理以及电子商务等领域.同时由于光照、视角、人的姿态以及背景的变化,同一个人的表观在不同的监控视频中往往变化很大,解决行人再识别问题也非常有挑战性.在设计行人再识别算法时,给定行人图像的特征,考虑到不同的特征分量具有不同的区分能力,学习合适的相似度度量非常重要.度量学习是一类学习相似度度量的主流算法,这些算法通过学习一个马氏距离相似度函数(Mahalanobis Similarity Function,MSF)来估计一对行人图像的相似度.然而MSF只与特征差分空间有关,忽略了一对图像中每个个体的表观特征,对于同一个人在不同场景中很大的表观变化的捕捉能力有限.为了加强相似度函数与每个个体的表观特征的联系,该文提出通过学习一个二次相似度函数(Quadratic Similarity Function,QSF),来估计一对行人图像的相似度.QSF是MSF的泛化形式,不但描述了一对行人图像的互相关关系,而且关联了一对行人图像的自相关关系,可以更好地捕捉同一个人在不同监控视频中很大的表观变化.为了学习QSF,该文分别从分类和排序的角度出发,设计两种不同的优化目标,提出了两种不同的学习QSF的算法.由行人再识别的公共数据集VIPeR和CUHK的实验表明,这两种不同的算法都可以学习到有效的QSF,识别性能优于已有的行人再识别算法.
[Abstract]:Pedestrian re-identification is a very important research issue of practical significance. It can be used in criminal investigation, finding lost children in public places, personal photo album management, electronic commerce, etc. At the same time, due to illumination, perspective, and so on, The appearance of the same person changes greatly in different surveillance videos, and it is also very challenging to solve the problem of pedestrian rerecognition. When designing the pedestrian rerecognition algorithm, the characteristics of the pedestrian image are given. Considering that different feature components have different ability to distinguish, it is very important to learn appropriate similarity measure. Metric learning is a kind of mainstream algorithm for learning similarity measurement. These algorithms estimate the similarity of a pair of pedestrian images by learning a Mahalanobis Similarity function MSF. However, MSF is only related to the feature difference space and ignores the apparent features of each individual in a pair of images. The ability of the same person to capture large apparent changes in different scenarios is limited. In order to enhance the relationship between the similarity function and the apparent characteristics of each individual, In this paper, we propose to estimate the similarity of a pair of pedestrian images by learning a quadratic similarity function, Quadratic Similarity function, QSF, which is a generalization form of MSF, which not only describes the cross-correlation relationship of a pair of pedestrian images. Furthermore, the autocorrelation of a pair of pedestrian images can better capture the large apparent changes of the same person in different surveillance videos. In order to study QSFs, this paper starts from the perspective of classification and sorting, respectively. Two different algorithms for learning QSF are proposed by designing two different optimization objectives. The experiments of VIPeR and CUHK, a common data set for pedestrian recognition, show that, These two different algorithms can learn effective QSFs, and the recognition performance is better than the existing pedestrian rerecognition algorithm.
【作者单位】：清华大学计算机科学与技术系;
【基金】：国家自然科学基金(61075026) 国家“九七三”重点基础研究发展计划项目基金(2011CB302203)资助
【分类号】：TP391.41

【相似文献】