基于改进的交叉二次判别分析的行人再识别  

Person Re-identification Based on Improved Cross-view Quadratic Discriminant Analysis

作　　者：李亚庆 柳婵娟[1] 马振磊 LI Yaqing;LIU Chanjuan;MA Zhenlei(School of Information and Electrical Engineering,Ludong University,Yantai 264039,China)

机构地区：[1]鲁东大学信息与电气工程学院,山东烟台264039

出　　处：《鲁东大学学报（自然科学版）》2018年第2期120-127,共8页Journal of Ludong University:Natural Science Edition

基　　金：国家自然科学基金(61170161;61300155);鲁东大学博士基金(LY2014033)

摘　　要：行人再识别的任务是匹配不同摄像机在不同时间、地点拍摄的人体目标.受光照条件、背景、遮挡、视角和姿态等因素影响,不同摄像机下的同一目标差异较大.针对交叉二次判别分析算法(XQDA)中的矩阵奇异问题,本文提出了一种基于PLDA改进的交叉二次判别分析方法,即PLDA-XQDA算法.在PRID 2011,i LIDSVID数据集上进行实验,结果表明:本文方法与已有的基于视频的度量学习方法相比,PRID 2011和i LIDSVID数据集匹配率分别提高了1.23%和1.07%;与已有的基于图像的度量学习方法相比,PRID 2011和i LIDS-VID数据集匹配率分别提高了12.92%和15.34%.The task in person re-identification is to match snapshots of people from non-overlapping camera views at different times and places.Intra-class images from different cameras show varying appearances due to variation in illumination,background,occlusion,viewpoint and pose.Due to the singularity of the matrix,an improved cross-view quadratic discriminant analysis method based on the PLDA(PLDA-XQDA)was proposed.Experiments were done on the PRID 2011 and iLIDS-VID datasets,and the results were as follows:the proposed method improved the rank-1 identification rates by 1.23%and 1.07%,compared with the existing video-based metric learning methods.Meanwhile,it improved the rank-1 identification rates by 12.92%and 15.34%,compared with the existing image-based metric learning methods.

关 键 词：行人再识别 度量学习 交叉二次判别分析 扰动线性判别分析 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]