面向行人重识别的跨视图最小分类误差二次判别分析方法  

Person Re-identification by Cross-view Minimum Classification Error-based Quadratic Discriminant Analysis Method

作　　者：江雨燕[1] 董映宇 郑炜晨 邵金 吕魏 JIANG Yu-yan;DONG Ying-yu;ZHENG Wei-chen;SHAO Jin;LV Wei(School of Management Science and Engineering,Anhui University of Technology,Ma'anshan 243002,China)

机构地区：[1]安徽工业大学管理科学与工程学院,安徽马鞍山243002

出　　处：《小型微型计算机系统》2021年第10期2125-2130,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(1172219)资助。

摘　　要：行人重识别技术旨在匹配不同的摄像机拍摄场景中属于同一个人的所有图片.近年来,核化跨视图二次判别法已在相关任务中取得优良的效果.然而,在处理高维小样本数据时,对于协方差矩阵逆的估计通常由于数据集较小的原因容易产生较大的偏差;在不同视图之间,人的外观经历复杂的非线性转换,因此导致识别精度较低.为解决此问题,本文提出一种将最小误差分类、平滑技术与核化跨视图二次判别法相结合的度量学习方法 MCE-kXQDA(minimum classification error-based kernel cross-view quadratic discriminant analysis),在非线性映射与跨视图二次判别法相结合的基础上将最小误差分类、平滑技术引入非线性维的核化空间中,实现非线性度量学习的同时有效提升协方差逆矩阵的估计精度.为验证MCE-kXQDA的有效性,我们在多个数据集上与其他相关方法进行了详细比较.实验结果表明MCE-kXQDA具有更优的识别精度和鲁棒性.The personal re-identification aims to match all the images of the same person in the scene taken by different cameras.In recent years,the kernelized cross-viewquadratic discriminant method has obtained better performance in the related tasks.However,when processing high-dimensional small sample data,the estimation of the inverse of covariance matrix is often prone to large deviation due to the small data set.Meanwhile,between different views,the appearance of a person undergoes complex nonlinear transformation,these results in lowrecognition accuracy.In order to solve this problem,a metric learning method(MCE-kXQDA) is proposed which fuses the classification of minimum error,smoothing technique with the kernelized cross-viewquadratic discriminant method,the nonlinear mapping method with cross viewquadratic discriminant method will be the basis of minimum error classification,smooth technology introduce nonlinear dimension nucleation in the space,to realize non-linear metric learning and improve the accuracy of inverse estimation of the covariance matrix.To verify the performance of MCE-kXQDA,we compare it to other related methods on multiple datasets.The experimental results showthat it has superior accuracy and robustness than other methods.

关 键 词：行人重识别 度量学习 最小误差分类 跨视图二次判别法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]