基于核fisher判别度量学习的行人重识别-中国科技论文在线.pdfVIP

中国科技论文在线 基于核Fisher 判别度量学习的行人重识别 李昭青，陈振学，李秀清** （山东大学控制科学与工程学院，济南 250061 ） 5 摘要：行人重识别的主要研究方法为特征表示学习和度量学习，由于度量学习比特征表示学 习更容易在行人重识别问题上取得显著效果，因此本文提出基于核Fisher 判别度量学习的行 人重识别方法。由于主成分分析法降维会降低特征的表达能力，因此首先运用核化的Fisher 判别算法对原始高维空间中的样本数据降维，然后使用L1 损失函数对样本进行约束来学习 10 距离度量矩阵，最后通过距离计算得到行人重识别结果。相较于MLAPG 方法，本文方法在 VIPeR 数据集、QMUL GRID 数据集和CUHK01 数据集上rank-1 的匹配准确率分别提升了 0.39%、0.13%和0.62%。三个不同数据集上的实验结果表明了本文方法的有效性。 关键词：行人重识别；度量学习；核Fisher 判别；损失函数；距离度量矩阵 中图分类号：TP391.4 15 Kernel-based Fisher Discriminant Metric Learning for Person Re-idengtification LI Zhaoqing, CHEN Zhenxue, LI Xiuqing (College of Control Science and Engineering, Shandong University, Jinan 250061) 20 Abstract: The research methods of person re-identificationare feature representation learning and metric learning. Compared with the feature representation learning, the metric learning is more likely to achieve significant effect on person re-identification.Therefore, this paper propose a person re-identification method based on kernel Fisher discriminant metric learning. Because the principal component analysis reduces the expression ability of the feature,this paper first uses the kernel Fisher 25 discriminant algorithm to reduce the dimensions of sample data in the original high dimensional space.On this basis, the L1 loss function is used as the objective function to learn the distance metricmatrix, and finally the person re-identification results are obtained by calculating distance. Compared with