基于核学习和距离相似度量的行人再识别

Person Re-identification using Kernel Learning and Similarity-distance Metric

行人再识别指的是在无重叠的多摄像机监控视频中,匹配不同摄像机中的行人目标.本文提出了一种基于核学习的测度学习的行人再识别方法,首先融合行人图像的颜色特征和纹理特征,并使用WPCA(PCA whitening)去除融合后的特征的冗余度,然后将处理过的特征通过核函数映射到更容易区分的核空间,并在核空间训练行人特征对之间的距离测度矩阵和相似度测度矩阵,结合距离测度函数和相似度测度函数来描述行人对之间的相似度.在VIPe R、iLIDS等数据集上的实验结果表明,本文的方法取得了较高的累积匹配得分,特别是第1匹配率,且对光照变化、行人姿态变化、视角变化和遮挡都具有较好的鲁棒性.

Person re-identification involves matching pedestrian images observed from different cameras in non-overlapping multi-camera surveillance systems.In this article,a person re-identification method based on kernelbased similarity metric learning is proposed.First,the dimension of person fusion features is reduced by PCA whitening.Second,the kernel trick is used to deal with reduced features.Finally,the similarity metric function,which combines similarity and distance functions,is applied to the system;it helps the system learn affine transformation,which shows pairwise contrast.The result based on VIPe R,i LIDS,ETHZ,and CUHK01 datasets shows a significant improvement in performance measured in cumulative match characteristic curves.The proposed method is robust to different viewpoints,illumination changes,varying poses,and the effects of occlusion.