典型相关分析在人脸姿态估计中的应用  

Facial Pose Estimation using Canonical Correlation Analysis

作　　者：赵松[1] 潘可[1] 张培仁[1] 

机构地区：[1]中国科学技术大学自动化系,安徽合肥230027

出　　处：《小型微型计算机系统》2009年第5期983-987,共5页Journal of Chinese Computer Systems

摘　　要：在单一物体角度估计中,典型相关分析(CCA)可以用来建立图像空间和角度空间的联系.基于人脸总体形状的相似性,提出了基于外观的人脸姿态估计方法.使用CCA建立姿态变化时人脸这一类物体的外观空间和姿态空间的关系.典型相关向量最大化人脸外观空间和姿态空间的相关性,张成了它们的相关子空间.在相关子空间中,通过线性回归的方法,估计未知姿态图像的角度.为了更好的处理外观空间的非线性问题,引入了该方法.在CUbiCFacePix数据库上的实验验证了这两种算法的有效性.In single object pose estimation problem, canonical correlation analysis （CCA） is used to fred the correlation between object image and its pose. A pose estimation method is proposed based on appearance model, considering the shape similarity of face universal set. CCA is introduced to derive correlation between appearance space of face set and pose space, whose canonical vectors span the subspace of them. Finally, the poses of face images are predicted by linear regression method between the two correlated subspaces. Kernel method is also introduced to handle non-linearity lies in appearance space. Experiments on CUbiC FacePix database validate the new pose estimation method.

关 键 词：典型相关分析 核典型相关分析 人脸姿态估计 线性回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]