从残缺的人体轮廓进行姿态估计  

Human pose estimaties from corrupt silhouettes

作　　者：沈春锋[1] 林学訚[1] 史元春[1] 

机构地区：[1]清华大学计算机科学与技术系,北京100084

出　　处：《清华大学学报（自然科学版）》2008年第10期1659-1662,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(60433030);国家"九七三"重点基础研究资助项目(2002CB312101)

摘　　要：通过背景差分和颜色分割得到的人体轮廓通常是残缺的,常会使传统的基于回归方法的人体姿态估计方法失效。该文提出了一种基于子流形投票的鲁棒统计方法。将人体轮廓分成几个小的部件之后,以Gauss过程训练获得其相应的子流形,并采用子流形投票的方式计算似然概率。通过融合似然概率和人体姿态的先验概率,利用Bayes推理框架就可以求出输入的人体姿态在流形空间的坐标。借助于Gauss过程,就可以从这个流形坐标推理出人体的姿态结构和外轮廓。通过对步行的姿态数据进行测试,该算法可以鲁棒地进行姿态估计,同时恢复残缺的人体轮廓。Human silhouettes obtained by background subtraction or color segmentation are always corrupt, which makes traditional regression methods fail. A robust statistical method named sub manifold voting was developed for pose estimation from corrupt silhouettes. After dividing the silhouette into small components their sub-manifolds are learned using Gaussian process, which are then used to calculate the likelihood probability by means of a sub-manifold voting strategy. The likelihood is fused with prior human poses in a Bayesian framework to specify the coordinates of the pose in the manifold space. The intrinsic pose structure and silhouette are then deduced from these coordinates using a Gaussian process. Tests on walking sequences illustrates that this method robustly estimates the pose structure and effectively recovers corrupt silhouettes.

关 键 词：图像识别 GAUSS过程 流形 姿态估计 鲁棒统计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]