基于加权局部梯度直方图的头部三维姿态估计  

Head Pose Estimation Using Weighted Localized Gradient Orientation Histogram

作　　者：崔汪莉 卫军胡[1] 纪鹏[2] 刘哲 

机构地区：[1]西安交通大学机械制造系统工程国家重点实验室,西安710049 [2]西安交通大学机械工程学院,西安710049 [3]西安邦威电子科技有限公司,西安710049

出　　处：《西安交通大学学报》2015年第11期71-76,共6页Journal of Xi'an Jiaotong University

摘　　要：在实时估计人的头部三维姿态时,基于局部梯度方向直方图的面部特征表示方法容易受到背景和环境的影响,其检测精度无法满足实际需求。为了减少图像或视频序列中背景和环境的影响,提出了一种新的对面部特征进行描述的方法,即基于肤色权值和高斯权值加权的局部梯度方向直方图特征表示方法。在具体计算时,首先进行人脸检测并将人脸区域缩放到统一大小,然后计算人脸区域每个像素点对应的梯度方向,接着计算肤色权值并利用肤色权值和高斯权值对梯度方向进行加权得到加权局部梯度方向直方图,从而强化面部特征在直方图中的比重,有效减小背景对头部三维姿态估计的影响,最后利用非线性支持向量回归机求解加权局部梯度方向直方图与头部三维姿态之间的关系。实验结果表明:该特征表示方法具有更高的检测精度。When used for real-time 3D head-pose estimation, the facial features based on the localized gradient orientation histogram are easily affected by the environment and background so that the detection accuracy cannot meet the practical requirements. To reduce the influence of environment and background in images and video sequences, this paper presents a new weighted localized gradient orientation histogram to represent the facial features. During the computation, faces are detected and made the same size firstly. The gradient orientations of every point in the facial area are computed and then weighted by its skin-color probability and a Gaussian random value. Based on these gradient orientations a weighted localized gradient orientation histogram is obtained, in which the role of facial area is increased and that of environment and background are reduced. Finally the relationship between the 3D head-pose and the new features is computed using nonlinear support vector regression method. The results of numerical experiments show that this new method has a reletively high detection accuracy.

关 键 词：三维头部姿态估计 肤色权值 高斯权值 局部梯度方向直方图 非线性支持向量回归机 

分 类 号：TH137[机械工程—机械制造及自动化]