融合二阶HOG与CS-LBP的头部姿态估计

针对头部姿态估计受光照变化、表情、噪声干扰等因素影响导致识别率低的问题,提出一种融合二阶梯度方向直方图(HOG)和中心对称局部二值模式(CS-LBP)特征的姿态特征,用于单帧图像的头部姿态估计。采用二阶HOG对人脸图像进行形状信息提取,得到人脸的轮廓特征;用CS-LBP进行局部纹理信息的提取,通过将二阶HOG提取的轮廓特征和CS-LBP提取的纹理特征进行融合,得到更有效的人脸特征;将融合的姿态特征通过核主成分分析(KPCA)变换非线性映射到高维核空间中,抽取其主元特征分量,采用支持向量机(SVM)分类器进行姿态估计。实验结果表明,方法和HOG、LBP、二阶HOG、CS-LBP方法相比有更高的分类准确率,对光照的变化有很好的鲁棒性。

Head pose estimation fusing the second order HOG and CS-LBP

In order to improve head pose recognition rate under variable illumination, expression, and noise, etc., a novel pose feature, fusing the second order histogram of the orientation gradient (HOG) with the center symmetric local binary pattern (CS-LBP) feature, is proposed in order to estimate head pose in a single frame image. The contour information of the facial image is extracted by the second order HOG, deriving the facial contour feature. CS-LBP is used to extract local texture information. More effective facial features can be obtained by fusing contour feature extracted by the second order HOG and the texture feature extracted by CS-LBP. Kernel principal component analysis (KPCA) is used to nonlinearly project the fused pose feature into a higher dimensional kernel space so as to further select the primary feature. A support vector machine (SVM) classifier is used for pose estimation. Experiment results show that the proposed method is more accurate than the HOG method and the LBP method. This method has good robustness for variable illumination.