基于二阶梯度朝向直方图特征的头部姿态估计

Head pose estimation based on a second order histogram of the orientation gradient

定义和提取与头部姿态紧密相关的特征是基于图像表观的头部姿态估计方法的关键步骤。该文提出将二阶梯度朝向直方图特征作为头部姿态图像表示,用于单帧图像的头部姿态估计。首先将图像划分成网格形式,对每个图像单元计算梯度朝向直方图,将相邻若干个图像单元组成图像块,对块内所有直方图元素之间进行成对组合得到成对关系向量,所有图像块的成对关系向量串联起来作为最终图像表示。该图像表示包含高阶的梯度朝向分布统计信息,有很强的姿态描述能力。实验结果表明:该方法比原始的梯度朝向直方图和GaFour等先进方法有更高的分类准确率。

One of the key problems of appearance based head pose estimation methods is to find the proper image descriptors which are closely related to the head pose variations.This paper describes a second order histogram of the orientation gradient（HOG） based head pose image representation for head pose estimation in still images.The image is divided into a grid of cells with the HOG computed for each cell.Then a pair wise relationship vector among the bins of neighboring HOGs is computed with all the pair wise relationship vectors then concatenated to form the final descriptor.The second order HOG is used to encode the second order statistics of the gradient orientation distribution to accurately describe the pose variations.Tests show that this method is more accurate than the two state-of-the-art methods using the original HOG and the GaFour method.