基于特征点识别的头部姿态计算

为了提升使用随机回归森林进行头部姿态分析的精度，提出了一种基于特征点识别分析头部姿态的计算框架.考虑到高误差投票的干扰，该计算框架以随机森林的特征点识别为基础从而避免异常投票干扰，将头部姿态计算问题转换为空间鼻尖特征点和朝向特征点的识别问题.在随机森林的训练中，决策函数使用了高斯曲率和平均曲率作为图形特征，根据微分熵的信息增益在随机生成的决策函数库中搜索最优化决策函数.在训练完成的随机回归森林的叶子节点中，通过分析保存的样本数据，可以得到目标特征点的高斯分布估计.根据实验测试结果，在适当的阈值设定的情况下，该方法可以实现较高的识别成功率，使用曲率后明显提高了识别精度，能够在一定程度上处理有遮挡的数据，并且该方法已经成功应用于虚拟座舱的实时头部姿态分析计算系统.

Head pose estimation framework based on feature point detection

In order to improve the precision of head pose estimation with random regression forest, an analysis framework based on feature point recognition was proposed for head pose estimation. In view of invalid votes disturbance, this framework recognized head position point and direction point to avoid accepting abnormal voting. The decision function used depth value, normal vector, Gaussian curvature and mean curvature as image features. An approximate optimized decision function search was executed in a binary test pool generated randomly according to information gain of differential entropy. The experiments focused on performance analysis on different occlusion rates of original data. The approach got high success rate in experiments with appropriate parameters, improved accuracy after using curvature, and enhanced the capability of handling with occlusion. The proposed framework has been applied for real-time head pose estimation system in virtual cockpits successfully.