一种基于随机森林的头部位姿估计算法  

Head Pose Estimation Algorithm Based on Random Forests

作　　者：曾霞霞[1] 李佐勇[1] 林文如[1] 

机构地区：[1]闽江学院计算机科学系,福建省信息处理与智能控制重点实验室,福建福州350121

出　　处：《福建师范大学学报（自然科学版）》2016年第4期29-34,共6页Journal of Fujian Normal University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61202318);福建省教育厅基金项目(JA14259);福州市科技局平台项目(2015-PT-91)

摘　　要：针对人头部位姿势估计问题,提出一种基于随机森林的头部位姿估计算法.对现有算法只能以高质量人脸深度图像为输入和对面部数据缺失敏感的缺陷,在随机森林分支节点分裂机制中,加入分类测度解决头部区域的分割,以及改进回归测度来估计头部位姿,提出结合两种测度的优化方法,同时在原有几何特征通道基础上加入纹理信息以优化识别率,完成构造基于随机森林的头部位姿估计模型.结合该算法搭建基于Xtion PRO的实时头部位姿估计软件系统进行实验,结果表明,提出的两种测度模型能够较好地解决头部分割和位姿估计问题,该系统能够实时准确的估计头部位姿,并对部分头部遮挡具有鲁棒性.A new head pose estimation method based on random forests was proposed. It put forward an improved method regarding the shortcomings of the existing algorithm,such as slow convergence and performance depend on high quality depth data which assumes that the face is the sole object in the field of view,sensitive to occlusions. This method extended the regression forests such that they discriminate depth patches that belong to a head（ classification） and used only those patches to predict the pose（ regression）,jointly solving the classification and regression problems. And it adds texture characteristic to feature channels which only contain geometric features before in order to optimize discrimination. And then head pose estimation algorithm with random forests model was constructed and proving system based on Xtion PRO was implemented. In the experiments,it showes that the approach can handle real data presenting large pose changes,partial occlusions. It has thoroughly evaluated the system on a publicly available database on which it achieve state-of-theart performance.

关 键 词：深度图像 头部位姿估计 监督学习 随机森林 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]