基于回归森林的面部姿态分析  

Regression Forests for Head Pose Estimation Analysis

作　　者：乔体洲[1] 戴树岭[1] 

机构地区：[1]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《计算机辅助设计与图形学学报》2014年第7期1151-1158,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"八六三"高技术研究发展计划(2009AA01Z333)

摘　　要：快速稳定地计算头部姿态的算法在很多应用领域都是非常重要的,为了寻求在飞行模拟器中实时跟踪操纵者头部运动的新方法,提出一个基于随机回归森林、使用深度数据来解决面部朝向的计算框架.该框架利用标注了真实头部位置和朝向的大规模人体面部模型数据库进行随机森林的训练,将携带标注真实参数值的随机采样图像块输入随机森林进行训练;在决策树叶子节点中得到姿态参数的高斯分布,再使用得到的随机森林进行面部姿态的计算,从而将面部姿态分析问题转换为待测试深度图像的随机采样子域的投票问题.测试了参数和引入计算的图像特征对识别性能的影响,并与相关算法进行比较,结果表明,该框架有较高的识别率和抗干扰能力,能够处理头部姿态大范围、快速变化、暂时性遮挡以及面部表情等数据.In order to tracking head pose motion of pilots in flight simulator, new and stable algorithms of head pose estimation are imperative and depth data are considered to be used to estimate head pose. We implement a random-regression-forest-based framework for head pose estimation. This framework uses massive human face scan model database annotated with head position and orientation, and patches are sampled randomly and sent to random forest for training. After training parts of leaf nodes save the Gaussian distributions for head position and orientation. Consequently head pose estimation is converted to searching for votable patches from test data, and votes of these patches are used to estimate head pose parameters. We analyze and demonstrate the effect of random forests’ parameters and image features. According to experiments the approach can handle real data when large and rapid head rotations, partial occlusions, and facial expressions exist.

关 键 词：随机森林 面部朝向分析 决策树 飞行模拟器 机器视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]