CASIA OpenIR  > 毕业生  > 硕士学位论文
基于人脸关键点标记的特征识别及姿态估计
成艺
学位类型工学硕士
导师郭大蕾
2017-05
学位授予单位中国科学院研究生院
学位授予地点北京
关键词人脸关键点标记 人脸识别 头部姿态估计 人机交互
摘要    面部特征识别和头部姿态估计可根据图像提供用户情绪状态、视线方向、操作意图等大量隐含信息,在心理分析和人机交互等领域应用前景广阔。人脸关键点标记作为面部图像分析中的基础研究,是面部特征识别和头部姿态估计的重要前提。本论文从最优化求解角度,围绕人脸关键点标记问题及其在面部特征识别和头部姿态估计中的应用进行研究,主要内容如下:
    首先,研究了人脸关键点标记问题。针对图像特征描述信息冗余问题,提出了PCA-SDM(Principal Component Analysis-Supervised Descent Method)人脸关键点标记方法,通过主成分分析将高维特征投影至低维子空间,缩短了模型训练时间,提升了关键点标记准确度。为了提高关键点标记鲁棒性,建立了联合表观和位置特征的asSDM(appearance-shape-Supervised Descent Method)关键点标记模型,借助岭回归法求解位置偏量,降低了关键点标记误差。
    其次,研究了基于人脸关键点信息的面部特征识别问题。针对眼部状态分类问题,根据眼部关键点位置,建立了基于边缘信息距离的眼部状态分类器模型,比较了多种分类器和距离特征的检测结果。同时,针对人脸识别问题,设计了基于关键点表观信息的面部特征,分析了不同识别算法和图像特征描述子对准确率的影响。
    再次,针对人脸关键点标记和跟踪分别在头部姿态估计和运动检测中的应用展开研究。建立了从三维空间到二维平面的头部姿态映射模型,分析了视频中头部姿态时序数据在不同尺度空间下的幅值特性,实现了连续参数的头部姿态估计和动作检测。
    最后,本论文实现了基于头部姿态和运动的计算机图文显示控制及三维模型旋转操作方案,完成了友好、便捷的人机交互。
其他摘要    In the field of computer vision, face alignment is among the most popular and well-studied problem. Accurate and efficient alignment algorithm provides a benchmark for face recognition, head pose estimation, facial expression analysis and 3D face modeling. This thesis employs the supervised descent method as the major tool, studies the face alignment issue and the application of facial feature detection and head pose estimation. The major contributions of this thesis are as follow:
    Firstly, face alignment algorithm is studied. Face alignment can be solved as a nonlinear optimization problem. This thesis proposes a modified method PCA-SDM(Principal Component Analysis-Supervised Descent Method) to extract the effective information and weak the noise by selecting the appropriate dimension. Most of the existing methods only rely on the current facial texture and it is unreliable when facial landmarks are partially occluded in unconstrained scenarios . To settle the issue asSDM(appearance-shape-Supervised Descent Method) is proposed , utilizing both appearance and shape information in learning regression functions. The performance of the proposed method is evaluated on different data sets and the results on benchmark databases demonstrate that the proposed method outperforms previous work for facial landmark detection.
    Secondly, the applications of facial feature recognition based on face alignment are studied. According to the facial key points, the eye state can be discriminated. Using texture information at the key points as facial feature instead of the single pixel value in an image, can achieve higher accuracy in face recognition algorithm, especially dealing with the images where there is a large variety on pose and expression.
    Thirdly, the applications of head pose estimation and motion detection based on face alignment and track are studied. By constructing the mapping from 3D model to 2D plane, head pose can be estimated using a continuous angular measurement across multiple degrees of freedom.
    Finally, utilizing head pose information to control the computer can provide users with a more natural, convenient and comfortable operating experience.
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14631
专题毕业生_硕士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
成艺. 基于人脸关键点标记的特征识别及姿态估计[D]. 北京. 中国科学院研究生院,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于人脸关键点标记的特征识别及姿态估计.(32360KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[成艺]的文章
百度学术
百度学术中相似的文章
[成艺]的文章
必应学术
必应学术中相似的文章
[成艺]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。