CASIA OpenIR  > 毕业生  > 博士学位论文
面向自然互动的表情与手势分析
其他题名Facial Expression and Hand Gesture Analysis for Natural Interaction
汪晓妍
学位类型工学博士
导师王阳生
2009-06-02
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业计算机应用技术
关键词表情分析 手势分析 在线纹理模型 模糊聚类 隐马尔可夫模型 Facial Expression Analysis Hand Gesture Analysis Online Appearance Model Fuzzy Clustering Hidden Markov Model
摘要自然互动是人机交互的发展方向,其目标是赋予计算机参照人类自然形成的与自然界沟通的认知习惯和形式来与用户进行沟通和互动的能力。作为人们非语言类交流中最为自然和直观的方式,表情和手势在社会活动中起着重要的作用,利用计算机视觉方法对表情和手势的研究有助于实现自然的人机接口。本文从自然互动的角度出发,对视频中的表情和手势进行分析,主要贡献如下: 1. 在人脸和表情建模方面,本文利用CANDIDE三维参数化模型对人脸形状和表情动作进行建模,通过弱透视模型对姿态进行建模,在模型中引入了生理结构约束,避免了生成一些实际不可能出现的表情,提高了后期跟踪识别的效率。另外根据人脸表情的特点选择出合适的表情动作参数用于人脸多特征跟踪和表情识别。 2. 在人脸多特征跟踪方面,本文从观测建模和模型匹配这两大问题出发,首先提出了一种基于灰度和边强度的综合特征和在线学习的自适应方法来进行观测建模,提高了跟踪的鲁棒性;其次为了解决模型匹配的迭代算法中耗时较多的问题,提出一种反转合成图像对齐算法结合在线纹理模型的迭代匹配算法,提高了模型匹配的效率。实验证明本方法能够适应复杂的光线和表情变化。 3. 在表情分析方面,提出一种基于高斯基距离度量的模糊聚类算法用于表情识别,基于改进方差率来设计高斯基度量中的惩罚因子,提高了聚类算法的性能。将跟踪与表情分类进行了统一考虑,直接将跟踪得到的表情参数作为表情识别的特征,从而排除了不同人的面部差异对于表情识别的影响。由于表情的复杂性和不确定性,在识别时给出了表情模糊性的描述。 4. 在手势分析方面,提出一种基于隐马尔可夫模型的动态手势轨迹的建模和识别方法,采用3次B样条拟合出手势轨迹曲线,求取曲线矩不变量全局特征和方向局部特征作为轨迹特征描述,引入阈值模型对非典型手势进行建模,并进行动态手势序列的自动切分和识别。本方法可以实现全自动的在线手势识别,并能对非典型手势进行成功拒识,实验证明加入曲线不变量全局特征比单独用方向特征的方法取得了更好的识别效果。
其他摘要Natural interaction is the evolutive direction of human computer interaction, and the goal is to give computer the ability to communicate and interact with user according to the cognitional habit and manner formed unselfconsciously when human is communicating with nature. As the most natural and intuitive interaction manners, facial expression and hand gesture play important roles in social activities. It will be very helpful for natural human computer interface realization to do research on facial expression and hand gesture using computer vision techniques. In this dissertation, we study on facial expression and hand gesture analysis, and the main contributions of this thesis are as follows: 1. As for face and facial expression modeling, a 3d parameterized model called CANDIDE is used to model the face and facial actions, a weak perspective projection method is used to model the head pose. Restriction based on physical structure is used in the model, so actually impossible facial expressions are avoided and the efficiency of latter tracking and recognition is promoted. In addition, appropriate facial action parameters are picked out for facial feature tracking and facial expression recognition according to the characteristics of facial expression. 2. As for facial feature tracking, we start from the two problems of observation modeling and model fitting. Firstly, an adaptive method based on online learning and robust feature combining edge strength and raw intensity is proposed to build the observation model, and the robustness of tracking is promoted. Secondly, to solve the time-consuming model fitting problem, an iterative model fitting algorithm combining Inverse Compositional Image Alignment with online appearance models is proposed, which improves the fitting efficiency. 3. As for facial expression analysis, a fuzzy clustering algorithm based on Gaussian basis distance measure to do expression recognition; while augmented variance ratio is used as the penalty factor of Gaussian basis measure and the clustering performance is promoted. We give a total consideration of the anterior alignment and tracking with the posterior facial expression classification, and the facial parameters obtained by tracking are directly used to do facial expression recognition, so the impact caused by facial diversity of different person is eliminated. Fuzzy description of facial expression is presented in recognition on account of its complication and uncertainty. 4....
馆藏号XWLW1385
其他标识符200618014629088
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6214
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
汪晓妍. 面向自然互动的表情与手势分析[D]. 中国科学院自动化研究所. 中国科学院研究生院,2009.
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
CASIA_20061801462908(2172KB) 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
[汪晓妍]的文章
百度学术
百度学术中相似的文章
[汪晓妍]的文章
必应学术
必应学术中相似的文章
[汪晓妍]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。