【作者】 王丽；

【导师】 李瑞峰；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2015， 博士

【摘要】 智能服务机器人在示教、娱乐、导游、助老助残等领域有着广阔的应用前景,是一个热门而又意义深远的研究领域。一般而言,自然和谐的人机交互能力是智能服务机器人必备的,因为丰富细腻的情感变化和心理活动信息的面部表情在人际生活中起着不可替代的作用。表情识别功能和仿人头部机器人是智能服务机器人系统中两个重要组成部分,前者用于交互人员表情输入,后者用于机器人表情输出。通过对表情识别算法和机器人头部实体研发实现自然高效人机交互,在理论研究和实际应用中均具有重要意义。为实现自然和谐人机交互,本文针对智能机器人表情交互系统中的表情识别技术和仿人头部机器人机构开展两方面研究工作:首先,分别对表情识别过程中实时自动分割人脸区域的方法、复合二值模式特征提取、无监督学习的LBP特征选择以及矩阵回归分析表情分类器设计进行研究,从而为智能化人机交互提供识别算法基础;在此基础上,进一步开展了关于仿人头部机器人表情再现实验系统的研制,经实验分析了相关原理方法的有效性。本文具体内容如下。人脸检测是表情识别中重要和必然的前期阶段,主要对识别目标实现精确面部区域分割。本文采用主动形状模型,以人脸五官轮廓为基础构建人脸识别模型,减少头部姿态变化对表情识别的影响并去除面部以外区域,从而提高人脸检测和人脸区域提取准确率。根据面部活动编码系统,利用脸部标志点提取人脸表情关键区域,精确定位人脸位置以及提取表情识别特征区域。有效表情特征提取对表情识别具有重要意义。通过研究局部特征算子中均匀二值模式,本文提出复合二值模式算子来描述人脸纹理的变化趋势和程度,通过揭示多邻域内不同方向数据结构的变化来提取面部主要表情区域的特征,为表情识别提供识别向量。减少表情特征的冗余信息是提高识别效果的一种有效方式。采用二值模式算子能够有效提取图像局部纹理特征,但其维数随着邻域中像素数量增大而陡然增加。因此,本文根据二值模式的直方图属性,采用2χ统计量构建特征空间邻域图,然后由二值权重矩阵求出拉普拉斯2χ统计值选择出更具表征性的特征,形成表情识别的新特征子集。分类器设计是表情识别的重要环节。线性判别分析作为种传统的分类算法无法有效解决特征空间的非线性问题,对大量高维数据样本的计算量巨大,且特征值求解中易受奇异值问题困扰。在图谱及核学习理论基础上,通过分析线性判别分析的权重矩阵与核空间中线性系数间关系,本文提出矩阵回归分析表情识别分类器,用于识别多种表情,避免分类器训练中特征值计算,有效降低计算时间和存储量。具备面部表情再现功能的智能服务机器人对实现智能化、情感化和人性化人机交互有着重大意义。从仿生角度出发,基于仿人头部机器人和表情识别技术构建智能机器人表情再现实验系统,分别在该系统上进行头部机构面部特征的多种动作实验及真实环境下表情识别算法的验证,在真实场景中利用识别的表情类别实现表情再现。本文针对相关的表情识别技术分别在数据库和实际环境中进行实验,验证了提出的复合二值模式算子、无监督特征选择算法以及矩阵谱回归分析表情分类器的有效性和可行性,和真实环境下自动人脸区域提取的实时性和稳定性。在数据库和真实环境中取得良好的识别效果,且实现真实环境下实时性人脸区域的自动提取,并通过表情识别结果利用仿人头部机器人再现表情。更多还原

【Abstract】 Intelligent service robot, which has a broad application prospect in teaching, entertainment, tour guiding, elderly or disabled people assistance and other fields, is a very popular and significant research field today. Generally, the ability of natural and harmonious human-computer interaction(HCI) is necessary for intelligent service robot. Since containing rich delicate information of emotional change and mental activity, human facial expression plays an irreplaceable role in social communication. Facial expression recognition(FER) and anthropomorphic robot head are two important parts of the intelligent service robot system, and the former is for the input of operator’s expression and the latter for the output of robot’s expression. It would make much sense in theoretical research and practical application to implement natural and efficient HCI through research and development of FER algorithm and robot head entity.In order to realize natural and harmonious HCI, we conduct the following research work by centering on FER and mechanical structure of humanoid robot head. Firstly, the automatic face region segmentation method of real time, compound local binary patterns(C-LBP), unsupervised learning of LBP feature selection and expression classifier design of matrix spectral regression analysis(MSRA) are studied in the process of FER, which provide recognition algorithms basis for intelligent HCI. Finally, the experimental system construction of expression representation for humanoid head robot is furtherly developed based on them and their effectivenesses are verified by experiments. The content in detailed is as follows.Face detection is the most important and necessary step of FER, which mainly provides accurate face region for object. The active shape model(ASM) is used to build models for face recognition according to the contour of face features, which reduce the influence of head pose change, remove the area outside face to improve rates of detection and extract face region. The landmarks on faces are applied to extract the key areas based on facial action coding system(FACS), and locate the face position and the feature region for FER exactly.It is great significant for FER to extract effective facial features. By research on “uniform” local binary patterns(LBP) form local feature operator, C-LBP operator is proposed by describing the tendency and extent of facial texture changes. It reveals the changes of data structure from multiple neighborhoods and directions to extract features in the main facial expression regions, which provides recognition vectors for FER.It is an effective technique that the recognition effectiveness can be improved by reducing the redundant information. Though LBP can extract the local texture features of images effectively, its dimensionality harshly grows companied with the number of pixels in the neighborhood region. Hence, the neighborhood graph in the feature space is firstly built by applying chi square statistic according to the histogram of LBP property in this paper. Then, the new feature subset is formed for FER by selecting more representative features according to 2LS- c scores calculated by applying binary weight matrix.Classifier design is the most important step of FER. Linear discriminant analysis(LDA), as a traditional statistical classification algorithm, cannot solve the nonlinear problem in the feature space effectively. Its computation amount is tremendous large for a great number of samples with high dimensionality in addition to the obsession from the singular values in the process of solving eigenvalues. Based on the theory of spectral graph and kernel learning, expression classifier MSRA is developed for multiple expression recognition after analyzing the relationship between linear coefficient in the kernel space and weighted matrix from LDA. It avoids calculating eigenvalues in the process of training classifier and reduces computation time and memory.It is very important that the intelligent service robot can perform facial expression for realization HCI of intelligence, emotion and humanlike. From bionics, the expression representation system of intelligent robot is constructed based on humanoid head robot and FER technique. Experiments of multiple movements on head mechanisms of this system and validations of FER algorithms in real situation are performed. Expression representation is realized by using expression class in real scene.In this paper, experiments for related FER techniques on databases and real situation are executed. The effectiveness and availability are validated on the proposed C-LBP operator, unsupervised feature selection algorithm for LBP feature and expression classifier MSRA. And the real-time and robustness of automatic face region extraction in real situation is validated. The well recognition effectiveness is obtained on databases and in real situation, and the real-time automatic extraction of face region in real situation is realized. Expression representation on humanoid head robot is achieved by FER results.更多还原