【作者】 李国栋；

【导师】 田国会；

【作者基本信息】 山东大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 随着科技的发展和人类社会老龄化趋势的日渐加剧,使服务机器人应用于家庭环境以完成聊天陪护、端茶递水等家政服务的诉求愈加迫切。以知识分布与智能分布为思想,以将摄像机、麦克风、温湿、瓦斯等传感器通过无线网络技术分布到环境中为手段的智能空间技术就是一种对服务机器人提供强有力支持,使之能够全面感知环境,正确理解人的意图,更为有效地为人类提供各种主动服务的技术。智能空间技术与服务机器人相结合,在扩展了服务机器人的感知和决策能力的同时,服务机器人作为智能空间中的可移动感知与执行设备,也丰富了智能空间的信息感知和服务执行功能。与人类通过眼睛获取外部环境80%有用信息的特点类似,智能空间通过摄像机这种核心传感器来感知环境信息,并利用机器视觉测量技术来完成诸如人的行为识别与理解、服务机器人的定位与导航、服务机器人对药品水杯等家庭物品的抓取与搬运等大部分家政工作。机器视觉测量技术是指根据不同视图的图像信息恢复出摄像机的欧氏运动(位姿信息)及景物的欧氏结构的图像处理技术,本文主要针对智能空间中的机器视觉测量技术及其应用进行了研究,主要研究内容和结果概括如下：(1)研究了世界平面单应矩阵和世界平面诱导的单应矩阵的计算及分解问题,这两种单应矩阵的分解是由2D图像信息恢复摄像机相对位姿和景物结构的有效途径。首先从Schur补的相关性质出发,给出了两种稀疏L-M迭代优化方法,并分别利用DLT方法和L-M迭代法对两种单应进行了计算；然后针对世界平面单应,在摄像机内参矩阵部分已知和完全未知的情况下,分别从单幅图像和多幅图像恢复出了摄像机的内参矩阵和摄像机坐标系相对景物平面固联坐标系的位姿,针对世界平面诱导的单应,提出了一种基于SVD的单应矩阵数值分解算法,利用这种算法可通过单应矩阵分解得到两组摄像机位姿和景物结构的物理可实现解。同时给出了景物结构已知情况下的单应矩阵分解算法,利用这种算法可直接得到唯一的摄像机相对位姿；最后推导了齐次图像坐标下的图像矩特征和齐次射影坐标下的图像矩特征间的转换关系,证明了两视图在2D仿射变换下其归一化图像几何中心矩为绝对对称逆变张量,并利用这一论断给出了一种全新的2D仿射单应的计算方法。(2)研究了服务机器人智能空间中的分布式摄像机系统对运动目标进行定位的问题。首先令机器人作几组已知的运动,并提取出对应的图像点,利用地面所在世界平面的单应分解方法对分布式摄像机的内外参数进行了标定；然后根据摄像机标定结果,设计了实现起来简单快速,但存在运动目标远离摄像机时精度下降缺点的运动目标单摄像机定位算法；最后引入了双目摄像机运动目标定位算法,针对利用简单的线性三角形法进行定位存在求得的运动目标点不在地面的缺点,进一步给出了带约束的线性三角形定位法,并为了进一步提高定位精度,最终采用了LM算法对检测到的运动目标的图像点进行优化,利用精化后的图像点和带约束的线性三角形法完成了运动目标的双目定位。(3)研究了基于顶棚激光投影循迹的室内服务机器人导航方法。首先建立了投影器的运动学模型,根据模型可以建立投影器各关节角位移、角速度和激光斑在地面的位移、速度间的非线性映射关系,同时为了将世界坐标系中的规划路径转换到投影器基坐标系中,提出了一种新的投影器外参数标定方法,完成了投影器的外参数标定,这样就可以使顶棚投影器在地面上投射出规划好的路径；然后将移动机器人抽象为一个三自由度机械臂,建立了它的运动学模型,并采用了一种基于世界平面分解的简单有效的方法标定出了机载摄像机的内外参数,同时计算出了以机载摄像机坐标系{c}为参考坐标系的地面所在的世界平面π；最后分别设计了自适应补偿跟踪控制律和非线性状态反馈控制律来控制移动机器人完成对运动激光斑的视觉伺服跟踪。(4)研究了服务机器人对QR Code人工地标这种智能空间中知识分布与智能分布的重要载体的定位与识读问题。首先按照功能将QR Code人工地标分为了用于全局语义地图生成的QR Code人工物标和面向局部导航地图生成的QR Code人工路标,并分别对两种地标的外围模式和内部信息编码方式进行了设计；然后分别利用QR Code外围模式的蓝色矩形框和红色同心圆部分给出了两种QR Code人工地标的定位方法；最后就求得的QR Code人工地标的位置信息构造了任务函数,并设计了基于位置的视觉伺服识读控制律,完成了服务机器人对QRCode人工地标的识读。(5)研究了搭载有机械臂的服务机器人对家庭环境智能空间中物品的搜寻与抓取、搬运操作。首先设计了辅助服务机器人抓取操作的QR Code人工物标,并对人工物标内的信息进行了编码,将对家庭物品的识别转换为对QR Code人工物标的识别,降低了服务机器人对家庭物品的操作难度；然后将搭载有机械臂的移动服务机器人抽象为一个广义机械臂,并为了完成此广义机械臂的统一控制,建立了广义机械臂的运动学模型,并分别利用解析法和数值法给出了其运动学逆解,同时提出了一种全新的机械臂手眼关系标定方法,这种方法通过令固联于机械臂末端执行器上的摄像机观测简单2D标定物,然后机械臂作两组任意运动,完成了手眼参数标定；最后考虑移动机器人运动的非完整性约束,设计了一种物品抓取的切换控制律：首先在眼注视的约束下逼近待操作物品,逼近到一定程度后,再利用单应分解加已知摄像机摄动来求取当前摄像机坐标系相对期望摄像机坐标系的位姿信息并切换到基于位置的look then doing视觉伺服控制方式,完成物品的抓取操作。更多还原

【Abstract】 With the development of technology and the intensified human society aging trend, the demands of using service robot in home environment to accomplish the domestic services such as chatting, accompanying, tea or water delivery have become increasingly urgent. The intelligent space technique, who uses knowledge distribution and intelligent distribution as the main thoughts, and uses the distribution of camera, microphone, temperature sensor, humidity sensor and gas sensor through wireless network technology into the environment as the main means, provides strong support for a service robot to fully perceive the environment, correctly understand people’s intent and more effectively provide active service. The combination of intelligent space technology and service robot technology extends the robot’s perception and decision-making ability. At the same time, as a removable perception and implementation equipment, the service robot enhances the functionality of intelligent space for information perception and service tasks execution.Human access to80%of the external environment information through eyes, similar to this characteristics, the intelligent space uses camera as the core sensor to perceive the environment information, and uses machine vision measure technique to finish most of the service tasks such as recognition and understanding of human behavior, localization and navigation of service robot, handling and delivering of medicines or cups and so on. The machine vision measure technique is a kind of image processing technique which is used to restore the cameras’ Euclidean motion information (position and orientation) and the3D scene structure information through different image views. This dissertation focuses on the machine vision measure technique and its applications in home environment intelligent space, the main contents and results are summarized as follows:(1)The computation and decomposition of two different kinds of homography, which are called homography from a world plane and homography induced by a scene plane, are discussed. The decomposition of the two kinds of hompgraphy is an effective way to acquire the camera relative pose and scene structure from2D image information. Firstly, based on the properties of schur complement, two different types of sparse LM iterative optimization methods are given, and computations of the two kinds of homography are given using the DLT method and LM iterative method respectively. Moreover, for the homography from a world plane, the camera intrinsic parameters, camera relative pose and scene structure are restored using single view or at least three different views in case of camera intrinsic parameters partly known or completely unknown respectively, for the homography induced by a scene plane, a singular value decomposition algorithm is proposed to acquire two solutions verifying a physical constraint, meanwhile a homography decomposition algorithm is given in case of knowing the scene structure, using which we can acquire the unique solution. Finally, the conversion relationship between homogeneous image coordinates based image moments and homogeneous projective coordinates based image moments is derived and the fact that the normalized central moments of the image are absolutely symmetrical inverter tensor is proved under the2D affine transformation, which is applied to educe a new method to compute the corresponding2D affine transformation.(2) The problem of using distributed camera system to locate the moving targets in service robot intelligent space is discussed. Firstly, the robot set to do several known movements, and extract the image correspondence. By using the decomposition of homography from a world plane, the internal and external parameters of the distributed camera system is calibrated. Then using the calibration results, a simple but fast monocular camera based moving target localization algorithm is designed, but the shortages of this algorithm reduces the accuracy when moving target is away from the camera. Finally a binocular camera based moving target localization algorithm is introduced, for the shortcomings of solved moving target’s position not on the ground using simple linear triangular method, a constrained linear triangular method is further given, Using the constrained linear triangular method and the LM method refined image points, the binocular camera based moving target localization is accomplished.(3) The problem of using ceiling projection navigation system to realize the visual track navigation method is researched. Firstly the kinematic model of the ceiling laser projector is established, using the model the nonlinear mapping relationship between projector’s angular displacement, angular velocity and laser spot’s linear displacement, the linear velocity on the ground can be acquired. At the same time in order to transform the planned route from the world coordinate frame to the projector’s base coordinate frame, an original projector’s external parameter calibration method is proposed. Secondly the mobile service robot is considered to be a3-DOF manipulator, its kinematic model is established, the internal and external parameters of the camera mounted on the robot are calibrated using decomposition of hompgraphy from a world plane, meanwhile consider camera frame as the reference coordinate frame, the3D parameter of ground plane is computed. Finally the adaptive Compensation Tracking Control Law and the Nonlinear State Error Feedback Tracking Control Law are designed respectively to control the service robot so that it can finish the visual servo tracking of the moving laser spot projected on the ground.(4) The problem of localization and recognition of the QR Code based artificial landmark is discussed, which is an important carrier for knowledge distribution and intelligence distribution. Firstly The QR Code based artificial landmark is divided into two categories:the QR Code based artificial object mark used for Global Semantic Map Building and the QR Code based artificial signpost used for Local Navigation Map Building, then the external pattern and internal information encoding are designed respectively. Moreover, by using the blue rectangle area and red circular ring area respectively, two different kinds of QR Code based artificial landmark location mthods are given. Finally, a task function is constructed according to the solved localization information and a position based visual servo control law, which is designed to control the service robot to realize the recognition of the QR Code based artificial landmark.(5) The problem of searching, handling and delivering of the objects dispersed in intelligent space for service robot with one manipulator is researched. Firstly, the QR Code based artificial object mark is designed, which is used to assist the service robot to finish the handling task, and the internal information is coded so that by converting the identification of household objects to the identification of corresponding QR Code, the object mark can reduce the difficulty of the operation for service robot effectively. Moreover by considering the mobile service robot equipped with a manipulator as a generalized manipulator, the kinematic model of the generalized manipulator is established, and certain analytical and numerical methods are given to solve the inverse kinematics, and an original manipulator’s hand-eye relationship calibration method is proposed as well. Finally take the mobile robot’s non-holonomic constraint into account, a switch object handling control law, which is composed of approximating control law and position based look then doing visual servoing control law, is designed. Under the approximating control law’s effect, the service robot approximates the object to be operated under the constraints of gazing, then using the decomposition of homography induced by a scene plane and the known camera motion, the position and pose information between current camera frame and desired camera frame can be computed, and the control law switches to the position based look then doing visual servoing control mode so that the handling operation can finally be accomplished.更多还原