节点文献
基于服务任务导向的机器人地图构建研究
Map Building Research of Robots Based on Service Mission Orientation
【作者】 吴皓;
【导师】 田国会;
【作者基本信息】 山东大学 , 控制理论与控制工程, 2011, 博士
【摘要】 随着服务机器人越来越多地走入我们的工作和生活,给人提供的服务种类也越来越广,如端茶送水、清运垃圾、货物搬运、传送信件等,因此构建适用于机器人服务任务的环境模型成为当前地图构建研究的热点。机器人传统的地图建模都侧重于空间几何结构的表示,适用于导航,但没有考虑机器人服务工作的区域性特征和局部服务空间的复杂性,更没有与人交互的语义信息。要使服务机器人具有智能,应从人对空间环境的表示方式上获得启发,研究能理解人的语言信息的服务机器人环境空间表示方法。人对环境的认识是分层次进行的:对于大空间环境构建了走廊和房间概念,而人操作的对象——物品的位置与房间应有必然关联;单位房间内大物品的摆放位置为人提供区域路径规划依据;小的局部复杂空间中物品的搜寻和抓取依靠人的视觉获得的立体空间模型完成。因此,模仿人的空间模型构建模式,针对机器人的物品管理、物品移动、物品操作三级服务任务导向,构建室内未知环境的三层环境地图是本文的研究重点。基于全局语义、区域规划和局部描述的三层环境地图,形成三级递阶规划系统,使服务机器人既能获得未知环境的房间分割模型,确定房间的功能及物品与房间的归属关系,又能实现走廊、房间之间和房间内的路径规划和导航,同时建立其操作对象及其周边环境的三维地图,使机器人能基于服务任务的导向,按“语义”进行逻辑推理,确定“目标”及其位置,为以服务任务为驱动的基于分层环境地图构建模式的“智能”导航奠定基础。由于多机器人系统具有的可拆分性、可重构性、容错性和鲁棒性等优点,多机器人协作构建地图的研究逐渐展开。本文在对三层分级环境地图构建的基础上,也对多机器人协作中备受关注的探测策略和定位问题进行了初步研究和探讨。本文的研究思路分四步展开。首先在总结分析当前环境地图构建的现状和问题基础上,提出基于服务任务导向的三层环境地图构建模式:反映房间拓扑关系的全局语义地图;描述区域内物品摆放关系的区域混合地图;勾绘局部复杂空间的三维栅格地图。然后引入基于QR code技术的自相似二维人工标签——人工物标和人工路标的设计及功能,阐明人工标签在三层环境地图构建中的作用。据此,给出三层环境地图构建模式的具体创建机制及其实现过程:利用双目视觉基于DSmT证据理论构建含体素概念的三维栅格地图;基于SIFT特征匹配算法形成无向加权图,再采用谱聚类算法构建具有房间分割功能的小范围空间语义地图;基于免疫网络算法构建“自主分布式表征”机制下的大范围结构化环境的认知语义地图。最后,初步探讨多个服务机器人同时构建地图时的协作策略及从研究中获得的启示。本文的主要研究内容和结果概括如下:1.针对单纯依赖本体视觉实现场景识别和理解的复杂性和局限性,提出基于QR code技术的新的室内环境空间认知手段,将移动机器人对环境的理解从几何结构的、模糊的、被动的感知层次,提高到语义的、精确的、主动的认知层次上,形成基于QR code技术的人工标签提供语义信息的地图构建新模式。设计了含外围模式和内部编码两部分的嵌入式人工标签,解决人工标签易被遮挡、QRcode编码远距离难以识别等问题;设计了含物品、房间、走廊等的功能属性和归属关系的QR code编码信息描述规约,为三级室内环境地图的构建提供丰富的语义信息。服务机器人基于高斯模型和椭圆拟合算法搜寻人工标签;基于模糊调整策略和投影透视原理逼近和对准人工标签。2.针对室内移动机器人的服务任务,提出一种包括全局语义层、区域规划层、局部描述层的三级室内环境模型,使机器人对环境信息的掌握不仅在面向导航的几何结构上,还增加了局部复杂空间的三维栅格信息及描述走廊、房间和物品功能及关联、归属关系的语义信息,并按机器人服务的活动区域和任务执行特点形成各层不同的地图表示形式。首先,对于房间内的小范围环境,机器人依据视觉获得的深度信息及QR code标签提供的物品操作功能信息进行三维栅格地图和物品功能图的构建,形成局域层空间描述;其次,基于贝叶斯估计算法构建区域层的二维栅格地图,同时形成无向加权图,构成区域规划层;最后,基于谱聚类算法构建具有房间分割功能的拓扑地图,结合物品功能图,获得房间功能及房间之间关联关系、物品与房间归属关系等的语义信息,形成全局语义拓扑地图。对于走廊的大范围环境,在获取QR code人工路标提供的语义及引导信息基础上,实现大范围区域的路径导航和认知语义地图的构建。3.针对室内局部区域的错层结构、立体装修、非结构化家具等复杂环境,构建三维栅格地图。在双目视觉获得深度信息的前提下,基于DSmT证据理论构建双目视觉信息不确定数学模型,并将新、旧信息按比例冲突分配规则进行融合,形成描述体素占有/空闲概率的三维栅格地图。基于特征点关系的稠密匹配获得三维场景的稠密重建,而体素对应匹配点的参数优化能获得最优占空值。在构建三维地图的同时,利用粘贴在大物品上的基于QR code技术的二维人工物标,为环境中的大物品添加语义标签,并基于大物品的尺寸更新对应的体素占空值,形成含大物品功能属性和归属关系描述的三维栅格地图。4.针对房间内的小范围半未知环境,根据机器人特有的服务任务和人机交互要求,结合大物品上粘贴的基于QR code技术的人工物标,构建半未知环境的以房间为单位的功能语义地图。首先基于谱聚类算法构建具有房间分割功能的拓扑地图。再利用QR code中存储的物品信息,建立物品信息库和物品归属关联图,最终获得包含物品信息描述、房间功能描述及物品-房间归属关系的语义拓扑地图。该地图为室内环境中物品的搜寻、管理和机器人服务提供完备的、拟人化的信息。基于语义地图机器人能理解人的语义命令,生成合理的服务路径,实现功能驱动的导航。5.针对走廊等大范围结构化半未知环境,提出了基于“自主分布式表征”机制的认知语义地图构建的思路和方法。按照人的路标导航模式,将基于QRcode技术的人工路标分布于环境的关键点处,形成认知导向点。机器人在大范围结构化环境中不断获取认知导向点处的人工路标信息,实现对环境的感知;从感知的信息中提取引导信息,确定目标和路径;对获取信息的逻辑分析,产生运动的行为指令。基于免疫网络算法形成“分布式表征”的环境认知机制,并基于颜色识别和SIFT特征匹配算法实现情景标签的记忆和认知,构建自主分布式的感知-引导-行为网络地图。按认知语义地图实现基于免疫算法的路径重建。6.针对多机器人协作构建地图的任务,提出了免疫网络探测算法的协作探索策略和基于无线传感网络的机器人融合定位算法。基于观测点融合的改进免疫网络探测算法(INEA),完成多机器人对未知区域的快速、高效的探索任务。该算法在大大减少通信量,又能准确计算各全局观测点花费的情况下,将不同机器人的局部观测点融合在一张地图上,使机器人的协作能力充分发挥。利用T细胞函数修正了免疫网络浓度模型,同时免疫模型参数考虑各观测点的扩散度和探索方向对系统性能的影响。基于RSSI算法对无线传感器网络节点测得的距离数据进行处理,融合基于粒子滤波算法获得的机器人位置数据,实现多机器人的协作定位。
【Abstract】 The emergences of service robots in our work and life bring us lots of conveniences more than ever before, such as tea service, garbage removal, freight handling, letter transmission and so on. Therefore, the environmental modeling of robots for service tasks has become the hotspot of current research for map building. The traditional methods of robot map building were focused on the description of spatial structure in spite of considering the functional characteristics of the environment that robots worked in and the complexity of the local space. Besides, they also ignored the semantic information to achieve communion with human. In order to make a service robot be intelligent, we should get inspired from the representation method of the space environment used by people, and study the spatial representation of service robots to understand human language information. Human understanding of the environment is carried out at different levels:the concepts of corridors and rooms are constructed for the large space, and the locations of the object operated by people must be associated with the room; the placement of large items in unit room should provide human the basis of region path planning; the searching and grasping of the objects in local complex space can be achieved by the accomplishment of three-dimensional space model receiving from human’s eyes. Therefore, this thesis will focus on the imitating spatial modeling method of the human, aiming to the three-level service task direction for the object management, moving and operations, and building indoor three-level environment maps in unknown environment. According to the three level environment maps of global semantic, regional planning and local description, three level step-up planning system is formulated, which makes the service robot can not only get room segmentation model for unknown environment to determine the accessorial relationship between room and objects, but also complete path planning and navigation in the corridor, the room and so on. At the same time, the three-dimensional map is built around the operated object. Based on the service mission direction, the robot can use "semantic description" to realize the logical reasoning, and determine the "target" and its location. It will lay the foundation for "intelligent" navigation which is driven by the service mission and based on hierarchical environment map modeling mode. As the multi-robots system have many advantages such as being splitted, reconfigurability, fault tolerance and robustness, the study of building map by multi-robots system is being unfolded. In this thesis, on the basis of three-level map building pattern, collaboration strategy and positioning problem of multi-robot system are discussed preliminarily.The main ideas of this thesis are divided into four steps. Firstly, aim at analyzing the status and problems of the current environment map building, a new three-tier space building pattern is proposed basing on robot service mission direction:the global semantic map reflecting room topology relation; the region hybrid map describing object location relation; three-dimensional grid map drawing local complex space. Then, some two-dimensional artificial tags (artificial object-mark and artificial signpost) based on QR code (Quick Response Code) technology are designed, and the function of artificial labels in the three-level environment map building is also discussed. Accordingly, the specific created mechanism and implementation process of three layer environment model are given:building three dimensional grid map based on DSmT(Dezert-Smarandache Theory) evidence theory which contains voxel concept and uses binocular visual; undirected weighted graph based on SIFT (Scale-Invariant Feature Transform) features match algorithm are formed and small range space semantic map is built basing on spectrum clustering algorithms which has room segmentation function; the cognizing semantic map of large range and structural environment is built basing on immune network algorithm which includes "independent distributed representation" mechanism. Finally, the preliminary study for collaboration strategies of multi-robots is done and inspiration from the research is introduced. The main research contents and results are shown as follows:1. A new environment cognition method based on QR code technology is proposed to solve the complexity and limitations of visual recognition and scene understanding which relies only on the robot’s vision. This method can extend the mobile robot’s understanding of the environment from the geometrical structure, vague, passively perceptive level, to semantics, accurate, actively cognitive level, and create a new model of building map with an artificial label providing semantic information. Artificial labels are designed with the external model and internal code to solve the problems such as labels sheltered and identifying QR code difficultly from long-range distance; containing items of QR code information description are designed to provide plentiful semantic information which includes the functional properties and adscription relations of objects, rooms, corridors. Service robot searches artificial labels based on Gaussian model and the elliptical fitting algorithm; approaches and aims at the artificial labels based on fuzzy adjustment algorithm and perspective projection principle.2. Three-level indoor space maps including global semantic layer, region planning layer and local describing layer are built for robot service mission. Use this space mapping pattern, the robot can not only know the plane structure of the environment for navigation, but also add three-dimensional grid map of local complicated space and semantic information which can describe the function, relationship and ascription of the room and the object. The different maps of every layer are formed according to the service region and task execution characteristics of the robot. Firstly, as the small scope inside rooms, depth information acquired by vision and object function information acquired by QR code label are used to build a three-dimensional grid map and an object function map which describe local space. Then a planar grid map is built basing on Bayesian estimate algorithm, and a non-direction-power-map is formed as well. Therefore, the region planning layer is achieved. Lastly, room-division topology map is built basing on clustering algorithms, semantic information including room functional information, relationship and object attributive relation are obtained, which constitute global semantic topology map. In the large scope as corridors, the navigation and cognizing semantic map building are realized according to the semantic and navigating information achieved from QR code artificial signpost.3. According to the complexity of indoor environments, such as staggered floor and tridimensional fitment and irregular furniture, a three-dimensional map is built for local space. Under the premise of gaining depth information with binocular vision, the uncertain mathematical model of binocular vision information is structured basing on the DSmT evidence theory, and the new and old information are fused by proportional conflict redistribution rule, then a three-dimensional grid map is formed which describes the occupied/free probability of the voxel. The dense matching based on feature point relations for image must be done to archive dense rebuilding of 3D scene, and the parameter optimization of the matching points in the voxels must also be done to calculate the optimal duty value.While structuring the three-dimensional map, QR code based object marks plastered on large objects are used to give semantic information, and the occupied/free values of the corresponding voxels are updated based on the large object’s dimension, then three-dimensional grid semantic map is formed which includes the function property and attributive relation of the large objects.4. Aim at robot special service tasks and man-machine conversation in the small semi-unknown environment as rooms, a functional semantic map is built using QR code based object mark plastered on large objects. First, a room-function topology map is built basing on the clustering algorithms. Then object information database and adscription relationship map are set up basing on object information stored in the QR code. Finally, a semantic map including object information description, room functional information and attributive relation between room and object is formed, which gives complete and personified information for object location, object management and robot service in indoor environment. The simulation results show that the service robot using semantic map can understand human semantic statement, produce reasonable service path, and achieve function-driven navigation.5. For large-scale structural and semi-unknown environments such as corridors, the ideas and methods of cognitive semantic map building are proposed basing on the "independent distributed representation". According to the navigation patterns of the human signpost, the artificial signposts based on QR code technology are distributed in the key points of the environment to form cognitive-oriented points. The robot constantly receives the artificial signpost information at cognitive-oriented points in a wide range of structured environment to achieve the perception of the environment; extracts leading information from the perception information to identify goals and paths; logically analyses the obtained information to make motion instructions. Immune network algorithm is used to form the environmental awareness mechanism of "distributed representation", color recognition and SIFT feature matching algorithm are used to achieve the memory and cognition of scenario tag, then the Cognition-Guide-Behavior Map is built as "independent distributed representation". The cognitive semantic map is used to path reconstruction by immune algorithm.6. For the task of building maps by multi-robots cooperation, the exploration strategy based on immune network exploration algorithm and fusion localization algorithm based on wireless sensing network are brought forward in the thesis. An algorithm named immune network exploration algorithm(INEA) based on observation point fusion is proposed to quickly accomplish m’ulti-robots’ exploration task for an unknown environment. In the case of reducing communication greatly and calculating general observation point costs efficiently, local observation points of individual robots fuse on a map, so robots’ collaboration ability is incarnated enough. Furthermore, T-cell function is used to update the immune network concentration model, and it is taken into account that system performance is affected by diffuse degree and exploring direction of observation points. Simulation results are provided to validate the efficiency of the complete exploration. The distance measured in WSN is calculated basing on the received signal strength indicators (RSSI), and location data received by particle filter algorithm. They are fused to realize multi-robot collaborative localization.
【Key words】 QR Code; Three-Level environment modeling; Semantic Map; Three-Dimensional Grid Map; Multi-Robots Coordination;