【作者】 王璐；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 计算机应用技术， 2007， 博士

【摘要】 本论文来源于国家自然科学基金重点项目“未知环境中移动机器人导航控制的理论与方法研究”(60234030)，属于其子任务“环境建模与定位”的研究方向。本文着重研究了移动机器人在未知环境中基于摄像机视觉的拓扑建模、定位及导航等问题，目的在于探索针对未知环境的视觉建模与定位方法，提高所建立拓扑模型的准确性，降低模型的存储需求，从而提高机器人在未知环境中的自定位精度和长时间生存能力。有效分析视觉传感器信息以获取环境特征是视觉环境建模与定位的至关重要的基础，本文首先基于视觉显著性研究了针对未知环境的自底向上的自然路标检测、表示与识别方法。提出了一种带反馈机制的显著性检测模型(feedback saliency detection model，FSDM)，并定义了能够更好地体现对比含义且充分考虑全局信息影响的级间对比度算子。在多尺度空间上计算颜色、纹理特征的对比度，经综合处理得到描述路标候选位置的显著性指示图。通过引入反馈通道，控制各类特征在不同场景图像处理中发挥的作用。基本性能测试、稳定性测试及抗干扰性测试等一系列实验表明，该方法具有非常优越的显著位置重复检测性能。此后，应用LOG算子进行自动区域尺寸选择，形成显著区域作为自然路标原型，选择梯度方向、二阶不变矩、归一化的色调作为自然路标的表示方案。目标识别实验表明，基于局部显著区域的自然路标稳定性好，能够容忍远近尺度、视角等变化引起的图像差异，识别准确率较高。在前述获得的自然路标的基础上，基于隐马尔科夫模型(hidden Markov model，HMM)提出了一种增量式视觉拓扑建模与定位方法。该方法首先通过单CCD摄像头扫视机器人当前所在环境来获得全方位图像序列，然后利用HMM建模所获得局部显著路标间的关系，以构造拓扑节点，于是定位问题可以转化为HMM的估值问题。设计了基于联合概率分布的初始定位方法，并且设计了基于最大后验概率(maximum a postefiori，MAP)的学习策略来处理定位的不确定性。该方法具有以下鲜明特点：采用局部图像特征而不是整幅图像特征参与环境识别，并利用HMM建模这些路标间的空间关系，因此在定位时能够对环境变化具有更高的容许程度，识别更加可靠；HMM的状态空间恒定，不会随着探索环境规模的扩大而增长，从而降低了概率定位的计算需求；支持在线增量式建立拓扑地图，同时进行定位，而目前大多数视觉拓扑建模系统都需要离线分析建模和在线定位两个阶段。实验表明，上述方法能够有效地提高场景识别的准确率，实现在线增量式拓扑建模与定位。为提高视觉拓扑建模与定位方法在包含各种动态因素的真实环境中应用的能力，本文侧重研究了存在运动物体(或称为运动目标)的动态未知环境拓扑建模，尽量消除其对建模的影响，以提高拓扑环境模型的准确性。首先提出了带运动补偿的运动目标检测方法，并从检测与跟踪任务的完整性考虑，以Kalman滤波器为基本单元构造了一个驱动摄像头的反馈控制系统。为快速提取完整目标，对模糊C均值聚类(fuzzy C-Means clustering，FCM)方法进行了改进。在此基础上，重新修改了自然路标提取策略：将所有检测出的路标划分为静态路标和动态路标两类，然后抛弃那些动态路标以滤除运动目标影响。实验结果表明，该方法能够有效地滤除噪声路标，提高了拓扑建模和定位的精度。在前述自然路标提取、增量式建模及动态目标检测研究的基础上，提出并实现了一个视觉拓扑建模及导航系统VOTMNS(Vision based Online Topological Mapping and Navigation System)，并在移动机器人平台MORCS-1上进行了实践。该系统主要包括自然路标提取、建模与定位、地图及路标库管理、规划等四个部分。根据可用性评价，提出了基于竞争学习的路标库管理方法，与更新全部路标存在状态的方法相比管理更加有效(提高了路标的使用率)，计算代价较小。设计了基于初始定位的导航不确定性管理方法，使机器人能够在导航过程中定位到非路径节点时重新进行规划。实验结果表明该系统能够在比较平坦的室内外未知环境实现稳定的环境建模及安全导航，路标库管理及系统的时间性能都表明该系统具备实时工作能力。更多还原

【Abstract】 The thesis was supported by NSFC key project "Research on theoriesand methods for navigation control of mobile robots under unknownenvironments"(60234030). Related works was subject to the part "mappingand localization" of the project. The thesis focuses on the problems ofvision(camera) based topological mapping、localization and navigation ofmobile robot in unknown environments. The objectives are exploring newapproaches of vision based mapping and localization in unknownenvironments, improving the accuracy of created topological map,decreasing the storage requirement of map, advancing the precision ofself-localization and ability of long-time surviving of mobile robot inunknown environments.It is absolutely necessary for vision based mapping and localizationthat images are analyzed to obtain features of environments. The thesisfirstly researches the approaches of natural landmark detection andrepresentation in unknown environments from bottom to up based on visualsaliency. A saliency detection model with feedback mechanism is presented.An opponency operator among scales is defined, which consideres theinfluence of global information. So the opponencies of color and textureamong multi-scales are computed and combined to obtain the saliency mappointing out candidate natural landmark’s position. By the feedback, thecontribution of each feature to image analysis on different scenes arecontrolled. Experiments such as basic performance test and stability testand anti-jamming test show that the approach has excellent repeatability ofsalient position detection. Then LOG operator is applied to selectappropriate size automatically to create salient region as natural landmarkprototype. The features involving gradient orientation and moment andcanonical hue are used to represent natural landmarks. Experiments ofobject recognition show that the natural landmark based on local salientregion has high stability and tolerance of image diversity caused by scaleand viewpoint etc changed. The accuracy of recognition is higher.After the natural landmarks are extracted, an incremental vision topological mapping and localization approach is presented based onhidden Markov model (HMM). Firstly a single CCD camera is drivenscanning current environment to obtain omni image sequence. While thesalient landmarks are detected from these images, HMM is used to modelthe space relationship among these landmarks and create a topological node.So localization problem can be transformed to be the evaluation problem ofHMM. A initial localization method based on joint probability distributionis designed. A leaming strategy based on maximum a posteriori (MAP) isalso designed to deal with uncertainty of localization. The approach hassome characteristics as follows. Local salient image features replace wholeimage features to contribute for environment recognition. And HMM isused to capture the space relationship among them. So localization (placerecognition) has more tolerance of environment change and more trustiness.The state space of HMM is invariable with the scale of exploringenvironments increasing, which decreases the computation requirement ofprobabilistic localization. The approach supports online incrementaltopologically mapping and locating simultaneously. While most visionbased topological mapping and localization systems have two distinctivestages of training for mapping offline and locating online. Experimentsshow that the approaches can improve effectively the accuracy of scenerecognition and realize online incremental topological mapping andlocalization.To improve the applicability of vision based mapping and localizationapproach in dynamic natural environments, the thesis studies the problemof topological mapping in dynamic unknown environments existingmoving objects. The objective is to eliminate the influence of movingobjects on mapping and improve the accuracy of topological map. Firstly amoving objects detection method with movement compensation ispresented. Considering the integrality of detection and tracking task, afeedback control system for driving the camera is built based on Kalmanfilter. To extract integral object fast, a modified fuzzy C-Means clusteringmethod is presented. Based on these the strategy of extracting naturallandmarks is modified as: classifying all detected landmarks as static landmarks and dynamic landmarks, then abandoning those dynamiclandmarks to avoid the influence of moving objects. Experiments show thatthe approach can filter noisy landmarks and improve the precision oftopological mapping and localization.Founded by the research of natural landmark extraction andincremental mapping and dynamic object detection, a vision based onlinetopological mapping and navigation system (VOTMNS) is presented andimplemented on mobile robot MORCS-1. The system includes 4 part ofnatural landmark extraction, mapping and localization, management of mapand landmark library, planning. In terms of evaluation of usability, amanagement method of landmark library based on competitive learning ispresented. Compared with the method updating all landmark’s existingstate, this method has more efficiency and fewer computation cost. Amanagement method of uncertainty in navigation is designed based oninitial localization, which make robot plan again when locates to a node notbelonging to the planning path during navigation. Experiments show thatthe system has the ability of mapping stably and navigating safely in flatunknown environments. Landmark library management and the timeperformance of the system indicate that the system has the ability ofreal-time work.更多还原