【作者】 郑重；

【导师】 汪增福；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 在移动机器人的各项研究中,导航是最基本及最重要的问题,移动机器人必须具备感知环境与自身状态,并实现在有障碍物的环境中面向目标的自主运动的能力,因此导航包含了面向目标的运动及运动过程中的实时避障二个基本任务。本文以视觉传感器为基础,以视觉归巢及障碍检测为主要的研究内容,对导航的二个基本任务进行了深入的研究。视觉归巢基本思想来源于生物学中昆虫导航的研究,它通过比较当前图像与目标图像计算返回目标的方向并以此指导机器人运动。相较传统的基于同步定位及构图的导航方法,该类方法需要的存储量及算法的复杂度都大大降低,因此成为极具吸引力的导航策略。然而目标图像是事先存储,目标与当前图像的拍摄时间可能相差很大,同时两幅图像的拍摄距离也可能较长,路标在图像间存在光照、尺度、旋转等多种变化,因此如何在光照强烈变化的情况下保持算法的实时性及鲁棒性依然是当前该领域面临的难题。本文首先构建了完善的视觉归巢算法的测试方法及平台,对当前主要的视觉归巢算法进行了详细的测试分析及比较,同时揭示了视觉归巢主要的两类算法:变形算法与RMS算法的内在联系,并在此基础上设计并实现了一种基于随机思想的变形算法变体,在保持变形算法性能的基础上,将计算效率提高一个数量级。为解决光照变化带来的影响,本文提出一种边缘高斯滤波器,它结合了基于特征提取匹配算法与基于原始图像算法的优点,借助于图像整体匹配的思想,将所有特征的位置信息转化为一幅新的图像,并利用滤波图像之间的量化差异来指导归巢任务。一方面通过特征提取来避免受到光照变化的影响,另外一方面避开了复杂及易错的特征匹配过程。在光照剧烈变化的情况下,算法显示了良好的性能。此外,针对非结构化环境下的实时障碍物检测问题,提出了一种基于角点的立体视觉障碍物检测方法。为提高角点提取与匹配的效率,采用基于Bresenham圆形模板和Moravec算法的两步角点检测与确认方法,并利用稀疏模板与极性约束进行角点匹配。考虑到单纯基于高度的障碍物的定义及其检测算法在月面等非结构化环境下不再有效,采用了一种基于角点间空间关系的障碍物定义,并设计了相关的障碍物检测算法。该算法具有以下特点:能实时运行于复杂的非结构化环境,且不需要环境的先验知识。实验结果证明了算法的有效性。更多还原

【Abstract】 Navigation is the most important and basic problem for mobile robot. It is necessary for mobile robot to have the ability of sensing extern environment and moving autonomously and safely in environment with obstacles. Thus navigation includes two basic tasks: goal-oriented movement and obstacle detection. In this dissertation we present our research on visual homing and obstacle detection.The basic idea of visual homing comes from insect navigation, visual homing calculates the direction from current location to goal location by comparing the images viewed from the two locations. Compared to the traditional navigation methods based on accurate geometric map and self-localization, the storage and complexity of visual homing are greatly reduced, thus visual homing has become an attractive strategy, however, the goal image is stored in advance, the acquisition times of current image and goal image are different; and the distance from current location to goal location are maybe long, the projection of landmarks in two images are different in light intensity、scales、rotation, so how to maintain the efficiency and robustness of visual homing algorithm is still a challenge problem.In this dissertation a test platform for visual homing algorithms is designed, the major visual homing algorithm are tested and compared detailedly,and the links between warping algorithm and RMS algorithm are discussed, which inspires a new method based on random searching, this algorithm can significantly improve the Computational efficiency with keeping the robustness in case of the original warping algorithm. Compared with the original warping algorithm, the computational cost of ours is about one order of magnitude lower.In order to alleviate the damage to the performance of visual homing algorithms caused by illumination changes, a new filter is proposed, which combines the ideas of feature extraction and origin image comparison, turn the coordinates of all edges in image plane into a new image, and calculate the homing vector from difference of filtered Images,which uses feature extraction to alleviate the impact of illumination changes, and avoid the complicated and error-prone process of feature matching. Experiments show good performance of this algorithm with dramatic illumination changesIn order to solve real-time obstacle detection problem in unstructured environment, a stereo vision algorithm based on corner features is presented. A fast two-step corner detection and confirmation algorithm is realized by using the Bresenham Circle template and Moravec filter, sparse filter and corner polar constraint are used for corner matching to reduce computing time. Considering that the obstacle definition by height can not be applied in unstructured environments, a new obstacle definition based on the space relationship of the corners is proposed, and the algorithm solve the problem of errors of corner matching. The algorithm is efficient, which do not need prior knowledge of the environment, and can be applied in unstructured environments. Experiments prove the effectiveness of the algorithm更多还原