【作者】 李科；

【导师】 朱齐丹；

【作者基本信息】 哈尔滨工程大学 ， 控制理论与控制工程， 2011， 博士

【摘要】 导航是实现机器人智能化和完全自主移动的关键技术。视觉导航具有探测范围广获取信息量大等优点,是移动机器人重要的研究方向。然而,传统的视觉导航方法需要建立环境的精确模型,解决地图创建与同时定位问题(SLAM),算法的空间复杂度较高,扩展应用困难。受昆虫类生物导航原理的启发,视觉归航方法将传统视觉导航中的定位与地图创建工作转化为对运行方向和停止位置的判断,不需要建立环境的精确模型,占用存储空间小,导航过程中没有累积误差,导航精度不受距离的影响,是视觉导航领域一个新的研究方向。本文采用全景视觉传感器对基于自然路标的机器人归航技术进行了深入研究。论文首先研究了全景图像中的自然路标提取问题。在非结构化环境中不存在人工路标,机器人要实现归航只能依靠环境中的自然路标。对比多种自然路标提取方法,图像局部特征具有多种不变性,特征点定位准确,非常适合作为自然路标点。SIFT和SURF是两种高效的斑点提取算法,以其优异的性能受到越来越多学者的关注。然而,这两种算法在全景图像中的特征提取性能,还没有被评估过。为了获得环境中性能稳定的自然路标点,本文采用重复率、匹配率、错误匹配率三个标准评估了SIFT和SURF在全景图像中的特征适应性。同时为了评估路标分布的均匀度,提出了特征分布均匀性的量化评估标准,并对SIFT和SURF在多种环境下的特征分布均匀性进行了评估。其次解决自然路标点之间的对应性问题。由于直接通过检索方法得到的最近邻匹配并不能保证完全正确,而错误的自然路标匹配会导致归航决策偏离正确方向,甚至会导致归航失败,因此必须对匹配结果进行提纯。原始匹配结果中存在多个路标对应一个路标的问题(多对一),文中提出了一种改进的匹配方法,消除路标匹配中的多对一问题,增强匹配的稳定性。针对全景图像中的错误匹配问题,提出了基于角度估算和基于最长公共子序列的误匹配消除方法,提纯自然路标。在完成了自然路标提取、匹配、提纯的基础上,本文给出了平均位移向量归航方法、平均路标向量归航方法的实现过程,首次完整表述了基于路标夹角差的归航方法。构建了完整的实验系统,采用推算式定位方法记录机器人在实验过程中的轨迹,对履带打滑进行补偿,利用全景视觉校正机器人的方位角,提高推算式定位方法的精度。在本文的实验环境中,对比了ADV、ALV和夹角差三种基于路标的归航方法,分析了每种方法的优缺点,并据此提出局部归航中的优化方法。针对ALV、ADV方法需要已知Home位置方位角的限制,采用基于角度估算的方法来减弱限制。对归航中的自然路标点进行优化,控制特征提取的数量、质量和均匀度。利用ADV和夹角差归航方法的优点,采用模糊控制策略,提出一种融合式归航方法。最后针对局部归航作用距离过短问题进行了初步研究。采用增加中间节点的方式,引导机器人沿着多个中间节点到达最终的Home位置。以拓扑地图的形式,组织中间节点,并使机器人在离开H0me位置时自动创建拓扑地图。更多还原

【Abstract】 Navigation is the key technology of achieving robot intelligence and autonomous mobile. Visual navigation can detect a wider range and get much more information, that is important to research mobile robot in the future. However, the precise model of environment must be established in traditional navigation methods. For example, SLAM(simultaneous localization and mapping) has a higher space complexity and more difficult extended application. Visual homing is a method of behavior-based navigation, which has been inspired by the principle of biological navigation. These methods change the simultaneous localization and mapping to determine the running direction and stop position. So it is more similar to intelligent life navigation. Visual homing does not need an accurate environmental model, and it occupies less storage space.. There is no accumulation error during navigation, so the accuracy is not affected from the distance. This paper deeply studied robot visual homing based on natural landmark used panoramic vision sensor.First, we studied on extracting natural landmark from the panoramic image. Because there are not artificial landmarks in the unstructured environment, so the robot can only rely on the natural landmarks of environment to achieve homing. Comparing a variety of methods of extracting natural landmark, local feature has many invariance and it is suitable as natural landmark points. SIFT and SURF are two efficient and robust spot extracting methods, so more and more scholars began to focus on them. But their performances in panoramic image had not been evaluated. In order to obtain stable natural landmark points, we evaluated the performance of SIFT and SURF using repeatability, match rate and mismatch rate. At the same time, a method of evaluating distribution uniformity of feature points has been proposed. And we assessed the distribution uniformity of SIFT and SURF feature points in different environment.Second, we calculated relevance of natural landmarks. Because mismatch will interfere with deviating from right direction during homing and even failed homing. So the matched natural landmarks must be purified. If using the previous matching results, there will be many landmarks matching to one landmark. And it is incorrect or unstable. In this paper, an improved matching method to solve this problem has been proposed. To solve the mismatching problem in panoramic image, two methods of eliminating mismatches have been also presented in this paper, which were based on angle estimation and the longest common subsequence. After completion of extraction, matching, purification of natural landmarks, the visual homing method based on natural landmarks were given, which included average displacement vector method, average landmark vector method. And the homing method based on angle difference was firstly completely described. Homing experiment system was constructed and moving track of robot was recorded by dead reckoning method. To improve the positioning accuracy, the robot tracked skid was compensated and azimuth error was corrected using angle estimation based on panoramic vision. We compared of ADV, ALV and the included angle difference based on landmarks in our test environment. Because ADV、ALV must have known(should know) the azimuth of home position, we reduced the constraints by angle estimation.Finally, we simply studied the long range visual homing. We increased the number of intermediate home targets, and used multiple intermediate nodes to guide the robot to reach the last home position. Intermediate nodes are organized in the form of a topological map, and the robot can automatically creates a topological map after leaving home position.更多还原