【作者】 陈伟；

【导师】 贺汉根；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2008， 博士

【摘要】 随着计算机科学、传感器技术、人工智能等学科的发展和制造水平的不断提高,移动机器人日益向着自主化的方向发展。移动机器人要实现自主化,其中的两个基本问题是自主定位和环境地图构建,这是移动机器人自主导航和环境探索的基础。定位与建图的精度是自主机器人能否在实际环境中成功应用的关键。本文围绕着单目视觉微小移动机器人在未知环境中的定位与建图进行了研究。重点是基于单目视觉的定位算法和粒子滤波SLAM(Simultaneous Localization andMap-building)算法及其在机器人上的实现。本文的成果和创新点包括以下几个方面:1)提出了一种新的单目视觉信息获取方法——视觉量角计。视觉量角计以环境标志点之间的视角作为所获取的信息。视角对机器人姿态具有不变性的特点,其大小只与机器人的位置有关,因而有利于实现机器人的位置跟踪。视觉量角计的提出使单目视觉在机器人定位中有了一种新的利用方式。2)提出了基于视觉量角计的卡尔曼滤波航迹修正算法。由于在微小机器人中受空间和载重能力的限制,当只有码盘和单目摄像机可用于定位时,本文提出了一种将两者的信息通过卡尔曼滤波相融合的定位算法。由码盘得到机器人初步位姿估计,同时单目摄像机以视觉量角计的方式获取环境信息,利用扩展卡尔曼滤波实现对码盘定位的修正。此算法避免了对环境标志点的三维计算,能较好的满足机器人定位的实时性要求,实验表明算法提高了定位精度。3)提出了基于视觉量角计的三角定位算法。针对微小机器人能独立获取航向信息的情况,提出了通过计算环境标志点坐标实现的三角定位算法。算法以获取的环境标志点坐标为基准,由稳定的视觉量角计信息实现对机器人位姿的最优估计。文中对算法误差进行了详细的理论分析,得出了定位误差上限的一个表达式,理论上保证了算法的可靠。同时算法得到了机器人导航实验的验证。4)针对未知环境中微小机器人的同时定位与建图问题,提出了利用单目视觉的改进粒子滤波SLAM方法。单目视觉图象结合码盘信息得到初步的环境标志点坐标。在一般的粒子滤波算法的基础上,调整了状态向量,使算法由一次高维滤波计算变为多次低维滤波计算。将多次获得的位姿取均值作为机器人的位姿估计。改进的算法大幅度减小了计算量,仿真实验表明此算法提高了定位精度的同时获得了更为准确的环境地图。5)将SLAM算法在自主研发的月球车原理演示样车上进行了性能测试,实现了月球车在非结构化环境中的定位与建图。针对月球车行走机构的运动特点,改进了状态转移方程,调整了粒子滤波中因月球车平面运动假设而导致扩大的误差采样范围。最后在月球车上对算法进行了验证,实验表明算法获得了使用特征点表示的环境地图,同时将定位误差减小了约三分之二,通过对比表明此环境地图整体与实际情况相吻合。更多还原

【Abstract】 With the development of computer science, sensor technology, artificial intelligence and the improvement of manufacture level, the robotics increasingly tends toward automation. Two of the essential problems to realize the automation of mobile robots are self-localization and map building. This is the foundation of autonomous navigation and environment exploring for mobile robots. The precision of localization and map is the key problem of whether mobile robots can be successfully applied in real environments or not. The intention of this dissertation is to describe the research on the localization and map building for monocular mobile robot in unknown environments. It principally introduced the algorithm of localization, the algorithm of SLAM based on particle filter, and the application of the algorithms to a mobile robot. The main contributions and innovation of this dissertation are as following:1) A new method of obtaining information from monocular camera called visual protractor is proposed in this dissertation. The information got from the visual protractor is the visual angle between two environment features. The visual angle has such a characteristic as invariability which does not relate to the pose of a robot, but merely to the robot’s location. This characteristic is very helpful for robot’s position tracking. The visual protractor provides a new method to use monocular camera for mobile robots localization.2) A algorithm based on EKF to correct the track of a robot has been presented. When there are only monocular camera and encoders in the robot for localization, the localization algorithm by fusing the information from above two sensors is described in this dissertation. The rough estimation of robot’s position is obtained with the encoders. The monocular camera is used as visual protractor to get environment’s information with which the algorithm corrects the primary position by EKF. The algorithm avoids calculating the coordinates of environment’s features, so it can well satisfy the real-time performance of the robot localization. The experiment showed the localization precision was improved by the algorithm.3) A triangulation localization algorithm with visual protractor also proposed in this dissertation. Under the situation that the robot can get its orientation independently, a new localization algorithm based on triangulation with visual protractor is presented. On the basis of calculated coordinates of environments’ land-marks, the optimal estimation of robot’s position can be obtained from the algorithm with the stable information from visual protractor. The algorithm is analyzed in detail and an upper limit expression is deduced, so the reliability of the algorithm is guarantied theoretically. This algorithm was testified by the experiment of robot navigation. 4) To the SLAM problem for a miniature robot in unknown environment, an improved SLAM method on particle filter is provided. The coordinates of land-marks are estimated imprecisely with the data from camera and encoders. On the basis of general particle filter, the state vector is adjusted to make the high dimension calculation become a few times of low dimension calculation. So the precise estimation of robot’s position is acquired by averaging the position estimations from particle filter. The improved algorithm reduced the cost of computation. The result of simulation experiment showed that the algorithm not only improved the precision of localization but also built a more accurate map.5) The SLAM algorithm has been testified on the lunar rover for principle demonstration and it made the robot realize localization and map building in an unstructured environment. According to the characteristic of the lunar rover’s movement structure, both the state equation and the extent of error enlarged by the hypothesis that the robot moved on a plane were adjusted. The algorithm was validated with navigation of the lunar rover. The experiment showed that the environment map expressed with feature points was obtained from the algorithm and the error of localization reduced to one third. The environment map accorded with the real environment by Compare between them.更多还原