【作者】 侯建；

【导师】 齐乃明；

【作者基本信息】 哈尔滨工业大学 ， 飞行器设计， 2007， 博士

【摘要】 根据我国的月球探测计划,在利用环月卫星对月球进行初步的探测之后,将需要使探测器在月球表面软着陆,并由月球车执行较大范围的详细探测任务,为载人登月和建立月球基地做好准备。月球车是各类探测仪器的平台,其安全的漫游是实现一切后续探测任务的基础,而安全漫游的实现需要准确的地形和自身位置信息,这就是本文视觉与导航系统的任务。作为室外移动机器人的一个特例,月球车的视觉与导航技术来源于通用视觉与导航技术,但又具有自身的特点,如对可靠性要求极高,希望系统的总质量、能耗等尽可能小,希望视觉处理速度尽可能快等。基于上述分析,本文选择被动的双目立体视觉作为视觉与导航系统的实现模式,根据月球车漫游任务要求,对系统硬件组成和配置参数进行了详细的分析,给出了一种配置方案。摄像机标定方面,因视场要求采用宽视场镜头,而较大的场景深度要求在较远距离处保证一定的精度,因此标定中必须采用考虑了镜头畸变的成像模型。现有的线性标定方法无法考虑镜头畸变,而基于优化的方法实现起来一般比较复杂。本文在前人研究成果的基础上,引入了扩展的径向准直准则,通过合理组织参数求解次序,给出了一种考虑了镜头畸变的迭代式优化标定方法。该方法实现简单,每一步只需用到最小二乘法,可以有效提高标定精度,并可以方便地进行扩展以包含更多或更少的畸变参数而不增加计算的复杂度。立体匹配是立体视觉中最为关键的一个步骤。月球车所处的月面自然地形中没有规则景物,场景图像中无纹理和少纹理区域较多,为匹配带来了困难。本文针对月球表面地形相对平坦的特点,在对图对进行立体校正以后,引入了一个适用于平面地形的视差线性变化约束,并以此为基础给出了一种多级匹配算法。算法首先对特征明显的点进行匹配,其结果通过视差线性变化约束辅助后续点的匹配。在具体匹配中,利用最有力的灰度、x方向梯度和梯度方向实现多判据匹配。算法对于多种不同类型的地形图像表现出了良好的适应性,在精度和速度方面较传统方法都有提高。在摄像机标定和立体匹配之后,可以重建得到场景中景物点的三维坐标。为得到月球车障碍规避和路径规划所需要的三维地形图,对重建后的三维坐标点首先去除孤立点,然后在原始数据和网格数据中进行两轮插值,在保留障碍物后遮挡空白的情况下给出了比较完整的场景表面。视觉导航部分,实现了以特征检测和跟踪为基础的视觉测量导航算法,分析了影响其精度的主要因素。以月球车立体视觉系统提供的局部地形图和轨道器或着陆器提供的全局地形图为基础,给出了一种利用两地形图内局部高点位置关系不变性实现定位的简便算法。这两种算法使得立体视觉的三维地形图不仅可用于路径规划和障碍规避,还可用于视觉导航定位。更多还原

【Abstract】 According to China’s lunar exploration program, a lunar satellite will be launch to execute an elementary survey of the moon. In the second step an unmanned lunar spaceship will be landed on lunar surface and a lunar rover will carry out a detailed research in a rather large range. These explorations will lay the foundation for manned lunar exploration and constructing lunar base. Lunar rover is the platform of scientific instruments and its safe roaming is the basis for subsequent exploration. The vision and navigation system, as the theme of this paper, provides the terrain information surrounding the rover and the position information of the rover, which is indispensable for lunar rover’s safe roaming. This paper deals with these two systems.In lunar rover’s vision system a fairly large field of view is needed and thus a wide-angle lens is need in a CCD camera. As the same time the measurement precision in a large distance needs to be guaranteed for safety of the lunar rover. These two requirements constraint the camera calibration technique to adopt a camera model with distortion. In this respect existing linear methods do not accommodate distorted model and nonlinear optimization methods are usually difficult to realize and dependent on initial parameter values. This paper inherits previous research and presents an optimized calibration algorithm for distorted camera modal. The algorithm works in an iterative manner and is easy to implement as it avoids nonlinear optimization procedures. It can improve calibration precision effectively and can be extended easily to include more or less distortion term for different applications without increasing computation complexity.Stereo matching is a critical step in stereo vision. In the surroundings of lunar surface, few structured objects and many areas lacking in texture exist. This condition presents large difficulty to correlation-based stereo matching algorithm. In this paper image rectification is executed for stereo pairs in the first step. Then based on the fairly-planar property of lunar surface a disparity linearity constraint is introduced and a multi-stage matching algorithm is proposed. In the algorithm points with large intensity gradient are matched first and their matching results are used to assist in matching of other points based on the constraint. In a point’s matching procedure, a multi-criterion matching technque including intensity, x gradient and gradient orientation is utilized to improve matching accuracy. The algorithm exhibits nice adaptability to various types of natural terrain and demonstrates advantage over traditional correlation-based matching algorithm both in precision and in efficiency.In reconstruction section image points are reconstructed to obtain the 3D coordinates of corresponding object points. In order to acquire the digital elevation map needed for path planning and obstacle avoiding, the isolated reconstructed object points are removed at first. Then the elevation at grid points are computed by interpolation both in original object points and grid points. As a result a fairly complete scene surface is formed with occlusions behind large rocks reserved.As to visual navigation, this paper realized the visual odometry method which is based on feature detection and tracking and analyzed the main factors affecting navigation precision. Then we proposed a compact terrain matching algorithm to position the lunar rover based on the terrain from lunar rover and orbiter or lander. These two algorithms make further use of the reconstruction digital elevation map.更多还原