【作者】 刘勇；

【导师】 李德仁； 李清泉；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2012， 博士

【摘要】 运动平台的自定位测姿是很多应用领域的一项重要任务。本文研究的基于影像的定位测姿是一种新的技术。目的是基于三维空间的位置和姿态估计,且输出的位置和姿态结果能用作其它传感器进行数据拼接和融合的相对空间基准。针对此目的,本文具体的研究工作如下：1)详细分析了单目视觉中,使用两幅和三幅影像时的相对位姿获取原理、方法,和计算过程。使用的双视和三视几何法是利用单目影像序列进行定位测姿的基础。对于双目立体相机影像,先得到三维空间点云,然后利用点云中匹配的三维点集来求解相对位置和姿态变化。2)详细讨论了如何对序列的单目和双目影像进行位姿的确定、运动轨迹的重建和场景中所提取特征点的三维重建。利用所建立的双视和三视几何关系,使用序列中的相邻两幅或三幅影像,可以获取整个序列中每幅影像的位置和姿态。具体如下：a)基于序列中相邻两幅影像的定位测姿方法。该方法能得到相对于参考坐标系的姿态,但相机的移动则缺乏全局统一的比例。b)基于序列中相隔一幅影像的定位测姿方法。该方法利用两幅影像作为当前组和上一组影像的约束关系,可以把位置的变化比例传递到整个影像序列中。c)基于序列中相邻三幅影像的定位测姿方法。该方法通过三个位移向量构成一个封闭三角形来形成约束,以恢复全局统一的位置变化比例。d)基于序列中相邻三幅影像的三焦张量法。该方法求解三幅影像间的三焦张量,并由此恢复各个相机矩阵。利用所得的相机矩阵,再恢复出相机的位置和姿态变化。3)使用单目影像和双目影像序列相结合的定位测姿方法。该方法的特点在于可以克服单目定位测姿的缺乏真实移动距离的缺点,以及降低定位测姿过程中的误差积累。也可以克服双目影像序列间隔大、采集频率低造成的时间分辨率不高的缺点。4)文中论述了获取影像序列中同名点的方法。对于单目视频序列影像,采用特征点跟踪的方法获得求解过程中所需要的影像间同名点,并通过关键帧的选择来确保计算结果的稳定和可靠。对于双目立体影像序列,则采用特征点匹配方法获得同名点,并通过特征点的匹配获得了点云中同名点的匹配。5)文中使用了计算机模拟数据和真实的车载影像数据进行了方法的验证。模拟数据能提供真值进行结果的对比。真实的车载影像数据能验证该方法的实际可行性、有效性和稳定性,并探讨其适用范围。本文研究的主要创新在于：1)提出完全基于影像传感器的运动平台三维定位测姿方法。与已有的基于影像传感器的定位测姿方法相比：①不依赖于其它的测距或定位测姿传感器,是完全基于影像数据的；②定位测姿是三维空间上的,而非平面上的坐标和方位。2)研究了几种基于单目影像序列进行定位测姿的算法,以及一种利用双目影像序列来进行定位测姿算法。研究了它们对影像中同名特征的要求,以及对运动的适应性。3)设计了一种基于多目影像序列的定位测姿方案。该方案适用于运动平台的任意运动模式,且利用单目影像的高频采集特点来提高定位测姿的敏感度；利用双目影像的定位测姿精度高提高精度和定位测姿的稳定性。更多还原

【Abstract】 Self-position and orientation of moving platform play a vital role in deep space exploration, indoor robot control and other application. Image-based self-position and orientation technology for moving platform is an arisen technology. It is a widely used for3D position and orientation estimation, and it provides the relative space benchmark for registration and fusion of other sensor data. Adopting such technology, the thesis reports the study in detail as follows:1) The relative positioning and orientation methods based on two or three-views have been studied thoroughly. This two or three-view geometry is the foundation of self-position and orientation from monocular image sequence. Besides, for stereo images,3D point clouds and relative position and orientation change estimation method based on these points are also discussed.2) The relative positioning and orientation for monocular image sequences and stereo image sequences have been detailed studied. For relative positioning and orientation, there are four methods as follows.a) The method based on two-view geometry with consecutive two images. In this situation, relative orientation change can be obtained with real angle relative to the last image. But the relative position change can be obtained without global consistent scale.b) The method based on two-view geometry with every other two images. In this situation, two-view geometry is calculated in every other two images. The image between the two images can be used to transfer the scale to the next triple images. So camera moving trajectory with a global consistent scale can be obtained.c) The method based on two-view geometry with consecutive three images. In this situation, two-view geometry is calculated one another in this triple images. The three displacement vectors can form a close vector triangle. With this constrain, moving trajectory with a global consistent scale can be recovered.d) The method based on trifocal tensor with consecutive three images. When the trifocal tensor has been calculated, camera projection matrixes of the three images are also obtained. With these matrixes, camera’s relative position and orientation are recovered consecutively.3) The relative positioning and orientation method based on monocular and stereo image sequences. On the one hand, using stereo image sequence, the real displacement in monocular image sequence can be obtained, and the accumulation error can be reduced. On the other hand, using monocular image sequence, the temporal sampling rate of stereo image sequence can be increased.4) The methods of corresponding point detection in image sequence have been introduced. For single camera video stream, feature point tracking method has been applied, and key frame selection step has been used for increase the stability and reliability of two-view geometry calculation. For both single camera image sequence and stereo camera image sequence, feature point detection and matching method have been applied. Besides, the corresponding point between two point clouds has been found through feature point matching between two stereo image pairs.5) Computer simulation image sequences and real image sequences have been applied to test the proposed relative positioning and orientation methods. Simulation test can provide the ground truth for comparison with the recovered moving trajectory, and real image sequences test can evaluate the method’s feasibility, efficiency and stability when it has been applied in mobile mapping system.The innovation of this search is as follows.1) Only image based relative positioning and orientation method has been proposed. Compared to those existing methods with image, this search has two characteristics:①It’s fully image-based method and independent on other range finding sensor, or positioning and orientation sensor;②the relative positioning and orientation is in3D space, not localization and heading only in plane.2) Four relative positioning and orientation methods based on monocular image sequence have been proposed and the comparison with one another has been made. Besides, their demands on corresponding points and movement have been studied. The stereo image has been used to calculate the relative position and orientation, not only reconstruct the3D scene structure. The corresponding point between two point clouds has been found through feature point matching between two stereo image pairs.3) The method which combines monocular image sequence based method with stereo image sequence based method has been proposed. This hybrid method can take the advantages of both high temporal sampling rates from single video camera and real displacement from stereo camera. It can produce high precision, high stability and high reliability results of relative position and orientation.更多还原