【作者】 陈卓；

【导师】 胡庆雷；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2015， 硕士

【摘要】 和地面机器人不同,微型无人机具有在三维空间中的机动能力,这使得它们能够进入人类很难到达的环境中完成许多危险的任务。对于监视、侦查、救援等任务,微型无人机必须能够不依赖预存的地图或者对环境的具体假设进行导航。本文为了解决微型无人机在未知环境中不依赖于GPS的导航问题,讨论了从单目视觉输入测量机体姿态和位置的方法,以及基于视觉和惯性传感器的卡尔曼滤波器的设计。并通过实验验证了算法设计的有效性。首先针对在采集图像序列时可能会引入的运动模糊、缺乏角点信息等问题,对特征各种特征检测算法进行了比较分析与试验。为了解决特征点集中分布在图像的一小块区域的问题,设计了分块特征检测的方案。然后针对从特征对应计算本质矩阵的问题研究了经典的五点算法,同时为了减小图像噪声等原因导致的误对应特征对本质矩阵计算的不利影响,设计了RANSAC算法与五点算法结合的方案用来剔除外点。考虑到五点算法对摄像机的姿态估计会随摄像机运动而漂移的特点,采取两点算法与航姿参考系统结合的方案来估计本质矩阵,并使用MATLAB测试了两点算法和RANSAC算法结合的位置测量效果。接下来针对从本质矩阵计算位姿矩阵的问题,研究了如何使用对应图像点三角化的方法来估计空间点位置,进而可以计算出位姿矩阵间的相对尺度,从而通过对经过尺度校正的相对位姿矩阵的逐次累乘最终得到摄像机位姿矩阵序列。最后针对上述位姿矩阵估计方法产生的漂移较大的问题,研究了局部光束平差法同时对空间点位置和摄像机位姿参数进行局部优化的方法,以期减小运动估计的漂移。最后为了改善单目视觉位姿测量的动态性能,同时为位置PD控制提供速度估计值,本论文引入了惯性测量元件,设计了视觉惯性组合导航算法。针对视觉惯性组合导航系统的设计问题,首先对比了两种过程模型的优劣,最后选择了直接IMU输入的过程模型,以此为基础设计了扩展卡尔曼滤波器,除了对机体相对于世界坐标系的位置、速度、姿态进行估计以外,还对惯性测量元件的零偏、视觉测程法引入的位置测量尺度因子、IMU坐标系与摄像机坐标系的相对位姿等参数进行估计。最后通过MATLAB仿真对滤波算法的有效性进行了验证。更多还原

【Abstract】 Unlike ground vehicles, micro aerial vehicles have the 3D maneuverability, which enable them to perform a lot of risky tasks in environments where humans cannot get access to. For tasks such as surveillance, rescue and reconnaissance, micro unmanned aerial vehicles(UAV) must be able to navigate without any prior knowledge of the scene map. In order to solve the navigation problem for micro UAV in unstructured environments where there is no GPS signal, we investigate a position and attitude measurement method from monocular visual input, as well as an Extended Kalman filter(EKF) based on visual and inertial measurement unit(IMU). Experiments are performed to demonstrate the validity of the algorithms.First, due to the possibility of motion blur and lack of corner in the input image, various feature detectors are analyzed and tested on every image from a challenging image sequence. Sometimes, feature might gather in a small region of an image. To solve this problem, we apply the feature detector to each block of an image. As for the computation of essential matrix from point correspondences, classic 5-point algorithm is used in combination with the RANSAC algorithm to eliminate the contamination of outliers.The next topic is the computation of the 5 DOF pose matrix from the essential matrix. Triangulation of point correspondences gives scaled 3D point coordinates. Coordinates of two consecutive triangulated 3D points can be used to calculate the relative scale factor of one pose matrix with respect to another, which is then used to rescale the latter pose matrix. Finally, concatenation of relative pose matrices results in a pose matrix sequence of the camera with respect to the world reference frame. In addition, local bundle adjustment was used to minimize the motion estimation drift.Finally, visual inertial navigation algorithm for a micro UAV is designed to improve the dynamic performance of the ego-motion measurement from an onboard camera. It additionally provides PD position control for the UAV with a velocity measurement. Pros and cons are listed for two process model candidates, based on which the process model with a direct IMU input is selected for the EKF to be designed. The EKF estimates such states as UAV body position, velocity attitude, IMU biases, scale factor induced by visual odometer. Finally, the algorithm was tested on a dataset with IMU and image information.更多还原