【作者】 冯国虎；

【导师】 吴文启；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2012， 博士

【摘要】 视觉/惯性组合导航以良好的互补性和自主性逐渐成为导航领域的一个新的研究热点和重要的发展方向。军民用领域的迫切需求使其凸显出重要的研究和应用价值，成为目前极具发展前景的导航技术。视觉/惯性组合导航系统具有强非线性，为了在载体线运动和角运动的条件下，能够高精度估计系统状态，需要深入、透彻地了解非线性系统的基本属性，特别是与状态估计性能密切相关的可观性特征。同时，现有视觉导航算法中大多数需要观测位置或尺寸已知的点特征或线特征，在未知环境下很难适用。因此，论文重点解决未知环境下单目视觉/惯性组合导航中两个问题，即单目视觉/惯性组合导航可观性分析和单目视觉/惯性组合导航动态滤波算法。主要研究工作内容如下：1）提出了基于线特征的纯单目视觉运动估计算法的改进方法。在基于线特征的Goddard方法的基础上，发现其公式推导错误，并将目标的加速度扩展为状态量，克服了Goddard方法只能在相对速度变化缓慢条件下应用的不足。2）分析了无惯性信息辅助条件下单目视觉导航的局限。对于高动态或长航程应用，基于点特征或线特征的滤波算法估计载体姿态的精度都不高，随着姿态误差的累积，长距离导航时系统的定位误差随距离呈非线性迅速增长。因此，在未知环境下，尤其是高动态或长航程应用时，纯单目视觉导航难以完成导航任务，需要引入其他传感器，如惯性系统。3）推导了基于高精度IMU的单目视觉/惯性组合导航系统的可观性条件。将误差方程线性化，采用EKF滤波估计误差状态。在摄像机高度已知的情况下，应用PWCS方法分别证明了基于点特征和线特征的组合导航姿态和速度误差的可观性。摄像机水平向前拍摄时，对于点特征，组合导航姿态和速度误差可观要求至少观测地面上三个不共线的点特征。对于线特征，组合导航姿态和速度误差可观要求至少观测地面上两条不平行的线特征。两者可观性条件等价。4）提出了基于高精度IMU的单目视觉/惯性组合导航滤波算法。对于点特征，单目视觉/惯性组合导航采用速度组合模式，利用惯性导航系统提供姿态信息，无需依靠视觉估计摄像机自身姿态。特征点不明显时，利用惯导系统短时间精度好的优势，确保导航的连续性。为提高视觉运动估计精度，通过组合导航实现摄像机标度因数误差的在线标定。与定期修正姿态的纯视觉算法相比，短距离的推车实验和长距离的跑车实验表明滤波算法定位精度更高，定位误差随时间增长率低。对于线特征，引入对偶四元数算子构造摄像机位置增量的表达式，将摄像机估计的速度与惯导系统测量的速度进行组合，建立滤波方程修正导航参数。实验表明滤波算法线特征匹配正确率高，能够连续多帧追踪。5）推导了基于低精度IMU和磁强计组合的单目视觉/惯性组合导航系统的可观性条件。针对低精度惯性系统，误差方程线性化会引入较大的非线性误差，采用矩阵滤波直接将姿态矩阵和速度作为状态进行估计，避免了非线性系统线性化带来的误差。以李导数理论为基础，构造可观性矩阵推导满足系统可观的特征观测集合，在摄像机高度已知且估计陀螺漂移和磁差的情况下，分析了可观性判据。对于点特征，组合导航姿态和速度可观要求至少观测地面上三个不共线的点特征，对载体运动的要求是：至少存在一个旋转自由度。对于线特征，组合导航姿态和速度可观要求至少观测地面上两条不平行的直线，对载体运动的要求是：至少存在一个旋转自由度。两者可观性条件等价。相对传统视觉导航的可观性结论，单目视觉/惯性组合导航的可观性条件减少特征数量的同时不再要求特征位置或尺寸已知，对未知环境下的组合导航系统设计和实现具有较强的实用价值。6）提出了基于低精度IMU和磁强计组合的单目视觉/惯性组合导航滤波算法。对于点特征和线特征，将姿态矩阵整体作为状态进行估计。相对四元数描述姿态，姿态矩阵的物理意义更为明显。相对其他非线性滤波算法，由于姿态矩阵的整体引入，原始非线性系统过程可以写成伪线性过程，即只含状态的一次项，不含二次项和高阶项，简化了滤波设计。更多还原

【Abstract】 Vision/SINS integrated navigation is an important topic in the integrated navigation,and has attracted considerable attentions from the civil and the military area. In order toestimate precise state of vision/SINS integrated navigation system, observability analysismust be done after filter design. Since vision/SINS integrated navigation system is a non-linear system, observability analysis is necessary when the system with linear and angulardynamics. Most of visual navigation algorithms need observations of known features. Itishardtosatisfyintheunknownenvironment. Motivatedbytheabove-mentionedinsuffi-ciencies, the main purpose of this dissertation is to address two important and challengingissues of monocular camera/SINS integrated navigation in the unknown environment:observability analysis and dynamic filter algorithms. The main contributions include thefollowing aspects.1. The estimate of position and orientation of moving object from monocular camerawas investigated. The EKF was designed according to Goddard method. Some mistakesin Goddard method were rectified. At the same time, relative translation and rotationsecondderivations of the object were added as state assignment. Comparedwith Goddardmethod, simulated data and actual experimental data show that the improved method canget better precision and robustness.2. The insufficiency of monocular vision navigation was analyzed. The improvedmethod still cannot estimate precise state when the system with high dynamics or overlong distance. It need combine with other sensors, such as SINS.3. For the monocular camera/high accuracy IMU integrated navigation system, EKFis used to estimate the error state. Under the condition of the camera positioning height isknown,observabilityoftheerrorsofattitudeandvelocityisprovedbyPWCSmethod. Forpoint feature, the observability of attitude and velocity errors need at least there featuresare detected on the ground. For line feature, the observability need at least two linearlyindependent features are detected on the ground.4. The EKF algorithm is proposed based on point and line feature, respectively. Forpoint feature, velocity integrated mode is used in the integrated navigation. With the helpof SINS, camera isn’t need to estimate orientation. SINS provide orientation with cameracontinuously. Camera’s velocity is solved after sensor alignment and time synchroniza-tion, and used to amend SINS’s velocity and position by Kalman filter. Integrated nav- igation calibrates camera scale factor on-line. Compared with visual navigation methodof periodic updates to the orientation estimate, the proposed method is more accurate andmore robustness, low rate of error growth. For line feature, dual quaternion was intro-duced to describe the position and attitude of camera. The proposed algorithm deducedthe increment formula between image feature and camera position. The velocity compu-tation difference between the SINS and the camera is chosen as observation of integratednavigation. A Kalman filter is used to correct the integrated navigation error includingthe camera scale factor. The experiment results show the proposed algorithm is accuratefor structural road.5. For the monocular camera/low accuracy IMU/magnetometer integrated naviga-tion system, the matrix Kalman filter(MKF) is used to estimate the state. Under the con-dition of the camera positioning height is known and gyro biases and magnetic variationsneed to be estimated, observability analysis is directly based on nonlinear system withLie derivatives. For point feature, the observability of attitude and velocity need at leastthere features are detected on the ground. At least one degree of rotational freedom isexcited. For line feature, the observability need at least two linearly independent featuresare detected on the ground. At least one degree of rotational freedom is excited. Theobservability conditions of attitude and velocity are equivalent for point and line feature.Compared with visual navigation, observability conditions of monocular camera/IMU in-tegrated navigation need less features. At the same time, the position of feature is no longneed to be known. There are valuable for the filter design and realization of the integratednavigation system.6. The MKF algorithm is proposed based on point and line feature, respectively. Itis designed such that the estimate of the state matrix is expressed in terms of the matrixparametersoftheoriginalplant. TheMKFpreservesthenaturalformulationoftheattitudematrixtorearrangetheoriginalnonlinearprocessmodelinapseudo-linearprocessmodel.更多还原