【作者】 章大勇；

【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， 2010， 博士

【摘要】 以惯性导航为核心的各种组合导航技术正得到广泛的研究与应用。近年来,基于外部环境测量的组合导航模式,由于具有自主导航能力强的特点,已经成为重要的研究方向。在各种外部环境测量手段中,激光雷达具备高精度重现三维环境的能力,测量信息丰富,不受复杂电磁环境影响,采用激光雷达和惯性系统构成的组合导航系统在军事及民用领域都有巨大的应用潜力和广阔的发展前景。本文研究了基于地标量测的激光雷达/惯性组合导航算法,对其中的关键技术和科学问题进行了深入分析,包括:地标特征提取中包含的一致性问题及其优化技术、激光雷达和惯性导航系统的量测一致性问题,组合导航最优估计的可观性问题和地标混合观测下的全局状态估计问题。论文的主要工作与创新点如下:(1)研究了基于地标量测的激光雷达/惯性组合导航算法的基本方法,包括不同特征地标的量测方程构造、空间非线性量测的线性化滤波器建模、点云扫描的归化问题,并分析了激光雷达/惯性组合导航中主要的误差源及其对系统的影响。分析认为,一致性与最优估计问题是影响激光雷达/惯性组合导航的关键因素。(2)研究了参数化点云特征提取及其参数空间一致性优化方法。首先研究了基于模型驱动的空间平面和圆柱面提取方法,深入分析了三维Hough变换的特点,指出传统三维Hough变换存在参数空间分割不一致问题。在此基础上,提出了基于对偶空间分割的三维Hough变换,在不改变计算量的情况下实现了参数空间的一致完备分割,对偶空间分割充分利用了传统Gauss分割的有效线性逼近,通过坐标旋转和重定义实现半球参数空间在测地线意义上的完备且一致分割,从而避免了多值性问题,保证了特征提取的一致性。实验表明,对偶空间分割能够唯一、有效检测半球面上任意方向的参数化地标。(3)研究了多尺度一致的点云局部特征提取算法。首先分析点云平滑算法,证明多尺度几何流点云平滑方法等价于三维网格上的高斯核方法,然后利用几何流平滑方法建立多尺度点云,并通过特征曲率检测方法提取不同尺度下的几何局部特征。然后在多尺度几何局部特基础上研究了相应特征描述符和匹配方法,最终实现了多尺度一致的点云局部特征提取与匹配。实验表明,算法能够应用于复杂环境下的导航地标检测和匹配,所提取特征在旋转、平移、尺度变化下能够保持一致性。同时,相比传统依赖网格的特征提取算法,本文算法的计算开销大为减小(4)研究了激光雷达/惯性系统的量测一致性问题,即二者的精确空间关系标定算法。首先研究了激光雷达和惯性系统的标定几何原理;针对控制点标定模型,建立了扫描线量测修正模型,得到基于对标定区域扫描激光强度图的改正量测;利用修正后的控制点量测构建了多矢量旋转参数估计方程,引入Wahba姿态确定算法,将旋转参数求解转变为四元数优化问题,得到了旋转参数的全局最优解,避免了小角度假设带来的理论缺陷;并以此为基础分析了参数解的最优化问题。最后利用实际激光雷达数据进行了验证,实验结果表明,在大角度标定条件下,相比传统平差方法,算法能够提高6.71%的精度,并且在一定范围内随着噪声增加,优化算法的精度没有明显降低。(5)从理论上分析了激光雷达/惯性组合导航的可观性问题。利用非线性全局可观性分析方法,针对不同类型地标,从全局可观的角度推导出满足系统可观的地标观测集合,将传统LOS量测导航的可观性结论从局部可观推广到全局可观,放松了对地标观测的要求。数值仿真和实际实验说明本文得到的全局可观性结论可靠、有效,全局可观性分析方法对组合导航系统设计具有重要的理论价值。组合导航估计的结果表明,激光雷达/惯性组合导航系统的水平定位精度优于5m,姿态角精度优于0.1~o,能够满足实际使用需求。(6)深入研究了激光雷达/惯性组合导航系统中的全局状态估计算法。研究了基于绝对/相对地标混合观测的组合导航算法,在其中利用状态扩展实现导航参数和地标位置的全局状态估计;在此基础上,详尽分析了绝对/相对地标混合观测条件下相对地标的收敛性,指出单纯的相对地标估计为弱收敛,即多个地标之间的相对位置收敛,而具备绝对地标参与的混合观测中,地标能够强收敛到绝对位置。进一步,分析了混合观测条件下导航参数的收敛性,证明系统导航参数精度在混合观测条件下最优。数值仿真表明,混合观测组合导航具有明显优势,全局状态估计算法在全程精度和最终精度上都优于单纯观测绝对地标和单纯观测相对地标的传统算法。更多还原

【Abstract】 Reasearchers have developed various of integrated navigation systems for large number of applications, which mainly designed with Inertial Navigation System as principal part. But systems using satelliates to aid INS suffer the problem of signal jamming and blocking. So independent integrated system, which sensing the environment to aid INS, are regarded as one potential solution. Compare to many kinds of senors, LIDAR(Light Detection and Ranging) is able to mapping 3D environment directly and accurately. So it is possible to integrate LIDAR and INS for precise navigation on military or civil area.In this dissertation, some key technologies of LIDAR/INS integrated system are offered, including consistency problem for 3D feateure detection, measurement consistency problem between LIDAR and INS system, global observability analysis and global steates estimation for integrated system. These four problems are tight parts for a whole. The main contributions include the following aspects:(1) The basic frame of integrated navigation is discussied, including measurements construction with three type landmarks, linearization model for filter, 3D points cloud transtition into key frame. Also the error source of integrated system is discussied. The analysis result indicates it is feasible to build the LIDAR aiding INS integetrated system.(2) Inorder to dectction regular landmarks such as patch or cylinder, the 3D Hough Transform is deeply analysed. For 3D case, the Hough Transform must divide the parameters on the spherical surface, but the Gaussian segmentation is only average on parameter space but not on sphere geodesic, which may cause singularity for patch’s or cylinder’s normal which point to high latitude area. If the singularity phenomena were appeared, the landmark detection is unconsistency. Then the dual segementation of sphere parameter is proposed to avoid the singularity phenomena, which divide the sphere into two approximate Gussian coordinates and define coordinates with different area and rotations. Based on dual segementation, the dual 3D Hough Transform can get complete and consistent segmentation for half sphere, then the landmark detection is nearly consistent for all area of sphere. The experiments show that dual 3D Hough Transform can dectect regular landmarks with arbitrary normal.(3) The dissertation discussed local interest point detector in points cloud with multi-scale consistency. The equivalence between Gaussian kernel on grid and the unorganized multi-scale geometry flow smoothing method is established. Based on multi-scale geometry flow smoothing, the interest point is detected with characteristic curvature, which is invariant to scale, rotation and transition. Also the descriptor of local curvature is proposed for points cloud matching. The experiment show that local interest point detector performs well in complicated environment.(4) The measurements consistency between LIDAR and INS is discussed, which also called calibration problem of integrated system. Firstly, the measurement geometry is proposed with control points and control area. Then the scan line correction algorithm is discussed for laser intensity image, which revise the raw control points to ground truth points. In order to slove the rotation matrix, the Wahba attitude determination alograthim is imported to estimate the matix with SO(3) restrict, which gets a global optimized result. The experiments using real LIDAR is implemented and the result shows that in large biases case, the algorithm can improve 6.71% accuracy than least squre method. Also with noise increase, the estimation is robust.(5) The newly theoretical analysis of nonlinear global observability of LIDAR/INS integrated system is propsed. Considering integrated system respectively with high precision INS and low precision INS, the dissertation gives the global observability conditions respectively for three types of landmarks. The result of observability extends the covariance anlyisis result of mutli-LOS(line of sight) observation and be more comprehensible than conventional conclusion. The observability analysis is straightforward and intuitionistic, which provide a theoretic fundation in designing the filter for integrated navigation. The simulation and experiments result shows that positioning accuracy of integrated navigation system is better than 5m, the attitude accuracy is better than 0.1o.(6) The hybrid estimation algorithm is propsed, which using global extended states to estimate both navigation states and landmark location. In conventional SLAM method, the location of landmarks can be estimated for revising the vehicle’s postion. For that LIDAR is a popular senor in SLAM, it is worthful to import this mehod into LIDAR/INS integrated navigation. The convergence of landmark in integrated navigation is discussed. The result shows that convergence in SLAM is weakly but this weakly convergence can transfer to strong convergence when there be observations of absolute landmarks. Furthermore, the strong convergence of landmarks can lead the location of vehicle to strong convergence, which improves the total location accurary for full navigation process. The simulation result shows the hybrid algorithm performs better than pure SLAM and pure integrated navigation mehod. Key Words:Integrated Navigation; LIDAR(Light Detection Radar); Strapdown Inertial Navigation System; System Consistency; 3D Interest Points Detection; Attitude Determination; Nonlinear Global Observability; Global States Extend; Estimation Convergence更多还原