融合波形和点云的机载成像激光雷达数据自动滤波与高精度三维地形信息提取研究

The project studies the theory and method of automatic filtering airborne laser scanner data and extracting high accuracy terrain models in low vegetation covered and hilly areas based on full-waveform data and point clouds. Since most LiDAR systems generate point clouds directly, it is impossible to filter them automatically without any manual intervention and the terrain points cannot be extracted when the height of vegetation covered is lower than the laser echo separation, which limits the accuracy of the final Digital Elevation Model (DEM). Likewise, in some hilly areas, the foot-print looks like an ellipse when a beam hits an incline plane. Especially if the inclination angle of the incline plane is large enough, the beam will be quantized as two echoes, which also introduces errors. To filter out lower vegetation and to increase DEM accuracy in hilly areas remains difficult in airborne LiDAR data processing, and therefore limits the application of the technology to these areas. This study aims to overcome the mentioned difficulties based on the combination of full-wave form data and discrete point clouds. First, point clouds are used to filter out high vegetation and man-made buildings that the heights are larger than the length of echo separation by using matured algorithms such as progressive TIN densification, then variational Bayesian method is utilized for automatic decomposition of full-waveform to derive parameters which could be used to extract lower vegetation. In according to the theory proposed by Wagner, a full waveform is superposed by several single Gaussian-like waves and therefore can be decomposed into several Gaussian functions. Conventional Expectation Maximum (EM) algorithms are usually adopted to carry out the decomposition process, which suffers from inefficiency or even fails to convergence since it requires relatively accurate initial parameters as inputs to the algorithm. Meanwhile, the number of components, which determines the number of discrete echoes that could be distinguished from the full waveform data, remains manually set in EM algorithm. Variational Bayesian is one of the newly developed machine learning algorithm which could be the candidate to overcome the drawbacks caused by conventional EM. It first deals with the full waveform as a single Gaussian function to fit the outline of the waveform, then superposing other Gaussians with different parameters to the first one in a recursive manner that the superposition of these multi-Gaussian functions fits the original waveform more and more accurately. The recursive process is controlled by splitting test. After the waveforms are accurately decomposed, weak echoes and neighboring echoes whose echo separations are too small to be quantized by a conventional discrete LiDAR system, could be detected and recorded, hence low vegetation or other objects whose height is less than the distance resolution of a LiDAR could be extracted. Filtering of low vegetation is implemented by first form feature vectors which is formed by the parameters extracted from various waveforms, and then analysis the feature vectors to classify them into two categories: vegetation and ground, so that the points from vegetation can be labeled as “vegetation” , thereby automates the filtering procedure. Furthermore, since the width of a single Gaussian component can be determined accurately by the algorithm, which relates directly the slope gradient of an incline plane, the slope gradient can be estimated as well. We will study the quantization methods in such a scenario, hopefully improve the final accuracy of the point clouds. Full waveform data acquired by Riegl LM 5600 will be used as experimental dataset in the study, which covers a hilly area with medium height to low vegetation.

本课题基于机载激光雷达（LiDAR）全波形（full waveform）和点云数据，研究在低矮植被覆盖区和地面起伏较大地区，自动滤波与高精度三维地形信息提取的理论与方法。工程应用中，目前都是使用由系统直接生成的点云数据，点云数据滤波需要人为干预，没有实现自动化，而且这种数据无法区分高度小于系统距离向脉冲分辨率的低矮植被，影响了数字高程模型的生成精度；同样，在地面起伏较大的地区，激光点相当于撞击在一个斜面上，从而使得地面光斑是一个长轴沿斜面方向的椭圆,如果斜面倾角比较大，就可能形成两次回波，同样引起结果数据的误差。本课题融合波形和点云数据，根据波形数据的高斯属性，运用改进的变分贝叶斯方法对波形进行全自动分解，提取相关参数；利用点云数据初步滤除主要地物类型的基础上，分析辐射参数，剔除低矮植被；建立回波波形参数与起伏地形地面倾角的关系，解决困难区域自动化滤波与高精度三维地形信息提取问题。

由机载激光雷达硬件系统提供的点云数据无法识别并剔除低矮植被，且精度受地形起伏影响较大。基于此，本课题研究低矮植被剔除及由倾斜地表所引起的误差改正的理论与方法。主要成果如下：.1、引入变分Bayes思想，提出利用F检验、循环叠加高斯函数的方法，自动确定波形分解过程中的高斯波形个数；提出分组LM算法，解决传统LM波形分解算法中Jacobian矩阵可能出现非数值元素的问题。实验表明，分组LM分解算法得到的地面点比传统分解方法得到的地面点多22%，且高程精度提高超过25%；.2、在低矮植被剔除研究中，首先利用渐进TIN加密滤波算法，获得初始滤波结果。然后根据波形分解得到的高斯波形参数，计算后向散射系数，得到低矮植被提取判据；由此得到的精化DEM与直接由系统点云生成的DEM相比，精度提高了30％以上；.3、研究了倾斜地表引起的回波波形展宽和地面倾角的定量关系，推导了波形展宽和地面倾角、飞行高度、激光入射角和光斑发散角之间的函数模型，并采用实际飞行数据，对该模型进行了实验验证。实验表明，在地面倾角小于5°时，由倾斜地表引起的误差可以忽略（经改正后，高程精度仅提高1.5%），而倾角大于10°时，该误差需要考虑改正（经改正后，高程精度提高7.2%）；.4、针对已有TIN迭代加密滤波算法的缺陷，提出一种顾及断裂线的点云滤波方法。该方法首先利用波形数据生成点云，而非直接使用系统提供的点云。经两块测试区实验结果表明相比较于已有的TIN迭代加密滤波算法，改进后的TIN迭代加密滤波算法总误差分别由8.91%、9.07%降低到6.3%、5.87%；同时利用变差函数提取地形因子对地形进行分类，在分类后的地形基础上确定迭代角，进行TIN迭代加密滤波，提高滤波算法的自适应性；.5、此外，本课题还研究了GPU加速的波形分解算法，取得较好的实验效果。.本课题融合全波形与点云数据，解决了低矮植被覆盖区域以及地形起伏较大区域三维地形信息精确提取问题，不仅具有重要理论意义，同时具备较大的应用价值。