一种地形自适应的机载LiDAR点云正则化TPS滤波方法

随着机载激光雷达成像技术(LiDAR)的不断发展,激光点云数据处理的相关研究也在不断深入。点云滤波是机载激光雷达点云数据处理的重要环节之一。针对多数经典滤波方法在复杂地形和地物条件下的滤波效果不够理想的现状,提出一种新的基于相对变异系数的地形自适应正则化薄板样条插值点云滤波方法。采用二维区域增长获取初始插值参考点后,基于线特征约束对参考点进行优化,去除部分低可靠性参考点以得到较准确、分布离散均匀的初始插值参考点集合,在此基础上通过正则化薄板样条插值方式来拟合地形点与地物点之间的滤波分类面,完成对机载激光点云的高精度自适应滤波。对比实验结果表明,本文的地形自适应滤波方法在2组实验数据的总体错误率分别达到4.14%和4.17%,在错误率和多地形综合表现等方面具有优势,且滤波运算效率在目前主流的滤波算法中处于较高水平。另外,实验结果验证了地形自适应滤波方法在斜坡、山脊等起伏较多的复杂地形与包含植被和建筑物的混合地形等处的点云滤波结果具有较好的准确性。

A Terrain-adaptive Airborne LiDAR Point Cloud Filtering Method Using Regularized TPS

With the continuous development of LiDAR technology, the research of LIDAR point cloud data processing is also in-depth. Point cloud filtering is one of the key steps in airborne LiDAR point cloud processing. The existing point cloud filtering algorithms often work well on some specific terrains, however, their filtering results are not satisfying in the cases of undulating terrains or mixed terrains, some post-processing measures are always needed. Based on relative coefficient of variation and regularized thin-plate-spline interpolation, a new terrain adaptive point cloud filtering method is proposed in this paper. The initial seed points are obtained by two-dimensional and 8-directional region-growing method, and then optimized by extracting line features from the point clouds, the points with low reliability are removed from the sets of reference points. After that the reference points are mostly reliable and scattered in the whole test area, and could be used to generate classifying surface. Finally, the classifying surface between ground points and non-ground points is fitted using thin-plate-spline interpolation. Classifying surface is used to absorb more ground points from point cloud, which could provide reference information for the next round of interpolation. In this process we use regularization item of adaptive coefficient to control the bending extent of classifying surface, in order to make the filtering algorithm adaptive to different types of terrains. Ground points are totally filtered after several iterations. The experimental results on point clouds from multiple devices show that the total errors of our proposed method were 4.14% and 4.17% in Guangzhou and ISPRS datasets, respectively. The result of the proposed filtering method is not the best, but it is more stable and has better terrain adaptability compared to state-of-the-art popular algorithms such as progressive TIN filter, cloth simulation filter, semi-global filter, etc. The proposed method outperforms other comparison methods in both error rate and overall performance in several complex or special terrains, as well as high computational efficiency. Additionally, the promising experimental results demonstrate that the proposed adaptive terrain filtering method is an accurate and efficient solution for airborne LiDAR point cloud filtering in complex terrains, such as slopes, ridges, and mixed terrains including vegetation and buildings.