基于点的多尺度形态学重建滤波方法

Point-based multi-scale morphological reconstruction filter

针对现有机载激光雷达(LiDAR)点云滤波算法难以准确分离复杂地形中地面点与地物点问题,提出了一种基于点的多尺度形态学重建滤波方法 PMMF (Point-based Multi-scale Morphological reconstruction Filter)。在初始尺度层次下,PMMF通过构建一种基于点的形态学重建对原始点云滤波,即先在掩膜点云约束下借助k邻域结构元素和高程缓冲区反复膨胀标记点云,获取潜在地面点;然后通过自适应坡度方法剔除潜在地面点中的非地面点,其中,坡度阈值随地形复杂度自适应变化。在上层滤波结果基础上,PMMF通过提升种子点选择的网格尺度重复上层滤波过程,直至结果收敛。以国际摄影测量与遥感学会(ISPRS)发布的15组基准数据为研究对象,将PMMF滤波结果与近5年(2016年—2020年)提出的15种滤波算法比较表明,PMMF有8组数据滤波效果占优,15组数据平均总误差和Kappa系数分别为2.71%和91.08%。使用4种不同地形特征的高密度机载LiDAR点云数据进一步验证PMMF的滤波效果,并将计算结果与简单形态学滤波(SMRF)、布料模拟滤波(CSF)、渐进加密三角网滤波(PTD)和多分辨率层次滤波(MHF)比较。结果表明,PMMF滤波性能最优,平均总误差为3.24%,较其他4种滤波方法分别减小了12.0%、59.1%、70.1%和53.2%。

Many airborne LiDAR point cloud filters have been proposed over the past decades.However,these existing filters are incapable of producing satisfactory results in complex landscapes,such as rugged slopes covered with low vegetation and discontinuous terrain.Thus,a point-based multi-scale morphological reconstruction filter(PMMF) is presented in this work to overcome these problems.In contrast with the classical morphological filters,PMMF takes raw point cloud rather than rasterized grids as the basic processing element.First,the potential ground points are obtained by repeatedly dilating the marker point cloud with the k-neighbor structural element and adaptive elevation buffer under the limits of the mask point cloud.Thereafter,the non-ground points mixed in the potential ground points are eliminated by a terrain-adaptive slope filter.Based on the filtering results from the upper scale,PMMF increases the grid scale for selecting ground seeds on the next scale and repeats the filtering process as the upper level until the result converges.The three main contributions of the new algorithm include a point-based morphological method rather than a grid-based one to avoid information loss caused by point cloud rasterization,a multi-scale geodesic dilation with a slope-adaptive elevation buffer to select potential ground points and reduce the omission error on a steep terrain,and a terrain-adaptive slope filter to eliminate commission errors mixed in potential ground points.PMMF was employed to filter the benchmark samples provided by ISPRS,and its results were compared with 15 filtering algorithms proposed in the last 5 years(2016—2020).Results illustrate that PMMF outperforms the other filtering methods on eight out of the 15 samples,and its average total error and Kappa coefficient were 2.71% and 91.08%,respectively.Moreover,PMMF was used to process four high-density airborne LiDAR point clouds with different terrain features,and the filtering results were compared with Progressive Morphological Filter(PMF),Cloth Simulation Filter(CSF),Progressive TIN Densification(PTD),and multiresolution hierarchical filter(MHF).Results show that PMMF with an average total error of 3.24% has the best performance.The total error of PMMF is reduced by 12.0%,59.1%,70.1%,and 53.2% compared with those of PMF,CSF,PTD,and MHF,respectively.A large number of experimental results show that PMMF has achieved satisfactory filtering results on various terrains,and the filtering accuracy is significantly higher than those of other conventional filtering algorithms.Experimental verification shows that the three innovations proposed in this work contribute to the higher accuracy of the new algorithm and overcome the imperfection of the existing algorithms.