一种灰度体素结构分割模型下的机载LiDAR 3D滤波算法

doi:10.12082/dqxxkx.2020.190326

摘要/Abstract

摘要：

针对现有的基于机载LiDAR数据的滤波算法未能充分利用数据提供的所有信息及其所采用的数据结构表达复杂、存在信息损失等缺陷,提出了一种灰度体素结构分割模型下的机载LiDAR 3D滤波算法。算法首先以综合利用机载LiDAR数据的高程及强度信息为目的将点云数据规则化为灰度（体素内激光点的平均强度的离散化表示）体素结构,然后基于各体素间的空间连通性和灰度相似性准则,将灰度体素结构分割并标记为若干个3D连通区域,最后依据地面与其它目标的高差特性提取与其对应的3D连通区域。算法优势在于:基于体素结构设计,为3D滤波算法;综合利用了地面目标的几何及辐射特征,对比传统滤波算法可应用于更复杂的场景;滤波结果为3D地面体素形式,可直接用于创建地面3D模型。实验采用国际摄影测量与遥感协会（International Society for Photogrammetry and Remote Sensing, ISPRS）提供的不同密度的机载LiDAR基准测试数据测试了邻域尺度参数的敏感性及提出的算法的有效性,并和其他经典滤波算法做对比。定量评价的结果表明,51邻域为最佳空间邻域尺度;点云密度为0.67点/m²的数据集1的滤波平均完整率、正确率及质量分别为0.9611、0.9248及0.8934;点云密度为4点/m²的数据集2的滤波平均完整率、正确率及质量分别为0.8490、0.8531及0.7404;对比其全经典滤波算法本文算法在高密度点云数据滤波时表现更佳。

关键词: 激光雷达, 滤波, 体素结构, 3D连通区域, 几何特征, 辐射特征, 分割, 高斯混合模型

Abstract:

Existing filtering algorithms based on airborne LiDAR data have many disadvantages, e.g., they fail to make full use of all the information of the LiDAR data, and their data structures are complex or suffer from information loss. In this paper, we proposed an airborne LiDAR 3D filtering algorithm based on the grayscale voxel structure segmentation model. The proposed algorithm ?rst regularizes a given LiDAR point cloud into a grayscale voxel structure to comprehensively utilize the elevation and intensity information of the LiDAR data, in which the voxel's grayscale denotes the discretized mean intensity of LiDAR points within the voxel. Then, the constructed grayscale voxel structure is segmented into multiple 3D connected regions depending on the spatial connectivity and grayscale similarity among voxels. Finally, the 3D connected regions corresponding to ground objects are detected based on elevation differences. The proposed algorithm has the following advantages. (1) It is designed based on the grayscale voxel structure and is a 3D filtering algorithm. (2) The average completeness, correctness and quality of the proposed filtering algorithm for dataset 1 with 0.67 points per square meter were 0.9611, 0.9248, and 0.8934. The average completeness, correctness and quality of the proposed filtering algorithm for dataset 2 with 4 points per square meter were 0.8490, 0.8531, and 0.7404. (3) Its filtering result is in the form of the 3D ground voxel, which can be directly used for creating a 3D model of ground. The sensitivity of "spatial adjacency size" parameter in the model and the validity of the proposed algorithm is analyzed by using two airborne LiDAR benchmark test datasets of different densities provided by the International Society for Photogrammetry and Remote Sensing (ISPRS), and the accuracy of the proposed filtering algorithm is compared with that of the other classical filtering algorithms. Our experimental results of quantitative evaluation show that: (1) The 56-adjacency was the optimal adjacent size. (2) The average completeness, correctness and quality of the proposed filtering algorithm for dataset 1/2 with 0.67/4 points per square meter were 0.9611/0.8490, 0.9248/0.8531, and 0.8934/0.7404, respectively. (3) Compared with other classical filtering algorithms, the proposed algorithm performed better in high-density point cloud data filtering.

Key words: LiDAR, filtering, voxel structure, 3D connected region, geometric characteristic, radiometric characteristic, segmentation, gaussian mixture model

王丽英, 赵元丁. 一种灰度体素结构分割模型下的机载LiDAR 3D滤波算法[J]. 地球信息科学学报, 2020, 22(11): 2118-2127.DOI:10.12082/dqxxkx.2020.190326

WANG Liying, ZHAO Yuanding. An Airborne LiDAR 3D Filtering Algorithm based on the Grayscale Voxel Structure Segmentation Model[J]. Journal of Geo-information Science, 2020, 22(11): 2118-2127.DOI:10.12082/dqxxkx.2020.190326

图/表 13

图1

图2

图3

图4

图5

表1

表2

图6

图7

表3

图8

表4

表5

参考文献 20

[1]	Wack R, Wimmer A. Digital terrain models from airborne laser scanner data-a grid based approach[J]. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2002,34(3B):293-296.
[2]	Gomroki M, Jafari M, Sadeghian S, et al. Application of intelligent interpolation methods for DTM generation of forest areas based on LiDAR data[J]. Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 2017,85(4):227-241.
[3]	Jiang J J, Zhang Z X, Ming Y. Filtering of airborne lidar point clouds for complex cityscapes[J]. Geo-spatial Information Science, 2008,11(1):21-25. doi: 10.1007/s11806-007-0157-2
[4]	Wang L Y, Xu Y, Li Y. Aerial LIDAR point cloud voxelization with its 3D ground filtering application[J]. Photogrammetric Engineering and Remote Sensing, 2017,83(2):95-107. doi: 10.14358/PERS.83.2.95
[5]	刘凯斯, 王彦兵, 宫辉力, 等. 机载LiDAR点云数据的二面角滤波算法[J]. 地球信息科学学报, 2018,20(4):414-421. doi: 10.12082/dqxxkx.2018.170489
	[ Liu K S, Wang Y B, Gong H L, et al. Dihedral angle filtering algorithm for airborne LiDAR point cloud data[J]. Journal of Geo-information Science, 2018,20(4):414-421. ]
[6]	王丽英, 王圣, 徐艳, 等. 结合体元数据结构的机载LIDAR建筑物检测[J]. 中国图象图形学报, 2017,22(10):1436-1446.
	[ Wang L Y, Wang S, Xu Y, et al. Airborne LIDAR building detection based on voxel data structure[J]. Journal of Image and Graphics, 2017,22(10):1436-1446. ]
[7]	张杰. 机载LiDAR点云滤波及特征提取技术研究[D]. 徐州:中国矿业大学, 2017.
	[ Zhang J. Study on the technology of feature extraction and filtering for airborne LiDAR points cloud[D]. Xu Zhou: China University of Mining and Technology, 2017. ]
[8]	孙美玲. 机载LiDAR数据滤波及城区汽车目标检测方法研究[D]. 成都:西南交通大学, 2014.
	[ Sun M L. Research on airborne lidar data filtering and urban car object detection method[D]. Chengdu: Southwest Jiaotong University, 2014. ]
[9]	孙蒙, 顾和和. 基于微分形态学断面的机载LiDAR数据滤波新方法[J]. 大地测量与地球动力学, 2016,36(7):591-594,599.
	[ Sun M, Gu H H. A new method for airborne LiDAR data based on differential morphological profiles[J]. Journal of Geodesy and Geodynamics, 2016,36(7):591-594,599. ]
[10]	Axelsson P. DEM generation from laser scanner data using adaptive TIN models[J]. International Archives of Photogrammetry and Remote Sensing, 2000,33(B4):111-118.
[11]	Sithole G. Segmentation and classification of airborne laser scanner data[D]. TU Delft: Delft University of Technology, 2005.
[12]	周晓明. 机载激光雷达点云数据滤波算法的研究与应用[D]. 郑州:解放军信息工程大学, 2011.
	[ Zhou X M. Research and application of airborne LIDAR point cloud data filters[D]. Zhengzhou: PLA information engineering university, 2011. ]
[13]	赵泉华, 石雪, 王玉, 等. 可变类空间约束高斯混合模型遥感图像分割[J]. 通信学报, 2017,38(2):34-43.
	[ Zhao Q H, Shi X, Wang Y, et al. Remote sensing image segmentation based on spatially constrained gaussian mixture model with unknown class number[J]. Journal on Communications, 2017,38(2):34-43. ]
[14]	ISPRS Commission III, the Working Group III/3. ISPRS test on extracting DEMs from point clouds [DB/OL]. https://www.itc.nl/isprs/wgIII-3/filtertest/downloadsites/, 2003 -08-05.
[15]	Cramer M. The DGPF test on digital aerial camera evaluation - overview and test design[J]. Photogrammetrie - Fernerkundung - Geoinformation, 2010,2:73-82.
[16]	Rutzinger M, Rottensteriner F, Preifer N. A comparison of evaluation techniques for building extraction from airborne laser scanning[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2009,2(1):11-20. doi: 10.1109/JSTARS.2009.2012488
[17]	Chen C, Li Y, Li W, et al. A multiresolution hierarchical classification algorithm for filtering airborne LIDAR data[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2013,82:1-9.
[18]	Shao Y C, Chen L C. Automated searching of ground points from airborne LIDAR data using a climbing and sliding method[J]. Photogrammetric Engineering and Remote Sensing, 2008,74(5):625-635.
[19]	Pingle T J, Clarke K C, Mcbride W A. An improved simple morphological filter for the terrain classification of airborne LIDAR data[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2013,77(1):21-30.
[20]	Mougus D, Žalik B. Parameter-free ground filtering of LIDAR data for automatic DTM generation[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2012,67(1):1-12.

样本	点数	异常点数	空间范围/m	体素分辨率/m	3D格网尺寸	非0值体素数
Samp31	28 862	83	174.2×161.9×117.0	1.0×1.0×0.3	176×164×123	22 129
Area3	237 875	2	151.9×236.3×24.5	0.4×0.4×0.1	382×593×243	168 412

样本	II_e					I_e					T_e
样本	6 邻域	18 邻域	26 邻域	51 邻域	56 邻域	6 邻域	18 邻域	26 邻域	51 邻域	56 邻域	6 邻域	18 邻域	26 邻域	51 邻域	56 邻域
Samp11	0.1901	0.1839	0.1982	0.3085	0.3652	0.1477	0.1434	0.1309	0.0967	0.0971	0.1658	0.1606	0.1596	0.1871	0.2115
Samp12	0.0575	0.0545	0.0721	0.0654	0.0858	0.093	0.0913	0.0558	0.0765	0.0805	0.0757	0.0733	0.0638	0.0711	0.0831
Samp21	0.0511	0.0532	0.0588	0.0675	0.0748	0.0056	0.004	0.0023	0.0020	0.0021	0.0157	0.0149	0.0148	0.0165	0.0182
Samp22	0.3041	0.1711	0.2025	0.2402	0.2908	0.0485	0.0644	0.0411	0.0583	0.0296	0.1282	0.0977	0.0915	0.1150	0.1110
Samp23	0.1270	0.1515	0.1492	0.1379	0.1506	0.1392	0.0485	0.0390	0.0222	0.0268	0.1335	0.0972	0.0911	0.0770	0.0854
Samp24	0.1215	0.1220	0.1283	0.1599	0.2265	0.0898	0.0709	0.0635	0.0480	0.0563	0.0985	0.0849	0.0813	0.0788	0.1031
Samp31	0.0821	0.0532	0.0504	0.039	0.0515	0.0161	0.0260	0.0255	0.0186	0.0252	0.0465	0.0385	0.0370	0.0280	0.0373
Samp41	0.0937	0.0875	0.0877	0.0445	0.0576	0.0616	0.0491	0.0491	0.0029	0.0143	0.0777	0.0683	0.0684	0.0237	0.0360
Samp42	0.0736	0.0637	0.0511	0.0169	0.0407	0.0760	0.0284	0.0241	0.0252	0.0178	0.0743	0.0533	0.0432	0.0193	0.0340
Area 2	0.0830	0.1056	0.1055	0.0578	0.0915	0.1577	0.1531	0.1414	0.1604	0.1466	0.1074	0.1215	0.1176	0.0901	0.1096
Area 3	0.0549	0.0440	0.0500	0.0561	0.0634	0.2289	0.2063	0.1966	0.1416	0.1391	0.0983	0.0835	0.0862	0.0775	0.0827

样本	R_com	R_cor	R_q	Kappa
Samp11	0.9033	0.7973	0.7346	0.6087
Samp12	0.9235	0.9368	0.8693	0.8578
Samp21	0.9980	0.9811	0.9792	0.9511
Samp22	0.9417	0.8964	0.8493	0.7236
Samp23	0.9778	0.8876	0.8701	0.8447
Samp24	0.9520	0.9402	0.8976	0.8003
Samp31	0.9814	0.9672	0.9498	0.9436
Samp41	0.9971	0.9572	0.9545	0.9526
Samp42	0.9748	0.9598	0.9366	0.9535
Area2	0.8396	0.8699	0.7459	0.7892
Area3	0.8584	0.8362	0.7348	0.7952

样本	孙美玲^[8]	Chen等^[17]	Shao和Chen^[18]	Axelsson^[10]	Pingle等^[19]	Mougus和?alik^[20]	Wang等^[4]	本文
Samp11	0.1330	0.1301	0.1188	0.1076	0.0900	0.1101	0.1949	0.1871
Samp12	0.0288	0.0338	0.0402	0.0325	0.0297	0.0517	0.0402	0.0711
Samp21	0.0140	0.0134	0.0467	0.0425	0.0146	0.0198	0.0205	0.0165
Samp22	0.0452	0.0467	0.0551	0.0363	0.0322	0.0656	0.0497	0.1150
Samp23	0.0461	0.0524	0.0480	0.0400	0.0748	0.0583	0.0591	0.0770
Samp24	0.0452	0.0629	0.0497	0.0442	0.0416	0.0798	0.0634	0.0788
Samp31	0.0124	0.0111	0.0121	0.0478	0.0220	0.0334	0.0158	0.0280
Samp41	0.0361	0.0558	0.0491	0.1391	0.0923	0.0371	0.0217	0.0237
Samp42	0.0096	0.0172	0.0214	0.0162	0.0305	0.0572	0.0107	0.0193
平均值	0.0412	0.0470	0.0490	0.0562	0.0475	0.0570	0.0529	0.0685

算法	Area2			Area3
算法	R_com	R_cor	R_Q	R_com	R_cor	R_Q
本文	0.8400	0.8700	0.7460	0.8580	0.8300	0.7350
CAL1	0.3260	0.9100	0.3160	0.3780	0.9310	0.3680
HANC1	0.8310	0.7260	0.6330	0.8970	0.6540	0.6080
LJU1	0.6870	0.8520	0.6140	0.8340	0.6060	0.5410
TUM	0.8050	0.8570	0.7100	0.7890	0.8420	0.6870
WHUZ	0.6040	0.6980	0.4790	0.4720	0.7220	0.3990
KNTU	0.6200	0.6470	0.4630	0.6430	0.7770	0.5430
ITCM	0.3720	0.8590	0.3510	0.5440	0.8350	0.4910
ITCR	0.6940	0.8220	0.6030	0.7560	0.7620	0.6120
CAL2	0.7480	0.8200	0.6430	0.7310	0.8150	0.6270