EN中文

Journal of Geo-information Science >

2016 , Vol. 18 >Issue 7: 878 - 885

DOI: https://doi.org/10.3724/SP.J.1047.2016.00878

Orginal Article

A Method for Semiautomated Segmentation of Building Facade from Mobile Laser Scanning Point Cloud Based on Feature Images and SVM

  • PENG Chen ,
  • YU Bailang ,
  • WU Bin ,
  • WU Jianping , *
Expand
  • Key Laboratory of Geographic Information Science, Ministry of Education, East China Normal University, Shanghai 200241, China
*Corresponding author: WU Jianping, E-mail:

Received date: 2015-07-19

  Request revised date: 2015-09-14

  Online published: 2016-07-15

Copyright

《地球信息科学学报》编辑部 所有
Fold

Abstract

Building facade is an important component of urban street features. Delineating and representing the building facade would benefit the urban building design and planning. As a new mobile mapping system, Mobile Laser Scanning (MLS) allows the quick and cost-effective acquisition of close-range three-dimensional (3D) measurements of urban street objects. This paper presents a semiautomated segmentation method for identifying the building facades from MLS point clouds data. The method consists of three major steps: (1) a horizontal grid system is built for the study area, and the multidimensional geometric features of 3D point clouds data, including the normal vector feature, omni-variance feature, geometric dimensionality of α1, α2 and α3, and eigen-entropy feature, are defined and calculated. Then, a feature image is created after projecting these features to the horizontal grid. (2) Building facades are roughly extracted using Support Vector Machine (SVM). (3) The rough extraction result is filtered according to the characteristics of grid including the shape coefficient, grid′s area, and the largest elevation. Two MLS point cloud datasets of Carnegie Mellon University (CMU) database were used in this study to estimate the feasibility and effectiveness of the method. It was found that this method performs well in extracting the building facades. The precision of the results is 0.88, and its recall rate is 0.90, which is better than some existing methods. Our method provides an effective tool for extracting building facades of streets from MLS point cloud data.

Key words： Mobile Laser Scanning system; building facades; segmentation; feature images; Support Vector Machine (SVM)

Cite this article

PENG Chen , YU Bailang , WU Bin , WU Jianping . A Method for Semiautomated Segmentation of Building Facade from Mobile Laser Scanning Point Cloud Based on Feature Images and SVM[J]. Journal of Geo-information Science, 2016 , 18(7) : 878 -885 . DOI: 10.3724/SP.J.1047.2016.00878

1 引言

随着三维空间信息获取技术的发展,激光扫描技术（Laser Scanning）已经成为获取城市地表三维空间信息最快最有效的手段之一。近年来,从激光扫描数据中提取建筑物点云已经成为计算机图形学和城市遥感的一个热点问题。
根据搭载平台的不同,激光扫描系统可分为星载、机载、移动和地面固定等类型。星载激光扫描数据精度较低,并不适合城市建筑物提取;机载激光扫描系统能较快获得较大范围的高精度数据,已有大量研究提出了从机载激光扫描数据中提取建筑物的方法[1-3]。虽然机载扫描数据能够覆盖大范围的城市区域,但是它无法获取建筑物的垂直立面。移动激光扫描能够获取海量的三维城市街区数据,相比地面固定扫描系统具有较强的可移动性,且点云密度相对均匀。但同时,由于道路两旁地物相互遮挡和数据的海量性,使从点云数据中高精度地提取建筑物十分困难。因此,研究从移动激光扫描数据中提取建筑立面的方法具有重要的理论意义和实际应用价值。
近年来,已有大量基于移动激光扫描点云数据的地物提取和分类的研究。根据分类对象的不同,大致可以分为3类：基于单个点[4-5]、基于体元[6-8]和基于特征图像[9-15]的地物分类提取。
基于单个点几何形状特征的分类方法,首先通过计算点云维数特征将点云划分为线状（一维）、面状（二维）和簇状（三维）,然后利用聚类算法或机器学习算法进行分类。例如,Lalonde等结合维数特征和高斯混合模型将点云分为线、面和簇3类,获得了较好的结果[5];考虑到直接使用固定邻域半径来估算维数特征存在较大的误差,Demantke等提出了通过最小化熵值来获得最佳邻域半径的方法[4];在该算法的基础上,Yang等增加了反射率信息获取了更佳的邻域半径,并使用支持向量机（Support Vector Machine,SVM）对点云进行分类,然后通过二次聚类的方法将不同的地物分开[16]。
基于体元或超体元的分类方法目前已经有大量研究。由于点云数据的海量性,基于单个点几何形状特征的分类方法计算时间较长。为了减少计算量,一些学者引入体元（Voxel）和超体元（Super Voxel）的概念,通过将三维空间体元化,达到数据压缩的效果。此外,还能利用八叉树等数据结构来加快查询和计算速度。例如,Wu等首先使用体元组织点云数据,然后使用分层区域生长的算法来提取单株行道树[8],无论是计算速度还是提取精度都有所提升;Lim利用超体元来加速条件随机场（Conditional random field,CRF）的计算[7];Aijazi结合点云特征和体元生成超体元,然后使用连接链（link-chain）代替区域生长算法来进行地物识别和分类[6]。
基于特征图像的分类方法虽然会损失一部分点云信息,但是相比于前2类算法的计算量更小。城市地物一般垂直于水平面,具有很强的2.5维特性,利用这一特点,部分学者提出了点投影算法,即在水平面构建一个虚拟的网格,采用正射投影的方法将三维点云投影到二维空间,形成一个二维的点云特征图像,再利用图像处理算法来进行点云的提取和分类。例如,史文中等提出了投影点密度法（Density of Projected Points,DoPP）来提取建筑物[12],但是这种算法很难设置一个合理的建筑物边界阈值;吴宾等对这种算法进行改进,提出了分层点密度法提取行道树[13]。高程是地物的一个很重要的特征,许多学者利用这一特点,将地物高程投影到二维平面生成高程特征图像[9-11]来进行粗分类。例如,杨必胜等对DoPP算法进行改进,认为每个网格的值应该由点云的高程和距中心点的水平距离决定,利用图像处理的算法提取单个地物[15],再使用PCA（Principal Component Analysis）等算法对地物进行分类和建筑物边界提取[14],对点密度投影法有较大的改进。
根据已有研究现状,本文提出了一种基于多维特征图像和SVM的建筑物立面点云提取算法。本方法的优点在于：（1）延续点投影法的优点,降低计算开销,同时将多种点云特征生成的特征图像与SVM相结合;（2）利用网格属性过滤法弥补了点投影法分类精度较低的缺点,在建筑物立面提取上取得了较好的结果。

2 建筑物立面点云提取算法

本文算法主要包含3个步骤：（1）构建研究区域水平网格;（2）生成点云特征图像;（3）精确提取建筑物立面。具体流程如图1所示。
Fig.1 Flow chart of the algorithm

图1 算法流程图

2.1 构建研究区域水平网格

按照一定间距,将二维平面分割成相同大小的网格,然后把三维点云按照一定规则垂直投影到网格上。虽然网格化可能会使不同地物的点云糅杂在一个网格里,对建筑物提取的精度造成影响,但是由于城市建筑物是一种具有相对规则形状的人造地物,立面通常基本垂直于水平面,且高度较高,垂直方向上不存在其他地物的点云。因此,提取出的建筑物网格对应的三维点云是完整的建筑物点云。
对于给定的激光扫描点云数据 P Lidar = { p 1 , p 2 , … , p n } 中的任意点 p k 都有 ( x k ， y k ， z k ) 三维坐标。首先定义二维平面网格大小长度 （ l ） 和宽度 （ w ） ,然后计算出研究区域最大和最小的三维坐标 （ X min 、 X max 、 Y min 、 Y max 、 Z min 、 Z max ） 。网格原点在三维空间的坐标定义为 （ X min , Y min , Z min ） ,从而保证点云都在网格的正上方。
根据最大和最小 XY 坐标可将水平网格分割成 I 行和 J 列,用 g ( i , j ) 代表第 i 行、第 j 列的网格,如式（1）所示。如果网格 g ( i , j ) 柱面上存在激光点则为非空网格,标记为1,否则为空网格,标记为0,不参与后续计算。对于研究区域的水平网格系统 G ,有。定义非空网格 g ( i , j ) 的点集为 p m 1 ≤ m ≤ M , p m ∈ g ( i , j ) } ,其中 M 为点集中点的个数。对于 p m 所属的网格,同样可将横纵坐标 ( x m ， y m ) 代入式（1）进行计算。
i = int ( ( x - X min ) / l ) j = int ( ( y - Y min ) / w ) （1）

2.2 生成点云特征图像

城市环境非常复杂,许多地物（如树木、电线杆和路灯等）都和建筑物的高程相似,而不同建筑物高程差异也较大,同时很多情况下地物会相互遮挡,导致数据缺失,因此,在很多情况下直接使用高程进行地物分割并不适用。对于同一类地物的点,局部几何特征具有一定的相似性,所以本方法选取点云的6种特征,分别为法向量与水平面夹角 θ 、全方差 v 、维数特征 α 1 、 α 2 、 α 3 和特征根熵 E f ,生成点云特征图像,作为分类。
估算一个点法向量和特征根近似于估计一个相切面法线问题,可以找到该点最近点云,使用最小二乘法进行平面拟合。对于点云中的每个点 p k ,使用最近邻法到最近的 K 个点,然后计算这 K 个点的协方差矩阵 C （式（2））, p ̅ 为所有点的重心。
C = 1 K ∑ i = 1 K ( p i - p ̅ ) ( p i - p ̅ ) T （2）
式中 ： p ̅ 为 K 个点三维坐标 ( x i ， y i ， z i ) 的算术平均。
通过数学变换可以发现,点 p k 的特征根为 λ k = ( λ 1 , λ 2 , λ 3 ) ,法向量估计近似于求取 C 最小特征根 λ 3 对应的特征向量 n k = e ̅ 3 = ( n x , n y , n z ) ,如式（3）所示。
C = e ̅ 1 e ̅ 2 e ̅ 3 λ 1 0 0 0 λ 2 0 0 0 λ 3 e ̅ 1 e ̅ 2 e ¯ 3 ( λ 1 ≥ λ 2 ≥ λ 3 ) （3）
建筑物立面垂直于地面,朝向并不固定,所以只需要提取点 p k 的法向量与水平面的夹角 θ 作为其局部几何特征之一,如式（4）所示。
θ = arccos ( n x 2 + n y 2 ) （4）
点 p k 的特征根在计算前必须使用式（5）进行归一化处理,归一化后的所有特征根之和为1,然后再使用归一化特征根来计算点局部几何特征。首先,根据式（6）计算点云的全方差 v ;然后使用维数特征 α 1 、 α 2 、 α 3 （式（7））表示点云属于线状、面状还是簇状;最后使用特征根熵 E f （式（8））描述点属于某一形状的概率。
λ i = λ i ∑ i = 1 3 λ i （5）
v = ∏ i = 1 3 λ i （6）
α 1 = ( λ 1 - λ 2 ) λ 1 α 2 = ( λ 2 - λ 3 ) λ 1 α 3 = λ 3 λ 1 （7）
E f = - λ 1 ln ( λ 1 ) - λ 2 ln ( λ 2 ) - λ 3 ln ( λ 3 ) （8）
综合上述所有特征,本文共构建了6维特征的向量 F （式（9））。对于任意一个包含了 m 个点的网格 g ( i , j ) ,将这 m 个点的6维特征分别进行加权平均便得到了网格特征 I i , j （式（10））。而对于整个网格系统 G ,一共可以生成6张特征图像 I = { I θ , I v , I α 1 , I α 2 , I α 3 , I E f } 。
F = θ , v , α 1 , α 2 , α 3 , E f （9）
I i , j = 1 m ∑ F ( p m ) , p m ∈ g ( i , j ) （10）

2.3 建筑物立面网格提取

点云特征通过网格化的方式生成了多维特征图像 I 。通过分析发现,建筑物的高程明显高于多数其它地物,而且建筑物立面与水平面的夹角较大,所以可使用线性分类器区分建筑物网格与非建筑网格。由于SVM相对于其他线性监督分类器能够较好地处理高维数据,并具有较好的稳定性[17],所以本文选用线性核函数的SVM来进行训练和分类。随机选取相同个数的建筑物立面与非建筑物立面网格点作为训练样本,训练建筑物网格分类模型;然后把六维特征图像 I 作为模型的输入特征,得到输出结果二值化图像 BI 。同时,直接使用SVM训练得到结果会出现很多错分类网格,必须使用建筑物立面网格的网格属性对非建筑物立面网格进行滤除。
首先,使用连通区域标记算法提取连续的建筑物区域;
然后,定义网格属性：
（1）建筑物立面投影多为线状或者是折线状,所以定义了形状系数 CI （式（11））。如果网格区域为一个正圆形,则 CI 接近1;如果 CI 为线状,则 CI 接近0。设置阈值 C I 0 ,令第 k 个区域 C I k < C I 0 ;
（2）建筑物属于较大型的地物,因此网格区域应该较大,所以设定参数区域面积 A （式（12））,并且设定阈值 A 0 ,令第 k 个区域 A k > A 0 ;
（3）建筑物属于较高的人造地物,因此其最大高程值 MH 必须高于指定阈值 M H 0 。
CI = 4 π A P 2 （11）
A = n × l × w （12）
式中： P 为网格区域的周长; l 和 w 分别为网格的长度和宽度。
最后,将所得到的建筑物立面网格反投影到三维空间,得到精确的建筑物立面。

3 实验结果与分析

3.1 数据源

本文采用美国卡内基梅隆大学的移动激光扫描点云数据库的3个数据集data-1、data-2和data-3,分别包含了943 596、999 466和1 032 542个点;每个数据集的研究区域覆盖范围分别是168 m×463 m、488 m×409 m和390 m×368 m。3个标准数据集预先通过人工精确分类,总共分为44小类（表1）,主要为建筑物立面、地表、树木、行人和电线等。图2为实验数据的概览图,其中图例为点云的相对高程。
Tab.1 Labeled id and name of CMU Oakland 3-D point cloud dataset

表1 卡内基梅隆大学移动激光扫描点云数据库的分类编号和类别

编号 类别 编号 类别 编号 类别 编号 类别
1001 undet 1109 fire_hydrant 1202 ground 1401 wall
1002 linear_misc 1110 post 1203 paved_road 1402 stairs
1003 surf_misc 1111 sign 1205 curb 1408 fence
1101 wire_bundle 1113 bench 1206 walkway 1409 gate
1102 isolated_wire 1114 lamp 1300 foliage 1410 ceiling
1103 utility_pole 1115 traffict_lights 1301 grass 1411 facade_ledge
1104 crossarm 1116 traffic_lights_support 1302 small_trunk 1412 column
1105 support_wire 1117 garbage 1303 large_trunk 1413 mailbox
1106 support_pole 1118 crosswalk_light 1305 thick_branch 1500 human
1107 lamp_support 1119 parking_meter 1306 shrub 1501 vehicle
1108 transformer 1200 load_bearing 1400 facade 9999 legacy
Fig.2 Overview of CMU Oakland 3-D point cloud dataset.

图2 卡内基梅隆大学移动激光扫描点云数据库

3.2 点云特征图像生成

为了避免点云特征丢失,同时有效地保持非空网格的连通性,通过多次试验,网格最合适的宽度为0.25 m×0.25 m,计算点局部几何特征选取最近30个点,生成的特征图像如图3所示。图3（a）是人工分类的原始点云数据,其中青色表示建筑物立面,黄色表示道路,绿色表示树木,蓝色表示车辆;图3（b）原始点云网格化生成的点云类别图像,绿色为建筑物网格,深红色为非建筑物立面网格。从特征图像（图3（c）-（h））可看出,建筑物立面与地面夹角较大,同其他地物有较大区别,全方差较小,主要的几何维数特征表现为面状。
Fig.3 Feature images generated by 3D point clouds

图3 点云特征图像

3.3 建筑物点云提取结果

网格分类所选用的分类器为LIBSVM,核函数选用线性核函数。从每个研究区域中随机选取1000个建筑物和非建筑物立面网格作为训练样本,构建分类模型。表2为使用模型粗分类的结果。粗分类的精度相对较高,3个数据集的精度分别为90.7%、95.6%和96.1%;但是召回率相对较低,分别为62.6%、41.1%和51.9%。结果表明,在粗分类的过程中,容易把大量的非建筑物立面网格误分建筑物立面网格。
Tab.2 Precisions and recalls of classification based on the building grids by linear SVM

表2 建筑物网格粗提取精度和召回率

数据集 网格类型 精度/(%) 召回率/(%)
Data-1 建筑物立面网格 90.7 62.6
非建筑物立面网格 94.7 99.0
Data-2 建筑物立面网格 95.6 41.1
非建筑物立面网格 94.3 99.8
Data-3 建筑物立面网格 96.1 51.9
非建筑物立面网格 93.6 99.7
为了滤除误分类的建筑物网格,对每个研究区分别设定合适的网格属性（形状系数、网格面积和最大高程值）。建筑物属于较大型的地物,所以所占用的网格数一般较大,设置为网格面积为20-44;建筑物网格的形状系数较小,所以阈值设定为0.45-0.5之间;而建筑物高程值一般大于3 m。利用以上阈值对粗分类结果进行精分类后,得到建筑物立面提取结果（图4）,云精度如表3所示。建筑物最终分类的精度均超过80%,分别为81.7%、89.8%和83.6%;召回率分别为89.4%、90.7%和90.0%。
Fig.4 Results of segmentation

图4 建筑物立面提取结果

Tab.3 Precisions and recalls of building facade segmentation from point clouds

表3 建筑物立面点云提取精度和召回率

数据集 A/m2 CI MH/m 精度/(%) 召回率/(%)
Data-1 1.25 0.45 3 81.7 89.4
Data-2 2.75 0.45 3 89.8 90.7
Data-3 1.25 0.45 3 83.6 90.0
对比图2可发现,data-1有一整栋建筑物没有提取出来,导致整体分类精度较低。针对此问题,对比实际建筑物照片和建筑物激光扫描点云（图5）,图5（a）为建筑物在谷歌街景中的照片,图5（b）为建筑物在点云中的形状。从图中可看出,该建筑物立面是由较为规则的多个小面片组成,而在点云中则表现为多个不连续的柱状,因此在提取和过滤时,建筑物被当做柱状物过滤掉了。为了验证该想法,首先将data-1中的该建筑物去除,过滤的网格属性与原来一致,分别设置为20、0.45和3,得到的精度和召回率分别为91.4%和91.5%,比原来的精度提高了约10%,召回率提高了2.1%。
Fig.5 The building which is not extracted from data-1

图5 data-1中未提取的建筑物

3.4 实验结果分析对比

为了验证本算法,选取了同样使用CMU数据集且较为先进的3种算法：S3DP（Stacked 3-D Parsing）[18]、LogR（K-class logistic regression）[18]和M3N（Max-Margin Markov Network）[19],其将点云数据分类为电线、杆状物、地面、植被、树干、建筑物和车辆7类,本文只针对建筑物这一类进行比较。由于分类结果在精度和召回率之间难以权衡,故设定综合评价指标F1（式（13））,F1较高时说明试验方法有效。
F 1 = 2 × P × R P + R （13）
式中： P 为精度; R 为召回率。对比结果汇总如表3所示。
从表4可看出,LogR算法表现最差,综合评价指标 F 1 为0.8,远低于其他算法;S3DP算法表现最好,精度为83%,召回率为93%, F 1 为0.88。由于受上述不规则建筑物的影响,本方法的建筑物立面提取结果略低于S3DP算法,精度为84%,召回率为90%, F 1 为0.87。去除不规则建筑物影响后,精度为88%,召回率为91%, F 1 为0.9,较修正前结果有较大提升。所以,本算法在较规则的建筑物立面提取上有较好的提取结果。
Tab.4 Precisions/recalls of different methods

表4 不同方法建筑物立面提取结果比较分析

计算方法 精度/(%) 召回率/(%) F1
S3DP 83 93 0.88
M3N 80 92 0.86
LogR 74 87 0.80
本文方法 84 90 0.87
修正后本文方法 88 91 0.90

4 讨论

4.1 网格大小的选取

不同网格大小会对结果产生较大的影响。一方面如果网格过大,不仅会导致同一个网格中包含较多点云,从而损失大量的细节信息,也会使单个网格里含有不同地物点的概率增大,特征图像的混合网格增多,影响分类的精度;另一方面,如果网格过小,不仅会使特征图像的网格个数增多,增加计算复杂度,还会使有效网格密度变小,造成特征图像相同地物邻域不连通的问题。
本文采用不同的网格大小对点云数据库data-2进行试验,试验结果如图6所示。结果表明网格大小为0.25 m×0.25 m和0.5 m×0.5 m的提取结果最好;网格大小为0.1 m×0.1 m的提取结果错分和漏分明显偏高;当网格较大（1 m×1 m）时,不同的地物会被分在一个网格中,不易区分,导致分类精度偏低。
Fig.6 The impact of grid size on the extraction results

图6 网格大小对提取结果的影响

4.2 网格属性的选取

网格属性的3个自由参数（形状系数、网格面积、最大高程）对最终结果也有一定影响。其中,建筑物的高程一般不会低于3 m,所以最大高程一般设置为3 m左右;形状系数和网格面积要根据研究区域而设定,如果取值过大,可能会滤除真正的建筑物,但是取值过小,达不到过滤的效果。根据实验得出,面积设置为1.5 m2,形状系数设置在0.45左右较为合适。

4.3 与同类型方法对比

本方法与文献[15]的方法属于同一类型方法,因此本方法也具有计算开销小,计算速度快的优点。不同的是,文献[15]只使用了高度这一维信息,因此在建筑物点云提取时会有2个缺点：首先,低矮的墙面由于高度较低,很容易被过滤掉,如图7（a）中左下角低矮的墙面并没有提取出来;其次,与建筑物高度的相似,并且在特征图像上与之相连的地物很容易被错分,如图7（a）中两侧的树木与建筑物高度类似,并且靠近建筑物,因此被错分为建筑物。而本文方法可以有效地避免上述2个问题（图7（b））。
Fig.7 Comparison of extraction results

图7 提取结果对比

5 结论

本文提出了一种从移动激光扫描数据中半自动提取建筑物立面的方法。此方法延续了DoPP算法中点投影的思想,将三维问题进行“降维”处理,并在此基础上改进了杨必胜等[15]的点投影特征图像算法,计算点的局部几何特征,得到多维点云特征图像;然后选取研究区域训练样本,使用SVM提取建筑物立面网格;最后使用网格属性（网格面积、形状系数和最大高程）过滤非建筑物立面网格,得到了较好的结果。以卡内基梅隆大学的移动激光扫描点云数据试验数据,本方法的提取精度和召回率分别为84%和90%。
本方法也存在一定的局限性,对一般建筑物立面提取比较理想,而对于比较特殊的建筑物立面,提取效果尚不完善。后续研究将对算法进行改进,使算法能够适应更加复杂的建筑物,并且将提取目标扩展到树木、地面、车辆等多种地物。

The authors have declared that no competing interests exist.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]
Haala ., Brenner C.Extraction of buildings and trees in urban environments[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 1999,54(2):130-137.In this article, two methods for data collection in urban environments are presented. The first method combines multispectral imagery and laser altimeter data in an integrated classification for the extraction of buildings, trees and grass-covered areas. The second approach uses laser data and 2D ground plan information to obtain 3D reconstructions of buildings.

DOI

[2]
Suveg ., Vosselman G.Reconstruction of 3D building models from aerial images and maps[J]. ISPRS Journal of Photogrammetry and remote sensing, 2004,58(3):202-224.

[3]
Yu ., Liu ., Wu ., et al.Automated derivation of urban building density information using airborne LiDAR data and object-based method[J]. Landscape and Urban Planning, 2010,98(3):210-219.Building density information is fundamentally important for urban design, planning and management and for urban environmental studies. This paper demonstrates that Building Coverage Ratio (BCR), Floor Area Ratio (FAR), and other building density indicators can be numerically and automatically derived from high-resolution airborne LiDAR data. An object-based method is proposed to process the LiDAR data for the building density information. The method consists of a sequence of numerical operations: generating the normalized Digital Surface Model (nDSM), extracting building objects, deriving object attributes, associating objects with the corresponding land lots, and computing building density indicators at land lot and urban district scales. The algorithms for these operations have been implemented as an ArcGIS extension module. The object-based method is applied to the processing of airborne LiDAR data over downtown Houston. Various attributes have been derived to quantify the building density, urban physical structure, and landscape morphological characteristics of the downtown area at three different spatial scales.

DOI

[4]
Demantké ., Mallet ., David ., et al. Dimensionality based scale selection in 3D lidar point clouds [J]. International Archives of Photogrammetry, Remote Sensin. and Spatial Information Science., Laser Scannin., 2011, XXXVIII-5/W12(97-102.

[5]
Lalonde J ., Vandapel ., Huber D ., et al.Natural terrain classification using three-dimensional ladar data for ground robot mobility[J]. Journal of Field Robotics, 2006,23(10):839-861.Abstract In recent years, much progress has been made in outdoor autonomous navigation. However, safe navigation is still a daunting challenge in terrain containing vegetation. In this paper, we focus on the segmentation of ladar data into three classes using local threedimensional point cloud statistics. The classes are: “scatter” to represent porous volumes such as grass and tree canopy; “linear” to capture thin objects like wires or tree branches, and finally “surface” to capture solid objects like ground surface, rocks, or large trunks. We present the details of the proposed method, and the modifications we made to implement it on-board an autonomous ground vehicle for real-time data processing. Finally,we present results produced from different stationary laser sensors and from field tests using an unmanned ground vehicle. 08 2006 Wiley Periodicals, Inc.

DOI

[6]
Aijazi A ., Checchin ., Trassoudaine L.Segmentation based classification of 3D urban point clouds: A super-voxel based approach with evaluation[J]. Remote Sensing, 2013,5(4):1624-1650.Segmentation and classification of urban range data into different object classes have several challenges due to certain properties of the data, such as density variation, inconsistencies due to missing data and the large data size that require heavy computation and large memory. A method to classify urban scenes based on a super-voxel segmentation of sparse 3D data obtained from LiDAR sensors is presented. The 3D point cloud is first segmented into voxels, which are then characterized by several attributes transforming them into super-voxels. These are joined together by using a link-chain method rather than the usual region growing algorithm to create objects. These objects are then classified using geometrical models and local descriptors. In order to evaluate the results, a new metric that combines both segmentation and classification results simultaneously is presented. The effects of voxel size and incorporation of RGB color and laser reflectance intensity on the classification results are also discussed. The method is evaluated on standard data sets using different metrics to demonstrate its efficacy.

DOI

[7]
Lim E ., Suter D.3D terrestrial LIDAR classifications with super-voxels and multi-scale Conditional Random Fields[J]. Computer-Aided Design, 2009,41(10):701-710.In this paper, we propose a new method for 3D terrestrial laser range data classifications. This functions as the first step towards virtual city model reconstructions from range data and is particularly useful for scene understanding. Classification of the outdoor terrestrial range data into different data types (for example, building surface, vegetation and terrain) is challenging due to certain properties of the data: occlusions due to obstructions, density variation due to different distances of the scanned object from the laser scanner, multiple multi-structure objects and cluttered vegetation. Also, the range data acquired are massive in size and require a lot of computation and memory. Recognizing the redundancy of labeling every individual data, we propose over-segmenting the raw data into adaptive support regions: super-voxels. The super-voxels are computed using 3D scale theory and adapt to the above-mentioned range data properties. Colors and reflectance intensity acquired from the scanner system are combined with geometry features (saliency features and normals) that are extracted from the super-voxels, to form the feature descriptors for the supervised learning model. We proposed using the discriminative Conditional Random Fields for the classification problem and modified the model to incorporate multi-scales for super-voxel labeling. We validated our proposed strategy with synthetic data and real-world outdoor LIDAR (Light Detection and Ranging) data acquired from a Riegl LMS-Z420i terrestrial laser scanner. The results showed great improvement in the training and inference rate while maintaining comparable classification accuracy with previous approaches.

DOI

[8]
Wu ., Yu ., Yue ., et al.A voxel-based method for automated identification and morphological parameters estimation of individual street trees from mobile laser scanning data[J]. Remote Sensing, 2013,5(2):584-611.As an important component of urban vegetation, street trees play an important role in maintenance of environmental quality, aesthetic beauty of urban landscape, and social service for inhabitants. Acquiring accurate and up-to-date inventory information for street trees is required for urban horticultural planning, and municipal urban forest management. This paper presents a new Voxel-based Marked Neighborhood Searching (VMNS) method for efficiently identifying street trees and deriving their morphological parameters from Mobile Laser Scanning (MLS) point cloud data. The VMNS method consists of six technical components: voxelization, calculating values of voxels, searching and marking neighborhoods, extracting potential trees, deriving morphological parameters, and eliminating pole-like objects other than trees. The method is validated and evaluated through two case studies. The evaluation results show that the completeness and correctness of our method for street tree detection are over 98%. The derived morphological parameters, including tree height, crown diameter, diameter at breast height (DBH), and crown base height (CBH), are in a good agreement with the field measurements. Our method provides an effective tool for extracting various morphological parameters for individual street trees from MLS point cloud data.

DOI

[9]
Golovinskiy ., Kim V ., Funkhouser T.Shape-based recognition of 3D point clouds in urban environments[C]. Computer Vision, 2009:2154-2161.

[10]
Serna ., Marcotegui B.Detection, segmentation and classification of 3D urban objects using mathematical morphology and supervised learning[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2014,93(7):243-255.We propose an automatic and robust approach to detect, segment and classify urban objects from 3D point clouds. Processing is carried out using elevation images and the result is reprojected onto the 3D point cloud. First, the ground is segmented and objects are detected as discontinuities on the ground. Then, connected objects are segmented using a watershed approach. Finally, objects are classified using SVM with geometrical and contextual features. Our methodology is evaluated on databases from Ohio (USA) and Paris (France). In the former, our method detects 98% of the objects, 78% of them are correctly segmented and 82% of the well-segmented objects are correctly classified. In the latter, our method leads to an improvement of about 15% on the classification step with respect to previous works. Quantitative results prove that our method not only provides a good performance but is also faster than other works reported in the literature.

DOI

[11]
Zhu ., Zhao ., Liu ., et al.Segmentation and classification of range image from an intelligent vehicle in urban environment[C]. Intelligent Robots and Systems, 2010:1457-1462.

[12]
史文中,李必军,李清泉.基于投影点密度的车载激光扫描距离图像分割方法[J].测绘学报,2005,34(2):95-100.对车载激光扫描系统获取的距离图像的分割进行研究,提出利用投影 点密度进行距离图像分割的方法.该方法不需要其他辅助数据就可以进行距离图像的分割,给出应用实例.图像的分割是目标识别、定位、特征提取与建模的基础和 关键.距离图像是离散坐标点阵列,不表达目标边界特征和拓扑关系,到目前为止还没有距离图像分割的成熟、可行方法.本文研究成果,展现了激光扫描技术在数 据获取、处理研究上的新方向.

DOI

[ Shi W ., Li B ., Li Q Q.A Method for Segmentation of Range Image Captured by Vehicle-borne Laserscanning Based on the Density of Projected Points[J]. Acta Geodaetica et Cartographica Sinica, 2005,34(2):95-100. ]

[13]
吴宾,余柏蒗,岳文辉,等.一种基于车载激光扫描点云数据的单株行道树信息提取方法[J].华东师范大学学报:自然科学版,2013,3(2):38-49.提出了一种基于分层网格点密度的单株树信息提取方法,从车载激光 扫描点云数据中提取出组成单株行道树的激光点,并计算树高、冠幅等特征信息.该方法由建立规则网格,基于高程的点云分层,逐层计算网格点密度,逐层提取激 光点,提取单株行道树和计算特征等步骤构成.通过实例证明,提取结果较好地保留了组成单株行道树的激光点,并能较准确地计算特征信息.该方法拓展了车载激 光扫描系统的应用领域,并可为城市绿化管理提供新的技术方法.

DOI

[ Wu ., Yu B ., Yue W ., et al.Method for identifying individual street trees from the cloud data of the vehicle-borne laser scanning point[J]. Journal of East China Normal University (Natural Science), 2013,3(2):38-49. ]

[14]
杨必胜,董震,魏征,等.从车载激光扫描数据中提取复杂建筑物立面的方法[J].测绘学报,2013,42(3):411-417.提出一种从车载激光扫描数据中提取复杂建筑物立面的新方法.该方 法首先利用“维数特征”方法确定每个扫描点的最佳邻域,进而计算得到每个扫描点精确的局部几何特征(法向量、主方向、维数特征)；然后基于“维数特征”对 扫描点进行粗分类,并设置相应的生长准则对不同类别的扫描数据分别进行分割；最后综合建筑物立面的语义知识对建筑物立面区域进行精确提取.试验结果和比较 分析表明,本文的方法不但能提取建筑物平面和非平面立面,而且可消除点密度差异(变化)对建筑物立面提取结果的影响,提高建筑物立面提取的正确率和完整 性.

[ Yang B ., Wei ., Li Q ., et al.Extracting Complex Building Facades from Mobile Laser Scanning Data[J]. Acta Geodaetica et Cartographica Sinica, 2013,42(3):411-417. ]

[15]
杨必胜,魏征,李清泉,等.面向车载激光扫描点云快速分类的点云特征图像生成方法[J].测绘学报,2010,39(5):540-545.车载激光扫描是空间数据快速获取的一种重要手段。车载激光扫描点云数据的分类和特征提取是目标识别与三维重建的基础。本文以车载激光点云数据为研究对象，提出了一种适合于其快速分类与目标提取的点云特征图像生成方法。该方法首先将扫描区域进行平面规则格网投影，通过分析格网内部点云的空间分布特征（平面距离、高程差异、点密集程度等）确定激光扫描点的定权，从而生成车载激光扫描点云的特征图像。利用生成的点云特征图像，可采用阈值分割、轮廓提取与跟踪等手段提取图像分割的建筑物目标的边界，从而确定边界内部点云数据，实现目标分类与提取。本文以Optech公司的车载激光扫描数据为实验对象，验证了本文提出方法的可行性和实用性。实验结果表明，该方法能快速有效分离出车载激光扫描点云中的地面数据、建筑物数据等。

[ Yang B ., Wei ., Li Q ., et al.A classification-oriented method of feature image generation for vehicle-borne laser scanning point clouds[J]. Acta Geodaetica et Cartographica Sinica, 2010,39(5):540-545. ]

[16]
Yang ., Dong Z.A shape-based segmentation method for mobile laser scanning point clouds[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2013,81(7):19-30.Segmentation of mobile laser point clouds of urban scenes into objects is an important step for post-processing (e.g., interpretation) of point clouds. Point clouds of urban scenes contain numerous objects with significant size variability, complex and incomplete structures, and holes or variable point densities, raising great challenges for the segmentation of mobile laser point clouds. This paper addresses these challenges by proposing a shape-based segmentation method. The proposed method first calculates the optimal neighborhood size of each point to derive the geometric features associated with it, and then classifies the point clouds according to geometric features using support vector machines (SVMs). Second, a set of rules are defined to segment the classified point clouds, and a similarity criterion for segments is proposed to overcome over-segmentation. Finally, the segmentation output is merged based on topological connectivity into a meaningful geometrical abstraction. The proposed method has been tested on point clouds of two urban scenes obtained by different mobile laser scanners. The results show that the proposed method segments large-scale mobile laser point clouds with good accuracy and computationally effective time cost, and that it segments pole-like objects particularly well.

DOI

[17]
张学工. 关于统计学习理论与支持向量机[J].自动化学报,2000,26(1):32-42.模式识别、函数拟合及概率密度估计等都属于基于数据学习的问题,现有方法的重要基础是传统的统计学,前提是有足够多样本,当样本数目有限时难以取得理想的效果.统计学习理论(SLT)是由Vapnik等人提出的一种小样本统计理论,着重研究在小样本情况下的统计规律及学习方法性质.SLT为机器学习问题建立了一个较好的理论框架,也发展了一种新的通用学习算法--支持向量机(SVM),能够较好的解决小样本学习问题.目前,SLT和SVM已成为国际上机器学习领域新的研究热点.本文是一篇综述,旨在介绍SLT和SVM的基本思想、特点和研究发展现状,以引起国内学者的进一步关注.

[ Zhang X G.Introduction to statistical learning theory and support vector machines[J]. Acta Autom Atic A Sinica, 2000,26(1):32-42. ]

[18]
Xiong ., Munoz ., Bagnell J ., et al.3-d scene analysis via sequenced predictions over points and regions[C]. Robotics and Automation, 2011:2609-2616.

[19]
Munoz ., Bagnell J ., Vandapel ., et al.Contextual classification with functional max-margin markov networks[C]. Computer Vision and Pattern Recognition, 2009:975-982.

Options
Outlines

/

〈 〉