Advanced Engineering Informatics 34 (2017) 1–16

Contents lists available at ScienceDirect

Advanced Engineering Informatics

journal homepage: www.elsevier.com/locate/aei

Full length article

Robust normal estimation and region growing segmentation of

infrastructure 3D point cloud models
Ali Khaloo, David Lattanzi ⇑
Department of Civil, Environmental, and Infrastructure Engineering, George Mason University, Fairfax, VA 22030, United States

a r t i c l e i n f o a b s t r a c t

Article history: Modern remote sensing technologies such as three-dimensional (3D) laser scanners and image-based 3D
Received 30 March 2017 scene reconstruction are in increasing demand for applications in civil infrastructure design, mainte-
Received in revised form 10 July 2017 nance, operation, and as-built construction verification. The complex nature of the 3D point clouds these
Accepted 11 July 2017
technologies generate, as well as the often massive scale of the 3D data, make it inefficient and time con-
Available online 2 August 2017
suming to manually analyze and manipulate point clouds, and highlights the need for automated analysis
techniques. This paper presents one such technique, a new region growing algorithm for the automated
Keywords:
segmentation of both planar and non-planar surfaces in point clouds. A core component of the algorithm
Segmentation
3D point cloud models
is a new point normal estimation method, an essential task for many point cloud processing algorithms.
Robust estimation The newly developed estimation method utilizes robust multivariate statistical outlier analysis for reli-
Outliers able normal estimation in complex 3D models, considering that these models often contain regions of
3D reconstruction varying surface roughness, a mixture of high curvature and low curvature regions, and sharp features.
Computer vision An adaptation of Mahalanobis distance, in which the mean vector and covariance matrix are derived from
Normal estimation a high-breakdown multivariate location and scale estimator called Deterministic MM-estimator (DetMM)
3D data processing is used to find and discard outlier points prior to estimating the best local tangent plane around any point
in a cloud. This approach is capable of more accurately estimating point normals located in highly curved
regions or near sharp features. Thereafter, the estimated point normals serve a region growing segmen-
tation algorithm that only requires a single input parameter, an improvement over existing methods
which typically require two control parameters. The reliability and robustness of the normal estimation
subroutine was compared against well-known normal estimation methods including the Minimum
Volume Ellipsoid (MVE) and Minimum Covariance Determinant (MCD) estimators, along with
Maximum Likelihood Sample Consensus (MLESAC). The overall region growing segmentation algorithm
was then experimentally validated on several challenging 3D point clouds of real-world infrastructure
systems. The results indicate that the developed approach performs more accurately and robustly in
comparison with conventional region growing methods, particularly in the presence of sharp features,
outliers and noise.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction algorithm [1], a process that converts 2D images into 3D point

Modern remote sensing technologies have enabled the genera-
tion of accurate three-dimensional (3D) geometric models that can
be used to augment and improve condition assessment, progress
tracking and documentation of civil infrastructure systems. While
there are many methods for creating such 3D models, terrestrial
laser scanning (TLS) is the most widely used. Photogrammetric
reconstruction via the Dense Structure-from-Motion (DSfM)
⇑ Corresponding author.
E-mail addresses: akhaloo@gmu.edu (A. Khaloo), dlattanz@gmu.edu (D. Lattanzi).
clouds, is often considered as a low cost alternative to TLS systems.
Accurate and efficient processing techniques of these often
massive 3D data is vital for extracting useful information. Point
cloud processing algorithms can be used for applications such as
automatic change detection and deformation monitoring of struc-
tures [2,3], or the generation of as-built 3D models in the form of
Computer-Aided Design (CAD) [4,5] or Building Information Mod-
els (BIM) [6–8]. In many algorithms, these procedures initiate by
segmenting the point cloud data into appropriate segments and
then recognizing primitives from the segments to generate 3D
models [9].

http://dx.doi.org/10.1016/j.aei.2017.07.002
1474-0346/Ó 2017 Elsevier Ltd. All rights reserved.
2 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
Segmenting point clouds of civil infrastructure systems is par-
ticularly challenging, as these massive 3D data sets are noisy, spar-
sely populated, and tend to have uneven point distribution (local
point density). Scanning noise is unavoidable and varies due to
laser scanning, imaging sensors, and radiometric scanning condi-
tions. The performance of the image-based 3D reconstructions also
vary due to camera imaging parameters, camera placement, scene
complexity, surface texture, environmental conditions and scene
size. Furthermore, the large-scale nature of most infrastructure
systems, their typically complex geometry (from sharp features
to highly curved regions), and the variety of textures due to varying
construction materials (from smooth to very irregular), makes the
segmentation of these 3D point clouds even more challenging. Seg-
mentation methods that are robust in light of such variability are
highly desirable.
The focus of this paper is on the development and validation of
a point cloud segmentation algorithm that is designed with these
considerations in mind. The algorithm uses a region growing
approach to isolate and segment both planar and non-planar sur-
faces in large scale 3D point clouds. As a first step in this larger
algorithm, point normals for each point in a cloud are estimated.
A new and more robust outlier detection process has been
designed to improve these estimations.
2. Related work
The state of research in as-built 3D modeling of infrastructure
systems and semi-automated segmentation routines were covered
in comprehensive review studies by Pătrăucean et al. [8] and Son
et al. [10], respectively. Previous work in this domain can be sepa-
rated into studies on normal estimation and/or point cloud
segmentation.
2.1. Prior work in normal estimation
Reliable estimation of the normal vector at each point in a 3D
point cloud has become a fundamental step in point cloud data
processing. Numerous algorithms rely on accurate normal estima-
tion, such as point-based rendering, surface reconstruction, 3D
piecewise-planar reconstruction, and 3D point cloud segmentation
[11]. Therefore, methods that provide a good trade-off between
computational efficiency and accuracy are essential. Normal esti-
mation in 3D point clouds can be classified into Voronoi/Delaunay
and Regression methods. The comparative study by Dey et al. [12]
showed that Voronoi methods are unreliable in noisy point clouds
or in clouds with variations in local density variation, making this
class of methods unsuitable for civil infrastructure applications.
Regression based normal estimation was pioneered in the work
by Hoppe et al. [13], as a module in their surface reconstruction
algorithm. Their method estimated the normal at a point p by fit-
ting a local plane under the assumption of planarity in the region of
the k-nearest neighbors of a given point, using Principal Compo-
nent Analysis (PCA) [14]. Due to the sensitivity of PCA to outliers
and the implicit smoothness assumption, this method is sensitive
to outliers and noise, and cannot handle sharp features. To avoid
strong assumptions for a local 3D neighborhood and reduce the
sensitivity of PCA, both Mitra et al. [15] and Wang et al. [16] sug-
gested an adaptive neighborhood size based on an iterative scheme
to find the optimal k closest 3D points. Pauly et al. [17] assigned
Gaussian weights to distances between point p and its k-nearest
neighbors. To make the normal estimation more consistent near
geometric singularities, such as corners and edges, Castillo et al.
[18] formulated PCA as a constrained nonlinear least squares prob-
lem (NLSQ) to assign appropriate weights to neighboring points in
order to minimize the contribution of points located along
discontinuities.
Recently, estimating point normals in 3D point clouds with
sharp features has been studied. Li et al. [19] implemented a statis-
tical estimator for local noise in tandem with RANdom Sample
Consensus (RANSAC) [20] to make the normal estimation proce-
dure more accurate in the presence of sharp features and planar
transitions. However, their method was not efficient for large 3D
point clouds with variations in local point density. In the work
by Boulch and Marlet [11], a version of Randomized Hough Trans-
form (RHT) was utilized for estimating point normals in point
clouds with sharp features. For more information on regression
based methods and their performance, the reader is referred to
comparative studies and detailed analysis by Klasing et al. [21]
and Song et al. [22].
Within civil engineering applications, it is possible to improve
the performance of many normal vector-based 3D model analysis
techniques such as automated crack detection [23], deformation
measurement [24] and 3D as-built modeling [25,26], through a
more accurate estimation of point normals.
2.2. Prior work in point cloud segmentation
Segmentation algorithms can be roughly classified into three
categories: Model fitting [27], Clustering [28] and Region growing
[29] methods.
Model fitting-based methods attempt to find sets of points in
the scene that fit 3D primitives. These methods generally use dif-
ferent adaptations of well-known algorithms such as Hough Trans-
form (HT) and RANSAC [30]. In most cases these techniques are
able to identify surfaces in cluttered scenes. However, they lever-
age shape primitives and are thus context specific. These methods
have been used to segment point clouds of civil infrastructure sys-
tems [31,32], but tend to perform poorly in the face of large data
sets or complex geometries [30] and were not considered further
in this study.
Clustering-based methods aim to segment sets of points based
on clustering local feature attributes of 3D point cloud data [33].
These techniques are computationally expensive for multi-
dimensional features in large datasets and their performances are
heavily impacted in the presence of a high proportion of outliers
[28]. To date these approaches do not perform well on points
placed on discontinuities (e.g. sharp features and planar transi-
tions), since the extracted feature vectors of these points often dif-
fer from other points located on adjacent planes.
Region growing-based methods [29] use local features
extracted from a neighborhood around each point to aggregate
nearby points with similar properties and segment a region of a
point cloud. Rabbani et al. [34] proposed a region growing tech-
nique based on smoothness constraints to extract smoothly con-
nected areas in unstructured 3D point clouds. This method uses
surface normal vectors and approximated local curvature values
(which depends on normal estimation and orientation) as mea-
sures of local geometry to determine each segment. Since the nor-
mal estimation in this approach was calculated by using the PCA-
based method [13], the algorithm is susceptible to noise and sharp
features within the point cloud model. In the work by Son et al.
[35], the aforementioned method was implemented to automati-
cally extract 3D points corresponding to pipelines in an industrial
plant. Walsh et al. [36] also implemented this algorithm to seg-
ment point clouds of small structural elements. In both of these
studies, the targeted 3D scanned data was relatively ‘‘noise-free”
and without any major challenges such as missing data due to
occlusion or significant variability in surface roughness and curva-
ture. Inspired by [34], Dimitrov and Golparvar-Fard [9] introduced
a region growing segmentation for building point clouds that
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
leverages multi-scale feature detection, such as roughness and cur-
vature, for each 3D point.
2.3. Contributions of this work
As shown by Yoon et al. [37], inaccurate estimation of point
normals is the most likely source of errors in segmentation,
emphasizing the need for advancements in normal estimation
algorithms, in particular for civil infrastructure system models
with a high number of sharp planar transitions. Current region
growing methods also require multiple user defined control
parameters, creating an algorithm sensitive to user inputs and sus-
ceptible to variations in surface roughness and cloud density. Over-
all, there is a need to develop accurate and efficient normal
estimation and segmentation algorithms capable of handling large
scale and noisy point clouds of civil infrastructure systems, which
often contain both highly planar and curvilinear objects.
This paper presents a new region growing method that
addresses the challenges of sharp features (planar transitions),
gaps due to occlusion and missing data, and outliers/noise. An effi-
cient and robust statistical outlier detection process is imple-
mented to more accurately estimate normal vectors near sharp
features, and in the presence of varying levels of noise and sam-
pling density. These improved estimates of point normals are then
used in a region growing segmentation that requires only a single
input parameter, minimizing parameter sensitivity. This is accom-
plished through the embedment of a locally adaptive criteria
within the region growth step, providing an intrinsic ability to han-
dle large variations in surface roughness without the need for a
control parameter.
The performance of the developed normal estimation algorithm
was first compared against a number of well-established alterna-
tive processes such as the PCA-based approach [13]. The overall
segmentation algorithm was then compared with the existing
region growing approach implemented in the Point Cloud Library
(PCL) [38]. The results of these comparative studies illustrated that
the new segmentation approach has a higher robustness in the
presence of noise, sharp features, and high curvature regions.
3. Methodology
The overall segmentation algorithm (Fig. 1) initiates by first
identifying the k-nearest neighbors (k-NN) for every point in the
dataset. The linear search solution for k-NN has a running time
of OðndÞ, where n is the number of points and d is the dimension-
Fig. 1. The automated robust segmentation process.
3
ality of the data. For dense point clouds of infrastructure systems,
which often contain hundreds of millions of points, computing the
k-NN of every point is computationally prohibitive. In order to
make this process more efficient, the k-d trees space partitioning
method is applied to the k-NN search [39], to reduce the search
complexity to Oðlog nÞ.
3.1. Normal and curvature estimation
After determining the k-NN for a point p, the approximate point
normal is then estimated. The salient local 3D features, the point
normal vector and curvature, are first estimated for each point p.
The conventional approach to normal estimation is to use Principal
Component Analysis (PCA) [13], with the point normal estimated
as the eigenvector associated with the smallest eigenvalue of the
symmetric positive semi-definite matrix:
1X k
1X k
C 33 ¼
ÞT ðpi  p
ðp  p
Þ;
¼
p p ð1Þ
k i¼1 i k i¼1 i

is the local data centroid
where C is the covariance matrix and p
(arithmetic mean) within the neighborhood N pi of point p. By per-
forming Singular Value Decomposition (SVD) [40] on the covariance
matrix, it is possible to compute the eigenvectors V (v 2 ; v 1 and v 0 )
and their corresponding eigenvalues k (k2  k1  k0 ). Within this
approach v 0 approximates the point’s p normal, while v 2 and v 1
span the tangent plane at p.
In addition, k0 quantitatively describes the variation along the
point normal, providing an estimate of the local noise level (a.k.a
surface roughness) as a description of how much the k points devi-
ate from the fitted tangent plane. The smaller the value of k0 the
better the quality of the fitted plane. Curvature (normalized sur-
face variation) can be defined as [17]:
k0
rðpÞ ¼ k2 P k1 P k0 ð2Þ
k0 þ k1 þ k2
if rðpÞ ¼ 0, then all the points lie in the perfect plane.
The PCA approach is relatively accurate when the k-NN of a
point p is located on a smooth surface, but tends to smear across
singularities such as sharp features or noise due to the contribution
to the PCA covariance matrix, C, of neighboring points across the
singularity (Fig. 2). This limitation then results in inaccurate seg-
mentations at the region growing stage. The challenges of sharp
features and scanning noise are particularly significant for point
clouds of civil infrastructure systems, which tend to exhibit many
of these issues.
4 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
Fig. 2. Normal estimation challenges.
However given an accurate normal estimation, sharp features
will be preserved and segmented surfaces will correctly represent
the physical object geometry. The method developed here accom-
plishes this through the use of statistical outlier detection to detect
singularities and outliers prior to choosing an appropriate set of
inlier neighboring points for PCA. A new multivariate statistical
approach called Deterministic MM-estimator (DetMM) is used to
determine inlier points within each N pi . The DetMM approach is
robust in the presence of noise and geometric complexities, includ-
ing sharp features (Fig. 3).
3.2. DetMM-based Normal Estimation (DMNE)
The classic distance-based discordancy measure for indicating
outliers is the squared Mahalanobis distance [41]. The (squared)
Mahalanobis distance is similar to the Euclidean distance, except
that it normalizes the data on the basis of the inter-attribute
correlations.
D2i ¼ ðfxi g  flx gÞT ½R1 ðfxi g  flx gÞ i ¼ 1; . . . ; n
T 1
A multivariate matrix ½X ¼ ðx1 ; . . . ; xn Þ describes an m-
dimensional local point neighborhood of size n, where
Fig. 3. Comparison of estimated normals on the CUBE dataset.
th
xi ¼ ðxi1 ; . . . ; xim ÞT is the i observation within that neighborhood.
The observation fxi g is thus a potential outlier under investigation,
P
where flx g ¼ 1n ni¼1 xi is the arithmetic mean of the sample obser-
P
vations, and R ¼ ni¼1 ðxi 
xÞðxi 
xÞT =ðn  1Þ is the sample covari-
ance matrix.
This approach works well if only a single outlier is present, but
suffers from a masking effect otherwise, because one far-away out-
lier can reduce D2i for other outliers [42]. This effect can be mea-
sured by a metric known as the breakdown point. The breakdown
point is the smallest fraction of outliers that may cause an arbitrar-
ily large effect on the estimation of the regression coefficients. For
example, the breakdown point for the Mahalanobis distance is only
2/n, which makes it extremely sensitive to outliers.
The problem of masking effect can be resolved by using robust
estimates of mean and covariance matrix that are less affected by
outliers. In the DMNE approach (Fig. 4), a variant of the squared
Mahalanobis distance is implemented to increase the breakdown
point of the distance estimate. Here, the deterministic multivariate
MM-estimator (DetMM) is used to compute the mean vector ( l b)
ð3Þ b
and covariance matrix ( R ):
D2i ¼ ðfpi g  f l
T b  ðfp g  f l
b gÞ ½ R b gÞ 6 v2m;0:975 i ¼ 1; . . . ; n ð4Þ
i
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16 5

Fig. 4. The DetMM-based normal estimation (DMNE) process.

1=2
As suggested by Rousseeuw and Leroy [42], hard-rejection 1X n
ððx  mÞT C1 ðxi  mÞÞ
weights are utilized by omitting any point whose estimated q1 ð i Þ ð7Þ
n i¼1 r~
squared Mahalanobis distance is larger than the 97.50 percentile
of the chi-square (v2 ) distribution with m = 3 degrees of freedom Over ðm; CÞ where m 2 R3 and C is a PDS matrix with jCj ¼ 1.
(v23;0:975 ¼ 3:075). The minimization is performed through iterative least-squares,
yielding ðl ^ Þ. The resulting DetMM estimator of mean and
^; C
covariance matrix are l ^ ¼r
^ and R ^ , respectively.
~ 2C
3.2.1. Deterministic MM-estimator (DetMM)
Similar to the MM-estimators of regression introduced by Yohai
[43], MM-estimators of multivariate location and shape [44] are 3.2.2. Robust normal estimation
based on two loss functions (q0 and q1 ). The following regularity Upon finding inliers for the local k sized neighborhood N pi of a
conditions are required for a loss function q: point pi , PCA then is applied to the inlier points in order to deter-
mine the best fit tangent plane for the cleaned neighborhood, as
(i) q is real, symmetric, twice continuously differentiable, and described previously (Fig. 4).
qð0Þ ¼ 0: Since the orientation of the estimated normals are ambiguous, a
(ii) q is strictly increasing on ½0; c for some c > 0 on ½c; 1½ for globally consistent orientation must be computed in a separate
some finite constant c. step. This can be accomplished through the use of a minimum
spanning tree (MST) approach to reorient raw normals using the
The loss function q can be defined as: term 1  jhni ; nj ij as the uncertainty measure for this
propagation-based approach, where ni;j are the two normals of
( adjacent points [13]. The key idea in this approach is to consider
x2 4 6
 2cx 2 þ 6cx 4 for jxj 6 c
qðxÞ ¼ 2
c2
ð5Þ that two data points pi and pj belonging to a smooth surface and
6
for jxj > c geometrically close, need to have their normals consistently ori-
ented (i.e. ni  nj 1). Although this MST-based approach works
Multivariate MM-estimators of the location, shape, and covari-
well for smooth and densely sampled surfaces with low curvature,
ance of the point neighborhood ½X ¼ ðx1 ; . . . ; xn ÞT  Rp it can fail at sharp edges and creases. In order to consistently exe-
(p-dimensional) can be robustly estimated by initially utilizing cute the re-orientation of the calculated normals in large noisy 3D
S-estimators [45]. An S-estimator of location and scatter is defined point clouds, the adaptation presented by Schertler et al. [47] was
as the couple ðl ~ Þ which minimizes jSj under the condition:
~; R implemented due to its performance in the presence of noise and
sharp features.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X n Fig. 3 illustrates the benefits of using outlier invariant normal
q ðxi  mÞT S1 ðxi  mÞ ¼ b ð6Þ estimation in comparison with the conventional PCA approach,
n i¼1 0
for a synthetically generated 3D point cloud (a perfect cube) with
Among all ðm; SÞ where m 2 R3 and S is a 3  3 positive definite sharp features. Unlike the conventional PCA approach, which does
symmetric (PDS) matrix with b ¼ 1n ½npþ1 . Recently, Hubert et al. not reject outliers, the DetMM normal estimation method is able to
2
reliably estimate point normals near the planar transitions of the
[46] proposed a deterministic S-estimator (DetS) that uses a
cube.
resampling strategy to improve the efficiency of the algorithm,
Due to the more robust estimation of point normals using
and their approach is used here.
DMNE, the values of the surface variation, rðpÞ, are more consis-
Deterministic MM-estimators (DetMM) [46] can then be
tent even for non-planar smooth surfaces (Fig. 5), when compared
derived by utilizing the DetS approach as the initial estimator for
with conventional PCA. Fig. 6 illustrates the benefits of the pre-
a multivariate MM-estimator, defined in two steps:
sented method. The boxes enclose the middle half of the results,
  i.e. the length of a box is the interquartile range with the ends of
(i) Let ðl; RÞ be a DetS estimator with loss function q0 . Denote the box at the lower/1st and upper/3rd quartiles. The line across
~ j1=ð2pÞ
r~ ¼ jR each box is the position of the median, and the ‘+’ markers repre-
(ii) The DetMM-estimator for location and shape ðl ^Þ
^; C sent outlying results. The box length shows the variability of the
minimizes: data.
6 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
Fig. 5. Comparison of estimated normalized surface variation (curvature) for a point cloud of a pipe.
Fig. 6. Boxplot of curvature values for the pipe model.
3.3. Alternative methods used for comparison
Alternative estimators of multivariate location and scatter
include: S-estimators [45], s-estimator [48], Minimum Volume
Ellipsoid (MVE) [49] and Minimum Covariance Determinant
(MCD) [42] estimator. They are all highly robust estimators with
breakdown points up to 50%, but they differ in terms of efficiency,
bias and outlier resistance. For detailed comparison of these
approaches, the reader is referred to [50]. In this work, the follow-
ing alternative methods were assessed relative to the utilized
DetMM approach.
3.3.1. Minimum Volume Ellipsoid (MVE) and Minimum Covariance
Determinant (MCD)
One of the earliest affine equivariant estimators to achieve a
breakdown point of approximately 50% is the Minimum Volume
Ellipsoid (MVE) estimator. The MVE estimator [49] of location
(mean) is defined as the center of the ellipsoid with minimal vol-
ume containing h out of m points (any n dimensional) with
½m=2 6 h 6 m. The corresponding scatter estimator (covariance
matrix) is given by the formula for this ellipsoid, multiplied by a
qffiffiffiffiffiffiffiffi
consistency factor c. Where c ¼ v2n;a with a ¼ h=m. According to
its definition, MVE estimates the mean and covariance of the h
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
most concentrated points in the dataset. This estimator has maxi-
mal breakdown value when h ¼ ½ðm þ n þ 1Þ=2 but it is not
asymptotically normal and more difficult to compute than the
MCD estimator [42].
An alternative to the MVE estimator that also has a breakdown
point of approximately 50% is the minimum covariance determi-
nant (MCD) estimator [42]. Rather than search for the subset of
half the data that has the smallest volume, MCD searches for the
half h > m/2 whose classical covariance matrix has the lowest pos-
sible determinant. Within the MCD approach, l is the mean of the
h observations, ½ðm þ n þ 1Þ=2 6 h 6 m, for which the determinant
of the sample covariance matrix is minimal and the scatter matrix
R is the corresponding covariance matrix multiplied by a consis-
tency factor in the case of multivariate normality. Generally, the
MCD estimates of location and scatter are affine equivariant which
means that they are also invariant under affine transformation
behavior. It should be noticed that the MCD estimator can only
be computed when h > n, otherwise the covariance matrix of any
h-subset will be singular. Since h P ½ðm þ 2Þ=2, this condition is
certainly satisfied when m  2n.
Although MCD is easier to compute in comparison with MVE, it
is still impractical for very large data sets since it requires the eval-
 
m image data under the surface, higher noise levels were achieved
uation of subset of size h. In order to make this approach
h at these locations which later resulted in over-segmentation and
scalable and suitable for large 3D point cloud datasets, an efficient
version of MCD with fast resampling called FAST-MCD algorithm
[51] was implemented in this study.
3.3.2. Maximum Likelihood SAmple Consensus (MLESAC)
RANSAC is a robust approach to classify data into inliers and
outliers which is implemented in many applications for extracting
shapes and estimating the model parameters from data that con-
tain noise/outliers [27,19]. Compared to RANSAC, MLESAC algo-
rithm adopts the same sampling strategy but chooses the
solution by minimizing the probabilistic loss rather than the num-
ber of inliers. Minimizing probabilistic loss renders the model ver-
ification threshold value invariant to the cardinality of the models
dataset. MLESAC is reported to be of higher accuracy and robust-
ness than RANSAC [52], which makes it more suitable for local
plane fitting. In this approach, first a minimal subset, e.g. three
points for a plane, is randomly selected and the required plane
parameters are then computed based on this subset. Next, MLESAC
evaluates the fitness of the hypothesized parameters using a prob-
abilistic model for the errors arising from noise contaminated data.
As discussed in [53,32], MLESAC measures the likelihood of the
generated model hypothesis by representing the error distribution
as a mixture model. The suggested probability distribution func-
tion (pdf) for inlier error is unbiased Gaussian distribution and
the pdf of outlier error is uniform distribution. Thus the probability
of error described as follows:
     
1 e2 1
PrðejModelÞ ¼ c pffiffiffiffiffiffiffiffiffiffiffi2ffi exp  2 þ ð1  cÞ ð8Þ
2pr 2r m To address this, median and Q n estimator [55] are implemented
where r is the standard deviation of Gaussian noise, e is the inlier
error, c is the prior probability of being an inlier and m is the size
of the available error space. When the PrðejModelÞ is larger than a
predefined threshold, then the model will be reestimated using only
inliers and MLESAC terminates.
3.4. Region growing segmentation
Once the point normals and curvatures are estimated, the
region growing segmentation starts by finding a seed point pi that
is then grown by incrementally adding new points to it using the
region growing test criteria. When growth is complete, a new seed
7
point is needed for the growth of the next segment. In this study,
the point pi with the minimum curvature value, rðpÞ, was consid-
ered as the initial seed to start the region growing process from a
part in the scene where the surface is smoother and, hence, the
surface variation is lower.
3.4.1. Spatial connectivity analysis in segment growth
While it has been used effectively in previous work, the
smoothness constraint based region growing method of Rabbani
et al. [34] suffers from several shortcomings. Firstly, due to the
PCA based normal estimation strategy implemented within, it is
inherently susceptible to noisy outliers. Secondly, this algorithm
requires two parameters/thresholds (for both curvature rðpÞ and
normal vector angular distance hij values). As suggested in
[36,34,35], choosing the curvature threshold as the 98th percentile
of the sorted normalized surface variation makes the segmentation
results data dependent. Fig. 7 illustrates the segmentation of a 3D
table model using this method [34]. Smooth areas are broken by
using solely the angular distance criterion between two adjacent
points. On the surface of the object due to the sensitive nature of
the DSfM process to environmental factors (e.g., humidity, temper-
ature and lighting) and lack of adequate lighting while collecting
failing of the curvature threshold.
Inspired by [54], and in order to overcome this difficulty, the
segmentation algorithm developed herein uses a single user-
defined parameter and two automatically calculated locally adap-
tive thresholds to account for variations in the local geometrical
features of each point.
th
Three distance criteria are defined between the i seed point pi
and the k-sized neighborhood, N pi , and used to add points to the
growing region. The test criteria are: (i) Coplanar Distance (CD),
(ii) Euclidean Distance (ED) and (iii) Angular Distance (h).
Point-to-plane coplanar distance (CD) measures the distance for
every neighboring point N pi of the pi to its best-fitted tangent local
plane using:

Þ T :n
CDðpi Þ ¼ ðpi  p ^ ð9Þ
where p
and n b are the local data centroid (mean) and the unit nor-
mal of the locally fitted plane, respectively.
In order to only add salient points during region growth, it is
important to define a robust threshold to identify outliers. As pre-
viously used for noise/outlier detection in point clouds [36,9], it is
possible to utilize mean (l) and standard deviation (r) in the form
of l þ 2r to distinguish inliers from outliers. Although mean is the
classic estimation of location, it is very sensitive to even one aber-
rant value out of the n data points which makes its breakdown
value equal to 1=n 0% for large n. In addition, standard deviation
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi
r¼
2
i¼1 ðxi  xÞ =ðn  1Þ also has a breakdown point of 0% since
a single outlier can make r arbitrarily large.
instead of mean and standard deviation to make the threshold for
the coplanar distance (CDth ) robust in the presence of noise. In con-
trast with mean, median has a breakdown value of about 50%,
meaning that it can resists up to 50% of outliers. Q n estimator is
a robust measure of scale which is defined as the first quartile of
all pairwise differences between two data points:
Q n ¼ 2:2219fjxi  xj j; i < jgðkÞ ð10Þ
   
h n
with k ¼ 4 and h ¼ bn=2c þ 1. The constant 2.2219 is
2 2
a correction factor that makes the estimator unbiased. Q n estimator
has a breakdown value of 50% and complexity of Oðn log nÞ.
8 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
Fig. 7. Estimated curvature values and segmentation of the Table model using Rabbani et al. [34] method with 98th percentile curvature values.
As discussed previously in the work by Li et al. [19], noise
within 3D point cloud data obeys Gaussian distribution. In compar-
ison with other robust scale estimators such as median absolute
deviation (MAD) and Sn (which has Gaussian efficiency close to
37% and 58%, respectively), Q n estimator has a high efficiency of
82% at Gaussian distributions [55] which makes it highly suitable
th
for point cloud analysis. Hence, the i point is considered a copla-
nar surface point if:
CDi < CDth ¼ medianfCDðNpi Þg þ 2  Q n fCDðNpi Þg ð11Þ
where fCDðNpi Þg is the set of CDs for all points in Npi .
The data points that have the closest distance to each other are
most likely to be within the same segment on the 3D point cloud.
Point-to-point Euclidean Distance (ED) computes the distance
between a pair of points (the seed point pi and one of its neighbors
pj in N pi ) as:
EDij ¼ kpi  pj k ð12Þ
Consequently, pj is considered to be spatially connected to pi if
in comparison with the majority of the points in N pi they are rela-
tively close:
EDij < EDth ¼ medianfEDij g ð13Þ
where EDth is the ED threshold and fEDij g is the set of all EDij s.
th
Finally, the angular distance is defined between the i seed
point pi and one of its neighbors pj as:
^ Ti :n
hij ¼ arccos jn ^j j ð14Þ
where n ^ i and n
^ j are the unit normals for points pi and pj , respec-
tively. As the direction of normal vectors has a 180 ambiguity, here
the absolute value of the dot product was used. This distance mea-
sure (h) plays a pivotal role in segmenting 3D point clouds at inter-
section edges were two surface patches meet. Two spatially close
points are co-plane if h is less than the user defined threshold hth .
Due to the robustness of the estimated normal vectors near sharp
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
features and surface discontinues through using the DetMM based
method (Fig. 3), the presented region growing segmentation is more
efficient in recovering sharp features in the presence of bilinear sur-
faces. The hth threshold is the only user-dependent parameter (as
CDth and EDth are locally adaptive and automatically derived)
within the presented algorithm, and can be intuitively selected
based on the expected geometric complexity of a scene with respect
to the presence of sharp planar transitions.
Fig. 8 shows an example segmentation result and curvature val-
ues based on the presented method. In contrast with the result
achieved using the method in [34] (Fig. 7), a much more accurate
segmentation with improved accuracy in regions with a relatively
higher noise level is apparent. This illustrates the robustness of the
algorithm in segmenting 3D datasets consisting of planar and non-
planar complex object surfaces while preserving the overall geom-
etry of the reconstructed physical object.
The complete region growing segmentation (Fig. 1) algorithm is
provided as pseudocode for repeatability (Algorithm 1).
Algorithm 1. The complete segmentation method with the
embedded DMNE process
4. Performance analysis
Both the DetMM-based normal estimation process (DMNE) and
the overall segmentation algorithm were evaluated. In this study
the implementation of the entire code was done in MATLABÒ, run-
ning on a MicrosoftÓWindows computer with a 3.4 GHz Intel i7-
3770 processor and 32 GB RAM.
4.1. DMNE evaluation
In order to assess the performance of the DMNE process, the fol-
lowing metrics are used: (i) Variation along the point p normal,
defined as the least eigenvalue k0 , (ii) curvature or normalized sur-
face variation rðpÞ, (iii) bias angle h ¼ arccos jn^ Ti :n
^ j j, between the
best fitted local planes with and without added noise/outliers in
9
order to quantify normal estimation sensitivity and (iv) computa-
tional time.
The efficiency and robustness of the DMNE algorithm was
tested on point cloud datasets with various geometric features,
noise levels and curvature values. The S1 model (Fig. 9) contains
highly curved pipes and support structures from a laser scanned
industrial facility scene with varying local point density, and con-
sists of 195,057 points. This model is from a publicly available
dataset by RAAMAC lab at the University of Illinois at Urbana
Champaign. The other dataset ‘‘Table” (154,100 points) was gener-
ated by using a hierarchical adaptation of the DSfM method [56],
and consists of both planar and non-planar complex objects with
variations in surface roughness (Fig. 9).
The average local point cloud density values for each of these
datasets are 50 points/m3 and 28 points/m3 , respectively. It is
important to note that the k sized neighborhood for each point p
should be large enough to negate noise while still small enough
to accurately estimate local properties. While greater neighbor-
hoods certainly help suppress noise in the sense that they smooth
out the estimated surface, they do not necessarily yield a better
surface normal estimate. In this paper k = 50 was used to fit a plane
to a local neighborhood at each 3D point.
Tables 1–3 shows the results of normal estimation and outlier
detection using a variety of estimators. The comparative analysis
demonstrates that, by finding and discarding outliers within the
local neighborhood of each point, it is possible to more accurately
estimate normal vectors for noisy point clouds by utilizing the
DMNE method. A disadvantage of utilizing MCD and MVE as mul-
tivariate robust estimators to detect and discard outlier points
within each local neighborhood, as illustrated in Fig. 10, is that
they tend to declare a larger number of outliers when the sample
size is small. This is because they make substantial use of the
asymptotic v2 approximation to the distribution of robust dis-
tances which may result in inaccurate results even in moderately
large datasets [42].
In addition, in this comparative study the best tangent local
plane was fitted for every point before and after contaminating
datasets by adding synthetic Gaussian noise with zero mean and
0.2 variance. The approach presented herein performs better (i.e.
smaller bias angles) than other methods in the presence of this
noise contamination (Table 1). Results in Table 3 show that the
DMNE algorithm is significantly faster than other methods in find-
ing the best-fit plane within each points neighborhood. This com-
putation can be performed in parallel using multiple threads to
take advantage of multiple CPU cores, although this was not done
in this work.
4.2. Evaluating segmentation accuracy
To determine the efficiency and robustness of the presented
point cloud segmentation algorithm, it was tested against the
smoothness constraint region growing method proposed by [34]
which has been implemented in C++, and published in the PCL
[38]. To validate the performance of the point cloud segmentation
algorithms, a set of ground truth 3D data were manually generated
by human experts, who were familiar with the surface segmenta-
tion process, to serve as a reference dataset. These reference seg-
ments were extracted based on experts’ judgment of surface
continuity and not object continuity.
4.2.1. Performance metrics
There are a number of available methods in the literature to
evaluate the agreement between two sets of segmented data.
Herein, the metrics precision, recall and F-score were utilized to
measure the overall accuracy and effectiveness of a segmentation.
10 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
Fig. 8. Estimated curvature values and segmentation of the Table model using the presented method with hth ¼ 4 .
These metrics are based on the following values: (i) True Positives
(TP) represents proper segmentation that matches with a manually
generated ground truth benchmark by human experts; (ii) False
Positives (FP) denotes over-segmentation within point cloud mod-
els that results in segmenting one ‘reference’ segment into several
parts; (iii) False Negatives (FN) happen when two or more separate
segments are wrongly grouped together to under-segment the 3D
model.
Precision represents the capability of the algorithm in properly
segmenting 3D point clouds, as given by:
jTPj
Precision ¼  100
jTPj þ jFPj
The recall metric allows to measure the ability of the 3D seg-
mentation algorithms to correctly retrieve ‘reference’ segments
according to the ground truth data:
jTPj
Recall ¼  100
jTPj þ jFNj
Finally, the F-score is the harmonic mean of precision and recall
values, summarized in a single factor:
Precision  Recall
F-score ¼ 2  ð17Þ
Precision þ Recall
Due to the sensitivity of the precision and recall metrics to the
existence of spurious elements and ‘reference’ data that are not
recognized by the algorithm, respectively, the F-score represents
the best metric to evaluate the robustness and effectiveness of
the segmentation process.
ð15Þ Since manually comparing extracted segments of point clouds
against a ground truth dataset of large infrastructure 3D models
is not a straightforward task, the automated object-based evalua-
tion method proposed by Awrangjeb and Fraser [57] was imple-
mented in this study. This method automates the process of
making one-to-one correspondences between extracted and refer-
ence segments of point clouds using the maximum overlaps. First,
ð16Þ
it establishes a pair of correspondences between extracted seg-
ments, S ¼ fS1 ; S2 ; . . . ; Si g, and ground truth segments,
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16 11

G ¼ fG1 ; G2 ; . . . ; Gj g. In this step, ðSi ; Gj Þ is considered a proper one-

to-one correspondence if each entity Si provides the maximum
overlap with Gj among all extracted parts and vice versa.

4.2.2. Experimental validations and discussions

A series of segmentation experiments were designed and run on
several complex 3D point clouds with different geometric charac-
teristics to evaluate the performance of the introduced algorithm
in comparison with the PCL implementation of [34], referred to
as PCL-RG (hth ; rth ). hth and rth are the angle threshold (in degrees)
and curvature threshold, respectively. Here, the rth were calculated
automatically using a specified percentile of the sorted curvature
values as suggested in [34–36].
In addition to the S1 and Table point clouds used for DMNE
evaluation, two corridor scenes (models T1 and T2), each contain-
ing densely packed plumbing equipment and featuring a signifi-
cant amount of occlusion and clutter, were used. T1 (554,150
points, density of 1230 points/m3 ) and T2 (1,360,945 points, den-
sity of 5768 points/m3 ) models, along with the S1 model were
selected to evaluate the segmentation algorithms. Fig. 11 illus-
trates the segmentation results on the T2 model. Due to occlusions
and the scan registration process, a large range of surface rough-
ness is apparent within this model which results in variability in
the normalized noise level. As shown in Fig. 11, there is no clear
set of input parameters for the PCL-RG method that can be deemed
best. In case of scenes containing variable levels of curvature,
where multiple objects are in close proximity, finding the proper
set of parameters will inevitably be a challenge. According to
Fig. 11, by setting the rth as the 95th percentile of the curvature
values, a noticeable amount of over-segmentation occurred in loca-
tions with excessive noise. Increasing the aforementioned user-
defined parameter to 98th percentile, resulted in an inappropriate
under-segmentation by falsely including points from different sur-
faces as one. The results from the presented algorithm reveal that
by incorporating DetMM-based normals and locally adaptive
thresholds using the spatial connectivity analysis, it was possible
to properly extract surfaces with various curvature values and even

Table 2
Comparison of accuracy in tangent plane fitting based on different methods.

Data set Methods Mean curvature rðpÞ Mean noise level k0 (mm)
S1 PCA 0.0128 3.7
MVE 0.0022 0.8
MCD 0.0027 1.1
MLESAC 0.0041 1.2
DetMM 0.0015 0.6

Fig. 9. 3D point cloud datasets. Table PCA 0.0615 21.1

MVE 0.0195 7.01
MCD 0.0251 9.3
MLESAC 0.2915 10.14
DetMM 0.0165 4.6

Table 1 Table 3
Bias angles h (in degrees) for different methods. Comparison of runtime in tangent plane fitting based on different methods.

Data set Methods Mean Median Data set Number of points Methods Execution time (minutes)
S1 PCA 21.38 12.88 S1 198,217 PCA 0.11
MVE 11.62 9.12 MVE 3187.21
MCD 12.71 12.26 MCD 1487.67
MLESAC 14.13 12.18 MLESAC 294.65
DetMM 7.28 5.61 DetMM 56.77
Table PCA 18.74 10.81 Table 154,100 PCA 0.09
MVE 8.42 7.35 MVE 2645.43
MCD 11.18 7.73 MCD 1207.25
MLESAC 15.83 9.45 MLESAC 274.47
DetMM 3.14 2.18 DetMM 40.12
12 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
Fig. 10. Index plot of robust Mahalanobis distances calculated using (a) MVE, (b) MCD and (c) DetMM estimators.
close parallel and adjacent surfaces of different sizes using a single
parameter segmentation.
Figs. 12 and 13 illustrate that the new method is capable of
cleanly separating objects in close proximity in both T1 and S1
models. The higher accuracy and robustness of the developed
method improves the segmentation quality in smooth regions
while preserving sharp features around corners and close to other
objects. The over-smoothing of point normals near or on sharp cor-
ners with the PCA-based approach limits the performance of the
conventional region growing method in recovering points on these
locations.
Since the region growing method implemented in the PCL was
utilized in this comparative study, the suggested point cloud
posted on [38] for benchmarking purposes was used to test the
new method. This synthetic 3D point cloud (Fig. 14) has 154,100
points and a local point density of 28 points/m3 and contains both
sharp and smooth transitions between mostly planar surfaces. As
illustrated in Fig. 14, PCL-RG had the tendency of splitting a
smooth patch into many smaller segments where even small gaps
appeared (e.g. close-up views in Fig. 14) in the final segmented
model. In comparison, the developed algorithm successfully
extracted closely connected planes with a variety of angular dis-
tances between them due to the less noticeable gaps between seg-
mented surfaces.
Table 4 presents the quantitative assessment of different seg-
mentation algorithms. Overall, the introduced method yielded con-
sistently better results for precision (i.e. correctness), recall (i.e.
completeness) and F1-score (i.e. overall accuracy) in all 3D data-
sets. Minimizing user defined parameters and considering the geo-
metric properties of the 3D point cloud data resulted in significant
improvement in adequately segmenting multiple objects in close
proximity. The overall lower F1-score values (Table 4) for the T1
and T2 models is due to the relatively higher rate of occlusions
and missing data in these models.
The segmentation results also show the capability and robust-
ness of the presented algorithm in dealing with a large range of
local point density and surface variation, a condition prevalent
for many point clouds of large infrastructure systems. Notably, as
shown in Figs. 8 and 11, the PCL-RG algorithm displayed poor per-
formance in over- or under-segmentation of point clouds with
varying surface roughness due to considering curvature (rth ) as
one of the user-defined values.
5. Conclusions
This paper presents a new and statistically robust method for
reliable normal vector estimation in noisy 3D point clouds with
non-uniform point distribution and high variance in the surface
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16 13

Fig. 11. Segmentation of a corridor scene (model T2) with various surface roughness.

Fig. 12. Segmentation of a corridor scene (model T1) with various surface roughness.

roughness and geometric complexity, referred to as DMNE. DMNE
incorporates a deterministic and efficient variant of a high break-
down MM-type multivariate location and scatter estimator to
improve the robustness of the Mahalanobis distance-based tech-
nique in detecting and removing outliers in the neighborhood of
each point of interest, resulting in improved normal estimation
near geometric discontinuities by incorporating only neighboring
points positioned on the same smooth surface in normal
computation.
Experiments based on several complex 3D point cloud datasets
generated using laser scanning and photogrammetric approaches
show that the DMNE method outperforms classical PCA as well
as more advanced methods such as MVE, MCD and MLESAC. Com-
pared with conventional PCA-based normal estimation, DMNE
14 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16

Fig. 13. Segmentation of the S1 model.

Fig. 14. Segmentation of the PCL dataset.

Table 4
Comparison of the conventional region growing against the presented segmentation method.

Models No. of ground truth segments Methods Metrics Total No. of segments
Precision (%) Recall (%) F1-score (%)
S1 13 PCL-RG (15°,95th) 100.0 57.1 72.7 7
PCL-RG (35°,98th) 100.0 25.0 40.0 4
Presented Method (10°) 92.3 100.0 96.0 13
T1 40 PCL-RG (15°,95th) 25.3 67.9 36.9 84
PCL-RG (35°,98th) 35.7 42.9 39.0 62
Presented Method (10°) 70.5 88.6 78.5 48
T2 34 PCL-RG (15°,95th) 3.3 100.0 6.4 392
PCL-RG (35°,98th) 14.1 95.0 24.5 136
Presented Method (10°) 38.9 100.0 56.0 54
PCL dataset 85 PCL-RG (3°,95th) 55.9 100.0 71.1 136
PCL-RG (5°,98th) 100.0 49.3 66.0 72
Presented Method (4°) 78.4 100.0 87.9 98
Table 49 PCL-RG (3°,95th) 51.3 100.0 67.8 78
PCL-RG (5°,98th) 48.1 64.1 54.9 66
Presented Method (4°) 90.2 100.0 94.8 51
 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
requires more computational time, though computation time is not
often a critical issue for point cloud segmentation.
In addition, a new region growing algorithm for the segmenta-
tion of unordered point clouds was also introduced. This approach
employs the robust normal vectors and other salient local features
estimated using the DMNE process. Due to the higher accuracy in
local tangent plane fitting and higher distinctiveness of point nor-
mals, the segmentation algorithm needs only a single user-defined
parameter to control the segmentation procedure. This is an
improvement over conventional methods which suffer from multi-
ple ambiguous input parameters to provide a trade-off between
under- and over-segmentation.
By utilizing a locally adaptive spatial connectivity analysis
within each segment growth step, it is possible to reach an appro-
priate segmentation solely based on geometric features. Conse-
quently, the presented method is not tailored for a specific 3D
dataset or context and has been validated on a range of validation
models to illustrate the flexibility for a variety of infrastructure
applications.
Future research can focus on optimizing and performing the
entire normal estimation and point cloud segmentation through
parallel processing schemes utilizing the Graphics Processing Unit
(GPU) to increase the performance for massive 3D models.
Acknowledgments
This material is based upon the work supported by the National
Science Foundation (NSF) under Grant No. CMMI-1433765. The
authors would also like to thank Prof. Mani Golparvar-Fard from
University of Illinois at Urbana-Champaign for graciously provid-
ing S1, T1 and T2 point cloud datasets and their corresponding
ground truth segmentation used in this research. Any opinions,
findings, and conclusions or recommendations expressed in this
publication are those of the authors and do not necessarily reflect
the views of the NSF.
References
[1] R. Szeliski, Computer Vision: Algorithms and Applications, Springer Science &
Business Media, 2010.
[2] M. Cabaleiro, B. Riveiro, P. Arias, J. Caamaño, Algorithm for beam deformation
modeling from lidar data, Measurement 76 (2015) 20–31.
[3] B. Jafari, A. Khaloo, D. Lattanzi, Tracking structural deformations via automated
sample-based point cloud analysis, in: International Workshop on Computing
in Civil Engineering 2017, ASCE, 2017.
[4] H. Son, C. Kim, C. Kim, 3d reconstruction of as-built industrial instrumentation
models from laser-scan data and a 3d CAD database based on prior knowledge,
Automat. Construct. 49 (Part B) (2015) 193–200, http://dx.doi.org/10.1016/j.
autcon.2014.08.007.
[5] A. Dimitrov, R. Gu, M. Golparvar-Fard, Non-uniform b-spline surface fitting
from unordered 3d point clouds for as-built modeling, Comput.-Aided Civil
Infrastruct. Eng. 31 (7) (2016) 483–498, http://dx.doi.org/10.1111/mice.12192.
[6] J. Jung, S. Hong, S. Yoon, J. Kim, J. Heo, Automated 3d wireframe modeling of
indoor structures from point clouds using constrained least-squares
adjustment for as-built bim, J. Comput. Civil Eng. (2015) 04015074.
[7] T. Gao, S. Ergan, B. Akinci, J. Garrett, Evaluation of different features for
matching point clouds to building information models, J. Comput. Civil Eng. 30
(1) (2014) 04014107.
[8] V. Pătrăucean, I. Armeni, M. Nahangi, J. Yeung, I. Brilakis, C. Haas, State of
research in automatic as-built modelling, Adv. Eng. Inform. 29 (2) (2015) 162–
171.
[9] A. Dimitrov, M. Golparvar-Fard, Segmentation of building point cloud models
including detailed architectural/structural features and mep systems,
Automat. Construct. 51 (2015) 32–45.
[10] H. Son, F. Bosché, C. Kim, As-built data acquisition and its use in production
monitoring and automated layout of civil infrastructure: a survey, Adv. Eng.
Inform. 29 (2) (2015) 172–183.
[11] A. Boulch, R. Marlet, Fast and robust normal estimation for point clouds with
sharp features, Comput. Graph. Forum 31 (5) (2012) 1765–1774, http://dx.doi.
org/10.1111/j.1467-8659.2012.03181.x.
[12] T.K. Dey, G. Li, J. Sun, Normal estimation for point clouds: a comparison study
for a voronoi based method, in: Point-Based Graphics, 2005. Eurographics/IEEE
VGTC Symposium Proceedings, IEEE, 2005, pp. 39–46.
15
[13] H. Hoppe, T. DeRose, T. Duchamp, J. McDonald, W. Stuetzle, Surface
reconstruction from unorganized points, in: Proceedings of the 19th Annual
Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’92,
ACM, New York, NY, USA, 1992, pp. 71–78, http://dx.doi.org/10.1145/
133994.134011.
[14] I. Jolliffe, Principal component analysis, in: Wiley StatsRef: Statistics Reference
Online, John Wiley & Sons, Ltd, 2002.
[15] N.J. Mitra, A. Nguyen, L. Guibas, Estimating surface normals in noisy point
cloud data, Int. J. Comput. Geomet. Appl. 14 (04n05) (2004) 261–276, http://
dx.doi.org/10.1142/S0218195904001470.
[16] Y. Wang, H.-Y. Feng, F.-t. Delorme, S. Engin, An adaptive normal
estimation method for scanned point clouds with sharp features,
Comput. Aided Des. 45 (11) (2013) 1333–1348, http://dx.doi.org/10.1016/
j.cad.2013.06.003.
[17] M. Pauly, M. Gross, L.P. Kobbelt, Efficient simplification of point-sampled
surfaces, in: Proceedings of the Conference on Visualization ’02, VIS ’02, IEEE
Computer Society, Washington, DC, USA, 2002, pp. 163–170.
[18] E. Castillo, J. Liang, H. Zhao, Point Cloud segmentation and denoising via
constrained nonlinear least squares normal estimates, in: M. Breu, A.
Bruckstein, P. Maragos (Eds.), Innovations for Shape Analysis, Mathematics
and Visualization, Springer, Berlin, Heidelberg, 2013, pp. 283–299.
[19] B. Li, R. Schnabel, R. Klein, Z. Cheng, G. Dang, S. Jin, Robust normal estimation
for point clouds with sharp features, Comput. Graph. 34 (2) (2010) 94–106,
http://dx.doi.org/10.1016/j.cag.2010.01.004.
[20] M.A. Fischler, R.C. Bolles, Random sample consensus: a paradigm for model
fitting with applications to image analysis and automated cartography,
Commun. ACM 24 (6) (1981) 381–395.
[21] K. Klasing, D. Althoff, D. Wollherr, M. Buss, Comparison of surface normal
estimation methods for range sensing applications, in: IEEE International
Conference on Robotics and Automation, 2009. ICRA ’09, 2009, pp. 3206–3211,
http://dx.doi.org/10.1109/ROBOT.2009.5152493.
[22] T. Song, F. Xi, S. Guo, Z. Ming, Y. Lin, A comparison study of algorithms for
surface normal determination based on point cloud data, Precis. Eng. 39 (2015)
47–55, http://dx.doi.org/10.1016/j.precisioneng.2014.07.005.
[23] M.M. Torok, M. Golparvar-Fard, K.B. Kochersberger, Image-based automated
3d crack detection for post-disaster building assessment, J. Comput. Civil Eng.
28 (5) (2013) A4014004.
[24] D. Lague, N. Brodu, J. Leroux, Accurate 3d comparison of complex topography
with terrestrial laser scanner: application to the rangitikei canyon (nz), ISPRS J.
Photogramm. Remote Sens. 82 (2013) 10–26.
[25] A.K. Patil, P. Holi, S.K. Lee, Y.H. Chai, An adaptive approach for the
reconstruction and modeling of as-built 3d pipelines from point clouds,
Automat. Construct 75 (2017) 65–78.
[26] T. Czerniawski, M. Nahangi, C. Haas, S. Walbridge, Pipe spool recognition in
cluttered point clouds using a curvature-based shape descriptor, Automat.
Construct. 71 (2016) 346–358.
[27] R. Schnabel, R. Wahl, R. Klein, Efficient RANSAC for point-cloud shape
detection, Comput. Graph. Forum 26 (2) (2007) 214–226, http://dx.doi.org/
10.1111/j.1467-8659.2007.01016.x.
[28] S. Filin, Surface clustering from airborne laser scanning data, Int. Arch.
Photogramm. Remote Sens. Spatial Inf. Sci. 34 (3/A) (2002) 119–124.
[29] P.J. Besl, R.C. Jain, Segmentation through variable-order surface fitting, IEEE
Trans. Pattern Anal. Mach. Intell. 10 (2) (1988) 167–192.
[30] F. Tarsha-Kurdi, T. Landes, P. Grussenmeyer, et al., Hough-transform and
extended ransac algorithms for automatic detection of 3d building roof planes
from lidar data, Proceedings of the ISPRS Workshop on Laser Scanning, vol. 36,
2007, pp. 407–412.
[31] M.F. Ahmed, C.T. Haas, R. Haas, Automatic detection of cylindrical objects in
built facilities, J. Comput. Civil Eng. 28 (3) (2014) 04014009.
[32] G. Zhang, P.A. Vela, P. Karasev, I. Brilakis, A sparsity-inducing optimization-
based algorithm for planar patches extraction from noisy point-cloud data,
Comput.-Aided Civil Infrastruct. Eng. 30 (2) (2015) 85–102.
[33] J.M. Biosca, J.L. Lerma, Unsupervised robust planar segmentation of terrestrial
laser scanner point clouds based on fuzzy clustering methods, ISPRS J.
Photogramm. Remote Sens. 63 (1) (2008) 84–98.
[34] T. Rabbani, F. Van Den Heuvel, G. Vosselmann, Segmentation of point clouds
using smoothness constraint, Int. Arch. Photogramm. Remote Sens. Spatial Inf.
Sci. 36 (5) (2006) 248–253.
[35] H. Son, C. Kim, C. Kim, Fully automated as-built 3d pipeline extraction method
from laser-scanned data based on curvature computation, J. Comput. Civil Eng.
29 (4) (2014) B4014003.
[36] S.B. Walsh, D.J. Borello, B. Guldur, J.F. Hajjar, Data processing of point clouds
for object detection for structural engineering applications, Comput.-Aided
Civil Infrastruct. Eng. 28 (7) (2013) 495–508, http://dx.doi.org/10.1111/
mice.12016.
[37] M. Yoon, Y. Lee, S. Lee, I. Ivrissimtzis, H.-P. Seidel, Surface and normal
ensembles for surface reconstruction, Comput. Aided Des. 39 (5) (2007) 408–
420.
[38] PCL, Region growing segmentation, 2016. <http://pointclouds.
org/documentation/tutorials/region_growing_segmentation.php>.
[39] J.H. Friedman, J.L. Bentley, R.A. Finkel, An algorithm for finding best matches in
logarithmic expected time, ACM Trans. Math. Software (TOMS) 3 (3) (1977)
209–226.
[40] G.H. Golub, C. Reinsch, Singular value decomposition and least squares
solutions, Numer. Math. 14 (5) (1970) 403–420, http://dx.doi.org/10.1007/
BF02163027.
16 A. Khaloo, D. Lattanzi / Advanced Engineering Informatics 34 (2017) 1–16
[41] P.C. Mahalanobis, On the generalized distance in statistics, Proc. Natl. Inst. Sci.
(Calcutta) 2 (1936) 49–55.
[42] P.J. Rousseeuw, A.M. Leroy, Robust regression and outlier detection, vol. 589,
John Wiley & Sons, 2005.
[43] V.J. Yohai, High breakdown-point and high efficiency robust estimates for
regression, Ann. Stat. 15 (2) (1987) 642–656.
[44] K.S. Tatsuoka, D.E. Tyler, On the uniqueness of s-functionals and m-functionals
under nonelliptical distributions, Ann. Stat. 28 (4) (2000) 1219–1243.
[45] P. Rousseeuw, V. Yohai, Robust regression by means of s-estimators, in: Robust
and Nonlinear Time Series Analysis, Springer, 1984, pp. 256–272.
[46] M. Hubert, P. Rousseeuw, D. Vanpaemel, T. Verdonck, The DetS and DetMM
estimators for multivariate location and scatter, Comput. Stat. Data Anal. 81
(2015) 64–75, http://dx.doi.org/10.1016/j.csda.2014.07.013.
[47] N. Schertler, B. Savchynskyy, S. Gumhold, Towards globally optimal normal
orientations for large point clouds, Comput.Graph. Forum (2016), http://dx.
doi.org/10.1111/cgf.12795, n/a–n/a.
[48] H.P. Lopuhaä, Multivariate s-estimators for location and scatter, Can. J. Stat. 19
(3) (1991) 307–321.
[49] S. Van Aelst, P. Rousseeuw, Minimum volume ellipsoid, Wiley Interdiscip. Rev.:
Comput. Stat. 1 (1) (2009) 71–82, http://dx.doi.org/10.1002/wics.19.
[50] R.A. Maronna, V.J. Yohai, Robust and efficient estimation of multivariate
scatter and location, Comput. Stat. Data Anal. 109 (2017) 64–75.
[51] P.J. Rousseeuw, K.V. Driessen, A fast algorithm for the minimum covariance
determinant estimator, Technometrics 41 (3) (1999) 212–223, http://dx.doi.
org/10.1080/00401706.1999.10485670.
[52] S. Choi, T. Kim, W. Yu, Performance evaluation of ransac family, J. Comput.
Vision 24 (3) (1997) 271–300.
[53] P.H.S. Torr, A. Zisserman, MLESAC: a new robust estimator with application to
estimating image geometry, Comput. Vis. Image Underst. 78 (1) (2000) 138–
156, http://dx.doi.org/10.1006/cviu.1999.0832.
[54] A. Nurunnabi, D. Belton, G. West, Robust segmentation in laser scanning 3d
point cloud data, in: 2012 International Conference on Digital Image
Computing Techniques and Applications (DICTA), 2012, pp. 1–8, http://dx.
doi.org/10.1109/DICTA.2012.6411672.
[55] P.J. Rousseeuw, C. Croux, Alternatives to the median absolute deviation, J. Am.
Stat. Assoc. 88 (424) (1993) 1273–1283.
[56] A. Khaloo, D. Lattanzi, Hierarchical dense structure-from-motion
reconstructions for infrastructure condition assessment, J. Comput. Civil Eng.
(2016) 04016047.
[57] M. Awrangjeb, C.S. Fraser, An automatic and threshold-free performance
evaluation system for building extraction techniques from airborne lidar data,
IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 7 (10) (2014) 4184–4198.