Localization and Quantification of Concrete SpallingDefects Using Terrestrial Laser ScanningMin-Koo Kim1; Hoon Sohn, M.ASCE2; and Chih-Chen Chang, M.ASCE3

Abstract: During construction and maintenance of concrete structures, it is important to achieve and preserve good surface quality of theircomponents. The current quality assessment for concrete surfaces, however, heavily relies on manual inspection, which is time demandingand costly. This study presents a new technique that can simultaneously localize and quantify spalling defects on concrete surfaces usinga terrestrial laser scanner. Defect-sensitive features, which have complementary properties to each other, are developed and combinedfor improved localization and quantification of spalling defects. A defect classifier is developed to automatically diagnose whether theinvestigated surface region is damaged, where the defect is located, and how large it is. Numerical simulations and experiments are conductedto demonstrate the effectiveness of the proposed defect-detection technique. Furthermore, a parametric study with varying scan parameters isperformed for optimal detection performance. The results demonstrate that the proposed technique can properly estimate the location andvolume of the concrete spalling defects. DOI: 10.1061/(ASCE)CP.1943-5487.0000415. © 2014 American Society of Civil Engineers.

Author keywords: Quality assessment; Concrete surface; Terrestrial laser scanner; Spalling defect; Defect localization and defectquantification.

Introduction

Surface defects such as cracks, corrosion, and spalling appear ona concrete structure during construction or in-service stages due toinappropriate curing, continuous external loading, or environmen-tal changes [Portland Cement Association (PCA) 2001]. These de-fects on a concrete surface may become severe as time passes andcan reach a certain level such that the steel rebars inside concretestructures are exposed and corrode, compromising serviceabilityand safety of the structures. Furthermore, the consequence of poorquality assessment can be expensive. It was reported that defects onconcrete components during construction, such as cracks and localmass losses, could result in rework costs up to 6–12% of the totalconstruction costs (Josephson and Hammarlund 1999).

Currently, defects on the surface of concrete structures are man-ually inspected by certified personnel. During the constructionstage of a concrete structure, inspectors use contact-type devices,such as a measuring tape or profilometer, to identify the sizeand location of surface defects. The profilometer moves above theconcrete surface in a predetermined pattern and measuresthe elevations of the surface, which are then compared with the

as-designed elevations (Tang et al. 2011). For an in-service struc-ture, visual inspection is carried out regularly by inspectors to iden-tify the extent of the defects and deterioration in the structure.According to the current National Bridge Inspection Standardsby the U.S. DOT [Federal Highway Administration (FHWA)2009], all public highway bridges in the United States shouldbe visually inspected at least once in every 2 years. During the in-spection, the inspectors evaluate the actual conditions of a bridgeby documenting and quantifying defects appearing on the primaryconcrete components, such as girders and piers. The inspectors usu-ally take photos of the defects to ensure that the locations of thesurface defects are known, and this information is used to ratethe condition of the bridge.

However, the existing methods for quality assessment of con-crete surface have several limitations. First, the results from manualinspections are subjective and may not be reliable [American Con-crete Institute (ACI) 2007; Patterson et al. 1997]. Second, manualinspection or a profilometer is costly and time consuming (Zhu andBilakis 2008). Third, a number of safety risks are associated withthe access to high-rise structures or heavy traffic zones duringmanual inspections. It is reported that 160 of 24,000 publiclyowned bridges in California are identified as those where climbingis the only means to perform bridge inspections (Sahs 2000).Hence, there is an urgent need to change the quality-assessmentparadigm for concrete surfaces from manual inspection to efficientand accurate sensing-based inspection.

Many researchers have explored noncontact sensing techniquesto overcome the limitations of the conventional inspection methods.The use of visual images is the most popular approach to detectexterior defects of a structure because it is speedy and inexpensive.As for the detection of cracks, Hutchinson and Chen (2006) pro-posed a probabilistic method based on Bayesian decision theoryfor automatic crack detection from images. Barazzetti and Scaiono(2009) used the RGB intensity to detect cracks and then computedthe crack width at a given cross section. As for the detection of airpockets, Suwwanakarn et al. (2007) proposed the use of threecircular filters to detect air pockets on the surfaces of concrete. Asfor the detection of spallings, Koch and Brilakis (2011) proposed a

1Ph.D. Student, Dept. of Civil and Environmental Engineering, HongKong Univ. of Science and Technology, Clear Water Bay, Kowloon, HongKong; and Korea Advanced Institute of Science and Technology, Daejeon305-701, Daehak-ro 291, Republic of Korea. E-mail: joekim@kaist.ac.kr;mkkim@ust.hk

2Professor, Dept. of Civil and Environmental Engineering, KoreaAdvanced Institute of Science and Technology, Daejeon 305-701,Daehak-ro 291, Republic of Korea (corresponding author). E-mail:hoonsohn@kaist.ac.kr

3Professor, Dept. of Civil and Environmental Engineering, Hong KongUniv. of Science and Technology, Clear Water Bay, Kowloon, Hong Kong.E-mail: cechang@ust.hk

Note. This manuscript was submitted on December 7, 2013; approvedon May 5, 2014; published online on July 16, 2014. Discussion period openuntil December 16, 2014; separate discussions must be submitted for in-dividual papers. This paper is part of the Journal of Computing in CivilEngineering, © ASCE, ISSN 0887-3801/04014086(12)/$25.00.

© ASCE 04014086-1 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

technique utilizing image segmentation and morphological thinning.While the image-based methods can offer good accuracy, their per-formance is heavily affected by lighting conditions. Furthermore,although identification of defect sizes such as length, width, depth,and volume is important for assessment of concrete structures, thesetypes of qualitative information cannot be retrieved using image-based methods without certain prior knowledge such as the distancebetween a camera and a target structure or the size of a referencetarget.

Recently, the terrestrial laser scanner (TLS) has gained attentionfrom the civil engineering discipline. TLS measures the geometricdata of an object’s surface by rotating the scanning head rapidlyalong both horizontal and vertical directions, and generates awealth of spatial information called a point cloud. There are twotypes of laser scanning technology, namely, (1) time-of-flight(TOF) and (2) phase shift. The TOF TLS sends out a laser pulseand measures the arrival time of the laser pulse reflected froma target point. By contrast, phase-shift TLS emits a continuoussinusoidal laser beam and estimates the distance by measuringthe phase difference between the emitted and reflected sinusoidallaser beams. Typically, the TOF TLS is suitable for long-rangemeasurement up to 6,000 m, has a measurement speed up to125,000 points=s, and 4–10 mm accuracy at a distance of100 m (RIEGL 2013). By contrast, the phase-shift TLS has a fastermeasurement speed up to 976,000 points=s, and is suitable forshort-range scanning (2–4 mm accuracy at a distance of 20 m;Olsen et al. 2010; FARO 2013). Because of these outstanding ad-vantages over the other conventional sensing techniques, TLS hasbeen widely adopted in various civil engineering fields, such asthree-dimensional (3D) modeling of structures (Son et al. 2002),historical preservation (Bernardini et al. 2002), and deflectionmeasurement of structures (Park et al. 2007; Gordon andLichti 2007).

Using the point-cloud data acquired from the TLS, researchershave proposed algorithms to detect and visualize defects onconcrete surfaces. Teza et al. (2009) proposed a defect-detectiontechnique based on the computation of the mean and Gaussiancurvatures of the surface. Tang et al. (2011) compared detectabilityof surface flatness defects with different algorithms and scanners.Olsen et al. (2010, 2013) proposed a technique that analyzed thevolumetric change of a reinforced concrete structure using I-SiteStudio 3.0 software program and discussed the potential of integrat-ing TLS scan points with intensity images for recognizing complexdefect patterns, respectively. Liu et al. (2011) proposed a distance-and gradient-based volume loss–detection technique for an in situconcrete bridge. Mizoguchi et al. (2013) estimated the depth ofscaling defects based on a customized region growing approach.

Despite a large volume of literature on TLS-based defect detectionof concrete surfaces, little attention has been paid on simultaneousand quantitative estimation of defect location and volume loss.Moreover, there has been no logistic guidance for optimal scan-ning-parameter selection, although it can be useful in practice.Hence, it is necessary to develop a technique that can accuratelylocalize and quantify concrete-surface defects, and identify an op-timal scanning parameter for enhancing the detectability of surfacedefects.

This paper presents a new and automated technique that can si-multaneously localize and quantify spalling defects on concretesurfaces using TLS. A defect classifier is developed to examinewhether each investigated surface region is damaged, where thedefect is located, and how large it is. Two defect-sensitive featuresare proposed to improve the accuracy of the defect-size estimation.A parametric study with varying scan parameters is performedwith the aims of enhancing the detectability of defects and provid-ing guideline for the optimal selection of scan parameters. Numeri-cal simulations and experimental tests are conducted on concretesurfaces to demonstrate the potential of the proposed simultaneousdefect localization and quantification technique. Compared withother previous studies, the proposed technique can offer thefollowing features: (1) autonomous and simultaneous spalling de-fect localization and volume estimation, (2) improved defectlocalization and quantification with a combination of complemen-tary defect-sensitive features, and (3) guidance for optimal scan-parameter selection.

This paper is organized as follows: First, the proposed techniquefor simultaneous spalling defect localization and volume estimationis formulated. Then, numerical simulations and experimentaltests are presented to demonstrate the effectiveness of the proposedtechnique. Finally, the paper concludes with a brief summary anddiscussions for future work.

Development of a Concrete Spalling Defect-Detection Technique

Fig. 1(a) presents an overview of the proposed concrete spallingdefect-detection technique. Here, importantly the term spallingused in this study is defined as a deeper surface defect than scaling,often appearing as circular or oval depressions on surfaces (PCA2001). The target object is assumed to be a rectangular concretestructure. The TLS is positioned at a certain distance from the targetconcrete structure and scans the concrete surface. First, a regionof interest (ROI) is selected, and then the TLS captures the 3Dcoordinate information of the scanned points inside the ROI.

Fig. 1. Overview of the proposed concrete spalling defect-detection technique: (a) TLS configuration for concrete-surface scanning; (b) overallprocedure

© ASCE 04014086-2 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

Fig. 1(b) shows the overall procedures for the proposed automatedtechnique, which include coordinate transformation, defect locali-zation, and defect quantification. The word automated used in thisstudy implies that, once the raw scan data within a ROI becomeavailable, the proposed technique operates in a fully automated wayfor all data processing from coordinate transformation to defectquantification. Details for each step are described in the followingsections.

Coordinate Transformation

Once point-cloud data are acquired from the TLS, their 3D coor-dinates with respect to the TLS are transformed to a new coordinatesystem with respect to the target object. The main purpose of thiscoordinate transformation is to remove unwanted background scanpoints behind the target so that only scan points within the targetsurface are retained for data processing. The removal can be easilyachieved after the coordinate transformation by setting a proper re-gion for each of three axes of the new coordinate system.

For the coordinate transformation, three pixel points near thecorners of the target structure are automatically determined fromthe range image, which can be generated from the point cloud.Here, each pixel of the range image and each scan point are coupledtogether so that each pixel of the range image holds the distancevalue from the TLS to the corresponding scan point. In addition,at this stage only pixel points near the exact corners are identifiedand used for the coordinate transformation because the range valuesof the exact corners have large measurement errors due to the mixedpixel problem (Hebert and Krotkov 1991). Fig. 2(a) shows therange image of the ROI of the target object. The target object closerto the TLS is shown in black while the background of the object isshown in white. To determine three pixel points near the corners ofthe target object, the exact corners of the target object are first ex-tracted by intersecting the edge lines obtained from the Houghtransform as shown in Fig. 2(b). Subsequently, the three pixelpoints used for the coordinate transformation in Fig. 2(c) are de-termined using the proposed algorithm in Fig. 3. First, candidatepixel points surrounding each corner point (C1, C2, and C3) arelimited to (1) the ones from each corner point within a certainmargin bound (ϵ) and (2) with a range distance less than a rangethreshold (TR; γ). The ϵ and γ are selected to be 10 pixels and themean range value of the range image such that only pixel points,which are on the target-object surface and 10-pixel distance fromeach corner point, are selected as candidate pixel points. An iter-ative search for the best three pixel points is then performed basedon a condition that the horizontal and vertical lines formed by thethree scan points corresponding to the selected three pixel pointsare orthogonal in the 3D space. It is important to preserve the or-thogonality condition between the two lines in order to avoid any

distortions of scan points after coordinate transformation. However,because a perfect right angle between the two lines might not beachieved due to limited spatial resolution of the TLS, three pixelpoints forming the angle closest to the right angle are determined asshown in Fig. 2(c). Once the three scan points corresponding to theselected pixel points are extracted, the rotation and the translationmatrices for coordinate transformation are determined, and thecoordinates of all scan points are transformed with respect to thetarget-object coordinate system. The unnecessary background scanpoints are then removed by setting a margin to each of the threeaxes of the new coordinate system such that only scan points on thetarget surface are retained for further data processing.

Defect Localization and Quantification

Defect-Sensitive FeaturesOnce the coordinate transformation and the elimination ofunwanted scan points are completed, defect localization and quan-tification processes follow. In this study, two defect-sensitive fea-tures, namely, angle deviation and distance deviation, are selectedand combined to enhance the detectability on the entire region ofspalling defects because each feature has its own unique yet com-plementary characteristics. Fig. 4 illustrates the definition of thetwo defect-sensitive features. The angle deviation is defined as theangle difference between the normal vector of a locally fitted planeand the reference direction and the distance deviation stands for thedistance difference between a scan point and the globally fitted

Fig. 3. Three-point selection algorithm for automated coordinatetransformation

Fig. 2. Determination of three points of the target object from a range image for coordinate transformation: (a) the initial range image of a targetspecimen within a ROI; (b) edge and corner detection in a binary image; (c) determination of three points near the corners of the target object

© ASCE 04014086-3 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

plane. Fig. 4(b) illustrates that the angle deviation from the refer-ence direction increases as the scan point within the defect areamoves closer to the defect edges. Therefore, the angle deviationis effective in identifying the defect boundaries. By contrast, asshown in Fig. 4(c), the distance deviation has a larger value nearthe center of the defect area rather than near the edges so that thedistance deviation is effective in identifying the inside region ofdefects.

First, the angle deviation for each scan point is computed asfollows:• The local plane of the ith scan point (pi) is generated from its

eight nearest neighboring points (pji ; j ¼ 1; : : : ; 8).

• The covariance matrix [CðpiÞ ∈ R3×3] of the eight neighboringpoints is computed as follows:

CðpiÞ ¼1

8

X8

j¼1

ðpji − p̄iÞ × ðpj

i − p̄iÞT ð1Þ

where p̄i = centroid of the eight neighboring points around pi.• A principal component analysis is performed on the covariance

matrix to estimate a normal vector of the local plane (Shakarji1998)

CðpiÞ × υmðpiÞ ¼ λmðpiÞ × υmðpiÞ; m ∈ f1; 2; 3g ð2Þ

whereυm and λm stand for themth eigenvector and eigenvalue ofthe covariance matrix, respectively. Because the covariancematrix (C) is symmetric and positive definite, its eigenvaluesare all real and nonnegative. When the eigenvalues are arrangedin an ascending order (0 ≤ λ1 ≤ λ2 ≤ λ3), the first eigenvector(υ1) corresponding to the smallest eigenvalue (λ1) approximatesthe normal vector þn ¼ fnx; ny; nzg or its opposite −n. Theambiguity on the sign of the normal vector is resolved by for-cing the orientation of all normal vectors to the upward pointingdirection (þz axis).

• A defect index [DI1ðpiÞ] for the first defect-sensitive feature,that is, the angle deviation from the reference direction, is de-fined as the average of the angle deviations of the eight neighborpoints surrounding pi

DI1ðpiÞ ¼1

8

X8

j¼1

θðpji Þ ð3Þ

where θðpjiÞ denotes the angle deviation of the jth nearest point

from the reference direction. Here, the reference direction is setto the þz direction perpendicular to the x-y plane. The DI1 in-creases as the scan point within the defect area moves closer tothe defect so that DI1 is effective in identifying the defectboundaries.As the second defect-sensitive feature, the distance deviation of

each scan point from a globally fitted plane is defined as illustratedin Fig. 4(c). Here, the globally fitted plane is obtained by least-square fitting a 3D linear plane into all scan points within the targetconcrete surface. Then, the distance deviation of a scan point (pi)from the global surface is computed as follows:

DI2ðpiÞ ¼����a × piðxÞ þ b × piðyÞ þ c × piðzÞ þ dffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

a2 þ b2 þ c2p

���� ð4Þ

where a, b, c, and d = coefficients of the globally fitted plane(axþ byþ czþ d ¼ 0); piðxÞ, piðyÞ, and piðzÞ ¼ x; y, and z co-ordinates of pi, respectively. The DI2 typically has a larger valuenear the center of the defect area rather than near the edges as men-tioned earlier. Therefore, DI1 and DI2 are complementary to eachother for defect localization.

Defect Identification and Quantification Procedures

The defect identification and quantification procedures aredescribed here and outlined in Fig. 5.1. Subdivision of the target surface area: The entire surface area

is virtually divided into a number of subdivisions. Because thespatial resolution of scan points within the target surface is notconstant due to different scan parameters (scan distance andincident angle) for each scan point within the target surface,this subdivision process becomes necessary to localize andquantify spalling defects. Here, the size of each subdivisionshould be small enough to achieve high defect-localizationaccuracy, but sufficiently larger than the spatial resolution oflaser scanning. The effect of the subdivision size on the defectlocalization is investigated in the experimental test section.

2. Calculation of two defect indices for each scan point: Basedon Eqs. (3) and (4), two defect indices, DI1 and DI2, arecalculated for all scan points. The global plane necessaryfor the calculation of DI2 is obtained using all scan points overthe entire target surface.

3. Calculation of two defect indices for each subdivision: Oncetwo defect indices for all scan points are obtained, the defectindices for each subdivision are computed by averaging theDI values of the scan points falling inside each subdivision.Hence, each subdivision holds two DI values [DI1ðSiÞ andDI2ðSiÞ], where Si stands for the ith subdivision.

4. Calculation of a unified DI for each subdivision ½DIðSiÞ�: Aunified DI for each subdivision is defined as follows:

DIðSiÞ ¼ α×DI1ðSiÞþ ð1−αÞ×DI2ðSiÞ; 0 < α < 1 ð5Þwhere α and (1–α) = weighting factors for DI1 and DI2, re-spectively. In this study, an equal value of 0.5 is assigned toα and (1–α) such that same contribution to detection of a

xy

Spalling defectScan point

z

x

z

z

x

Distance deviation from theglobally fitted plane

Globally fitted plane

Locally fitted plane

Normal vector of a locally fitted plane

Concrete surface

Angle deviation from the reference direction

Reference direction(+ z axis)

(a)

(b)

(c)

Fig. 4. Definitions of defect-sensitive features: (a) overview of scanpoints lying on a concrete surface; (b) definition of the angle deviationfrom the reference direction in the x-z plane view; (c) definition of thedistance deviation from the globally fitted plane in the x-z plane view

© ASCE 04014086-4 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

defect is given to DI1 (the edge region) and DI2 (the inner re-gion). DI1 and DI2 are combined here because they are com-plementary to each other as discussed in the previous section.Therefore, the combined index can preserve the individual ad-vantages of the two indices. Here, it is important to note thateach index is normalized by dividing its value with the max-imum value among all subdivisions, so that each index rangesfrom 0 to 1.

5. Calculation of a TR: A TR for defect diagnosis is computedfrom the intact subdivisions of the surface. There are two stepsto determine a TR value from the intact subdivisions. First,selection of candidate intact subdivisions is undertaken.Because actual intact subdivisions are initially unknown, can-didate intact subdivisions, having DI values lower than the TRvalue associated with 95% confidence level from the Weibulldistribution of all subdivision DI values, are determined. Notethat the confidence level is selected based on two assumptions:(1) The surface area of intact subdivisions is much larger thanthat of defect subdivisions; and (2) the intact subdivisions haveboth low DI values. Once the candidate intact subdivisionsare obtained, as the second step, each TR (TR1 and TR2) forDI1 and DI2 is established with 99% confidence level from theWeibull distribution of DI1 and DI2 values obtained fromthe candidate intact subdivisions. Finally, the unified TR isobtained as follows:

TR ¼ α × TR1 þ ð1 − αÞ × TR2 ð6Þ

The form [Eq. (6)] of calculating TR should be derived inthe same form as that of Eq. (5) because the TR is nothing but

the DI value corresponding to a certain statistical confidenceinterval. In addition, it is important to note that the candidateintact subdivisions are iteratively updated through the recur-sive defect-localization step [Step (7)] and finally converged.

6. Initial defect localization: Once DIðSiÞ and TR are computed,initial defect diagnosis is undertaken based on the followingstatement: “If DIðSiÞ exceeds TR, the corresponding subdivi-sion (Si) is diagnosed as damaged. Otherwise, the subdivisionis classified as healthy.”

7. Recursive defect localization: In this step, recursive defectlocalization is performed for improved defect localization.The initial defect-localization result may not be accurateenough because the global plane necessary for the calculationof DI2 may even include scan points from defect areas. Toimprove the defect-localization accuracy, the global plane isrefitted at this stage excluding the scan points within the defectareas identified in the previous step. Then, DI2 is updated, andSteps (3–6) are repeated until there is no difference betweenthe previous and current defect-localization results.

8. Defect quantification: Once all damaged subdivisions areidentified, the total volume loss caused by defect is estimatedby multiplying the size of each subdivision with the defectdepth (the distance deviation of each subdivision from theglobal plane) and summing up this product over all damagedsubdivisions.

Numerical Simulation

Numerical Setup

The proposed concrete spalling defect localization and quantifica-tion technique was first validated through a numerical simulation.Fig. 6 shows the configuration of the numerical simulation model.In the simulation, it is assumed that the shape of laser beam iscircular and the spatial resolution between two adjacent scan pointsis equal within the target surface. A 50 × 50 mmvirtual panel witha spatial resolution of 1 mm (total 2,601 points) was generated tosimulate a set of point-cloud data lying on a concrete surface. Twodifferent types of spalling defects, which are the most common inpractice, were considered: (1) concave-shape defect with dimen-sions of 10 × 10 mm (defect I); (2) flat-top defect with dimensionsof 10 × 10 mm (defect II). A maximum thickness of 2 mm at thecenter of the defect and a uniform thickness of 2 mm were simu-lated for defects I and II, respectively. To examine the effect of themeasurement noise of the TLS on defect-localization and quanti-fication results, Gaussian random noise with zero mean and fivedifferent levels of standard deviations (STDs; 0.1, 0.2, 0.3, 0.4,and 0.5 mm) were added to all virtual scan points.

Simulation Results

Fig. 7 shows the defect-localization results from the numericalsimulation. The results are obtained with 0.3-mm STD noise.Figs. 7(a and b) show the effectiveness of DI1 and DI2 in detectingthe edges and the inner regions of the defect areas, respectively.When the two defect-sensitive features are combined, the overalldetectability of the defects is enhanced as shown in Fig. 7(c).For the defect classification shown in Fig. 7(d), a TR value corre-sponding to a confidence interval of 99% is used. Table 1 summa-rizes the results of defect localization and volume-loss estimationwith varying measurement noise levels. All simulations are re-peated 10 times, and the average results are reported.

To evaluate the defect-localization performance of the proposedtechnique, recall and precision ratios are computed (Su 1994). Fig. 8

Fig. 5. Procedures for concrete-surface defect identification

© ASCE 04014086-5 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

Fig. 6. Configuration of numerical simulation model: (a) Two types of spalling defects, namely, concave-shape defect with a maximum deviation of2 mm (damage I) from the surface and flat-top defect with a uniform 2-mm deviation (damage II); (b) cross-sectional view of the concave-shape andflat-top defects in the x-z plane

Fig. 7. Defect-localization results (numerical simulation): (a) angle deviation from the reference direction (DI1); (b) distance deviation from theglobally fitted plane (DI2); (c) combination of two defect indices (DI); (d) defect classification (the solid lines indicate the boundaries of the actualdefect areas, and the regions in white are the detected defect regions)

Table 1. Summary of Defect-Localization and Quantification Results(Simulation)

NoiseSTD(mm)

Defect I Defect II

Recallratio(%)

Precisionratio(%)

Volume-estimationratio (%)

Recallratio(%)

Precisionratio(%)

Volume-estimationratio (%)

0.1 100.0 65.4 97.1 99.3 91.0 98.20.2 98.2 78.5 95.8 98.4 88.2 99.10.3 92.1 87.2 97.0 97.9 88.8 98.80.4 86.3 91.8 95.3 95.8 92.8 96.20.5 52.3 98.2 63.0 70.2 91.8 70.1

B

B+C

B

A+B

Fig. 8. Definition of recall and precision ratios used for the evaluationof defect-localization performance; recall ratio represents the ratio ofthe correctly detected defect area (B) to the actual defect area (Bþ C),and the precision ratio denotes the rate of the correctly detected defectarea (B) to the estimated defect area (Aþ B)

© ASCE 04014086-6 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

illustrates the definitions of these two ratios. The recall ratio rep-resents the ratio of the correctly identified defect area (B) to theactual defect area (Bþ C), whereas the precision ratio denotes theratio of the correctly identified defect area (B) to the detected defectarea (Aþ B). For the evaluation of defect-volume-estimation per-formance, the volume-estimation ratio is defined as the ratio of theestimated volume loss to the actual volume loss.

From Table 1, the following observations can be made: (1) Asthe noise level rises, the recall and volume-estimation ratios de-crease and the precision ratio increases. This trend is attributedto the fact that the TR value is mainly affected by the noise level.With a lower noise level, the TR value becomes lower and the de-fect size is overestimated. By contrast, with a higher noise level,only severely damaged subdivisions inside a defect are detected,resulting in low recall and volume-estimation ratios but a high pre-cision ratio; (2) The overall best performance is obtained with anoise level of 0.3 mm; (3) The flat-top defect (defect II) is moreeasily detected than the concave-shape defect (defect I).

Laboratory Experiments

Description of Laboratory Specimen Tests

Laboratory tests were conducted to further validate the proposedspalling defect localization and quantification technique. The over-all test configuration and details of the test specimen are shown inFig. 9. FARO Focus-3D, which offers a distance accuracy of�2 mm at 20 m, was used for data acquisition (FARO 2013).The TLS was mounted on a tripod at a height of 1.5 m, which wasthe same height as the center of the test specimen as shown inFig. 9(a). A planar Styrofoam with dimensions of 600 × 350 mmwas used as the test specimen, and eight surface defects of varyingsizes (10–100 mm) and depths (3–7 mm) were introduced at

multiple locations on the specimen surface as shown in Fig. 9(b).The eight defects were either flat top (defects 1–4) or concave shape(defects 5–8). The size and depth of each defect were selectedto simulate the typical defects found during common concrete-surface inspections (Ytterberg 1996). The subdivision size was se-lected to be 5 × 5 mm2 for defect localization and quantification.

Table 2 presents the scan parameters used in laboratory testing.To examine the effect of scan parameters on the performance of theproposed technique, a total of 18 scans were performed by varyingthe following scan parameters: (1) scan distance (4, 8, and 12 m)between the TLS and the test specimen; (2) angular resolution(0.009°, 0.018°, and 0.036°) of the TLS; and (3) incident angle(0°, 15°, and 30°) between the TLS and the test specimen. Themeasurement noise of the TLS was 0.3 mm at 10 m based on thevendor specification. The total scanning time was around 72 minfor 18 scans (average 4 min =scan). The scanning time mainly de-pends on the selection of angular resolution, and scans with denserangular resolution (0.009°) required more time than scans withcoarser angular resolution (0.018°).

Laboratory Test Results

Fig. 10 shows the test results obtained with 8-m scan distance,0.009° angular resolution, and 0° incident angle. Figs. 10(a and b)show the different sensitivities of DI1 and DI2 to the boundaries andinner areas of the defects, respectively. To minimize the mixed pixelproblem that occurred at the edges of the test specimen, both DI1and DI2 at the specimen’s boundaries were set to zero. As before,the localization performance was improved with combination of thetwo defect indices [Fig. 10(c)]. Then, defect classification on eachsubdivision was undertaken [Fig. 10(d)].

Table 3 summarizes the defect localization and quantificationwith 18 sets of scan parameters. Each entry in Table 4 is the averagevalue obtained for all eight defects. The following observationsare drawn from Table 4: (1) The incident angle is the most criticalparameter affecting the localization and volume-estimation accu-racy. In fact, the accuracy deteriorates as the incident angle in-creases from 0° to 30°; and (2) increase of the angular resolutionor the scan distance leads to a higher recall ratio and a lower pre-cision ratio.

Fig. 11 illustrates how the scan parameters affect the defect-localization performance. The localization and quantification resultfor defect 5 is presented in Fig. 11 as a representative example.Comparison of Fig. 11(a) with Fig. 11(b or d) reveals that scandistance increase (12 m) or coarser angular resolution (0.018°) re-sults in widening of the scan spacing and enhancement of thedefect-localization performance, particularly the recall ratio. Thisphenomenon can be explained as follows: As the scan spacing

TLS

Test Specimen

Flat-top defect

600 mm

Concave-shape defect

Defect 1 100 x 100 x 7 Defect 2 50 x 50 x 4Defect 3 - 30 x 30 x 3 Defect 4 - D100 x 5 Defect 5 - D100 x 7 Defect 6 - D100 x 5Defect 7 - D50 x 5Defect 8 - D10 x 4

Dimensions (mm)

(a) (b)

1

2

34

5 6

7

8

350

mm

Fig. 9. Laboratory test configuration and test specimen: (a) test configuration; (b) dimensions of the specimen and induced spalling defects

Table 2. TLS Parameters

Scan parameters Values

Distance (m) 4, 8, 12Incident angle(degrees)

0, 15, 30

Angular resolution(degrees)

0.009, 0.018

Measurement rate 976,000 points=s (0.009° angular resolution),488,000 points=s (0.018° angular resolution)

Measurementnoise (STD)

0.3 mm at 10 m, 0.5 mm at 25 m

Scanning time Total 72 min for 18 scans

© ASCE 04014086-7 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

increases due to a longer scan distance or coarser angular resolu-tion, the number of scan points lying inside each subdivision isreduced. Because DI1 of each scan point is determined by its eightneighboring points, the decrease in the number of scan points lyingnear edge subdivisions results in the increased sensitivity of DI1 tothe edge subdivisions of a defect. For instance, Fig. 12(a) illustratesthat the edge subdivision with a longer distance or coarser angularresolution has less scan points (2 intact and 2 defect points in thisexample) within the subdivision, and the corresponding DI1 is moresensitive to edge defects because the DI1 values of all the pointsare affected by the defect points. By contrast, the other edge

subdivision with a short distance or denser angular resolution in-cludes nine scan points (6 intact and 3 defect points) within thesubdivision, and results in a relatively lower DI1 value becausethe left three intact scan points are not influenced by the defectpoints. As for the incident angle effect, Fig. 11(c) shows that, asthe incident angle increases from 0° to 30°, the location of the esti-mated defect is shifted to the right of the actual defect. Fig. 12(b)illustrates that, as the incident angle increases, the scan spacing be-comes coarser and subdivisions at the edge region of the defect maynot include any scan points. In addition, this phenomenon becomesmore serious when the defect depth increases at the defect edges.Therefore, the scan spacing with respect to the subdivision sizeshould be properly selected for optical defect-localization andquantification performance.

Actual Concrete-Panel Test

To further examine the feasibility of the proposed localizationand quantification technique for concrete spalling defects, an ex-periment was performed on an actual concrete panel with dimen-sions of 1,200 × 900 × 150 mm. Fig. 13 shows the overall testconfiguration and the details of the concrete panel. The sameTLS used for the laboratory specimen test was utilized and posi-tioned at a fixed distance of 10 m from the concrete panel as shownin Fig. 13(a). The concrete panel was fixed to a concrete wall, andthe TLS scanned the panel at two different angular resolutions,0.009° and 0.018°. Nine artificial spalling defects with different

Fig. 10.Defect-localization results (laboratory test with 8-m scan distance, 0.009° angular resolution, and 0° incident angle): (a) angle deviation fromthe reference direction (DI1); (b) distance deviation from the globally fitted plane (DI2); (c) combination of two damage features (DI); (d) defectclassification (the solid lines indicate the boundaries of the actual defect area, and the regions in white are the detected defect regions)

Table 3. Summary of Defect-Localization and Quantification Results(Laboratory Test)

Angularresolution(degrees)

Scanningdistance(m)

Incidentangle

(degrees)

Recallratio(%)

Precisionratio(%)

Volume-estimationratio (%)

0.009 4 0 80.6 97.1 85.40.009 4 15 76.0 94.7 74.50.009 4 30 68.2 90.3 73.80.009 8 0 84.7 90.5 88.40.009 8 15 84.5 90.1 89.20.009 8 30 79.6 88.7 85.20.009 12 0 88.1 93.0 88.90.009 12 15 84.7 90.1 87.50.009 12 30 72.9 84.5 78.70.018 4 0 87.7 92.6 86.80.018 4 15 87.3 93.2 81.90.018 4 30 76.0 87.7 81.60.018 8 0 91.6 89.3 90.10.018 8 15 89.6 85.7 90.30.018 8 30 79.5 83.4 81.70.018 12 0 92.3 90.7 90.00.018 12 15 71.8 96.8 74.90.018 12 30 49.8 85.9 50.9

Table 4. Defect-Localization Results (Concrete-Panel Test)

Scanningdistance (m)

Incidentangle

(degrees)

Angularresolution(degrees)

Recallratio(%)

Precisionratio (%)

10 0 0.009 89.1 89.810 0 0.018 94.6 88.1

© ASCE 04014086-8 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

dimensions and depths were introduced on the concrete-panel sur-face as shown in Fig. 13(b). The defects were either flat top (defects1–5) or concave shape (defects 6–9). The size of each subdivisionwas chosen to be 5 × 5 mm2 as before such that the whole surfacearea is divided into 43,200 subdivisions.

Fig. 14 shows the defect-localization results obtained with0.009° angular resolution. As expected, Figs. 14(a and b) show theeffectiveness of each defect-sensitive feature for detecting the edgeand inner areas of the defects, respectively. Combining the twodamage indices, enhanced defect-localization results are obtainedin Fig. 14(c), followed by the defect classification [Fig. 14(d)]. Thedefect-localization results show that all defects were successfullyidentified except defects 3 and 4, which have very small thicknessdeviations (3 and 2 mm). The manual inspection of the concretepanel revealed that the upper left corner of the panel has a notice-able nonflat surface, and this nonuniform surface condition isattributed to difficulties in detecting shallow spalling defects (de-fects 3 and 4). Tables 4 and 5 show the defect-localization and vol-ume loss–estimation results for the concrete panel, respectively.Each entry in Table 4 is the average value of the seven successfullylocalized defects. The recall and precision ratios are 91.9% and88.9%, respectively. In Table 5, the average volume loss–estimationratio is 84.8% in the case of 0.009° angular resolution. The resultsdemonstrate that the proposed technique is able to successfullylocate and quantify all surface defects except defects 3 and 4.

Discussion

All the simulation and test results presented in this study clearlyindicate the importance of proper scan-parameter selection. It isrecommended that the scan distance, incident angle, and angularresolution be no more than 12 m, 15° and 0.018°, respectively, toachieve correct localization and volume loss–estimation ratios ofover 80% for spalling defects using the proposed technique. Forexample, if the TLS is 10 m away from a concrete structure, anallowed scan area corresponding to a maximum incident angleof 15° becomes 5.4 × 5.4 m2.

Fig. 11. Effects of the scan parameters on defect localization: (a) reference scan parameters (8 m distance, 0° incident angle, and 0.009° angularresolution); (b) 12 m distance; (c) 30° incident; (d) 0.018° angular resolution

Fig. 12. Effects of scan distance, angular resolution, and incident angleon scan spacing: (a) the number of scan points within each subdivisiondecreases with a longer scan distance or coarser angular resolution;(b) with an increasing incident angle, subdivisions located at the defectedges may not have any scan points

© ASCE 04014086-9 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

Fig. 13. Test setup and actual concrete panel: (a) test configuration; (b) dimensions of the specimen and induced defects

Fig. 14. Defect-localization results (concrete-panel test): (a) angle deviation from the reference direction (DI1); (b) distance deviation from theglobally fitted plane (DI2); (c) combination of two defect indices (DI); (d) defect classification (the solid lines indicate the boundaries of the actualdefect areas, and the regions in white are the detected defect regions)

Table 5. Defect Volume Loss–Estimation Results (Concrete-Panel Test)

Aspects

Defect number

1 2 5 6 7 8 9

Design Estimation Design Estimation Design Estimation Design Estimation Design Estimation Design Estimation Design Estimation

Volume loss(10−6 m3)

486.0 428.3 72.0 80.2 254.5 276.4 353.4 304.4 150.8 160.7 18.9 14.7 1.9 1.3

Estimationratio (%)

88.1 88.1 88.6 88.6 91.3 91.3 86.1 86.1 93.3 93.3 78.0 78.0 68.3 68.3

Estimationaverage (%)

84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8 84.8

© ASCE 04014086-10 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

Second, the relation between the subdivision size and scan spac-ing needs to be considered carefully. The size of the subdivisionshould be small enough to achieve precise defect localization,but sufficiently larger than the laser scan spacing. Theoretically,the subdivision size should be larger than the following scanspacing (Δ):

Δ ¼ L × αR

cosðαIÞð7Þ

where L denotes the scan distance between the concrete surface andthe TLS, and αR and αI denote the angular resolution and theincident angle, respectively. For example, if the subdivision sizeis 5 × 5 mm2, and L, αI , and αR are set to 15 m, 30°, and 0.018°,respectively, the scanning can be problematic because Δ (5.4 mm)is larger than the subdivision size. Based on the experimental re-sults of this study, it is recommended that at least four scan pointsare included in each subdivision. Therefore, the scan parametersshould be carefully selected to ensure a good accuracy of spallingdefect localization and quantification.

Third, scanning time also needs to be considered. The scanningtime is mainly dependent on selection of angular resolution, whichdictates horizontal and vertical speeds of TLS. The size of a targetstructure also influences the scanning time. If scan parameters suchas angular resolution and scan distance are assumed to be fixed,bigger size of target surface requires more scanning time thansmall-sized ones. Therefore, proper estimation of scanning timeshould be conducted before actual inspections for effective spallingdefect detection.

Conclusions

This study presents a new defect-detection technique that cansimultaneously localize and quantify spalling defects on concretesurfaces using a TLS. Two defect-sensitive features, which arecomplementary to each other, are proposed and combined forimproved spalling localization and quantification performance.A defect classifier is also developed to automatically diagnosewhether the investigated surface region is damaged, where thedefect is located, and how large it is. Numerical simulations andexperiments are conducted to demonstrate the effectiveness ofthe proposed defect-detection technique. The results demonstratethat the proposed technique can accurately estimate the locationand volume of concrete spalling defects. Furthermore, a parametricstudy with varying scan parameters of TLS is performed for opti-mal scan-parameter selection. The proposed technique can offerthe following features: (1) autonomous and simultaneous spallingdefect localization and volume estimation, (2) improved defectlocalization and quantification with a combination of complemen-tary defect-sensitive features, and (3) guidance for optimal scan-parameter selection. In addition, the authors believe based on theresults obtained that the proposed technique can be used to detectany spalling defect, which is bigger than 3 mm in both length anddepth. Currently, the applicability of the proposed technique is,however, limited to flat concrete surfaces and two types of defects(concave-shape and flat-top defects), and shallow spalling defects(less than 3 mm deep) can hardly be detected. Further investigationis underway in the following directions: (1) extension of the appli-cability of the proposed technique to more complex structures andvarious types of defects, (2) enhancement of the detectability onshallow defects, and (3) integration of the TLS data with intensityimages for more comprehensive analysis of spalling defects.

Acknowledgments

This work was supported by a grant (Grant No. 13SCIPA01) fromthe Smart Civil Infrastructure Research Program funded by theMinistry of Land, Infrastructure and Transport of the KoreaGovernment and the Korea Agency for Infrastructure TechnologyAdvancement, and a grant (Grant No. 2010-0017456) from Mid-career Researcher Program of the National Research Foundation ofKorea (NRF) funded by the Ministry of Education, Science andTechnology. The first author likes to acknowledge supports fromthe Postgraduate Student Award from the School of Engineeringof Hong Kong University of Science and Technology and a JointPh.D. Degree Program between Korea Advanced Institute ofScience and Technology and Hong Kong University of Scienceand Technology.

References

American Concrete Institute. (2007). “ACI manual of concrete inspection.”ACI Committee 311, SP-2(07), Detroit.

Barazzetti, L., and Scaiono, M. (2009). “Crack measurement: Develop-ment, testing and applications of an automatic image-based algorithm.”J. Photogramm. Remote Sens., 64(3), 285–296.

Bernardini, F., Martin, I., Mittleman, J., Rushmeier, H., and Taubin, G.(2002). “Building a digital model of Michelangelo’s Florentine Pieta.”IEEE Comput. Graphics, 22(1), 59–67.

FARO. (2013). “Focus-3D technical specification.” ⟨http://www.faro.com/site/resources/share/944⟩ (Dec. 6, 2013).

Federal Highway Administration. (2009). “National Bridge InspectionStandards.” Fed. Regist., 74(246), 68377–68379.

Gordon, S., and Lichti, D. D. (2007). “Modeling terrestrial laser scannerdata for precise structural deformation measurement.” J. Struct. Eng.,10.1061/(ASCE)0733-9453(2007)133:2(72), 72–80.

Hebert, M., and Krotkov, E. (1991). “3D measurements from imaging laserradar: How good are they?” Image Vis. Comput., 10(3), 170–178.

Hutchinson, T. C., and Chen, Z. Q. (2006). “Improved image analysis forevaluating concrete damage.” J. Comput. Civ. Eng.,10.1061/(ASCE)0887-3801(2006)20:3(210), 210–216.

I-Site Studio 3.0 [Computer software]. Glenside, SA, Maptek.Josephson, P. E., and Hammarlund, Y. (1999). “The causes and costs of

defects in construction: A study of seven building projects.” Autom.Constr., 8(6), 681–687.

Koch, G. M., and Brilakis, I. K. (2011). “Pothole detection in asphalt pave-ment images.” Adv. Eng. Inf., 25(3), 507–515.

Liu, W., Chen, S., and Hauser, E. (2011). “LiDAR-based bridge structuredefect detection.” Exp. Tech., 35(6), 27–34.

Mizoguchi, T., et al. (2013). “Quantitative scaling evaluation of concretestructures based on terrestrial laser scanning.” Autom. Constr., 35,263–274.

Olsen, N. J., Chen, Z., Hutchinson, T., and Kuester, F. (2013). “Opticaltechniques for multiscale damage assessment.” Geomatics Nat.Hazards Risk, 4(1), 49–70.

Olsen, N. J., Kuester, F., Chang, B. J., and Hutchinson, T. (2010). “Terres-trial laser scanning-based structural damage assessment.” J. Comput.Civ. Eng., 10.1061/(ASCE)CP.1943-5487.0000028, 264–272.

Park, H. S., Lee, H. M., Hojjat, A., and Lee, I. (2007). “A new approachfor health monitoring of structures: Terrestrial laser scanning.”Comput.-Aided Civ. Infrastruct. Eng., 22(1), 19–30.

Patterson, A., Dowling, G., and Chamberlain, D. (1997). “Building inspec-tion: Can computer vision help.” Autom. Constr., 7(1), 13–20.

Portland Cement Association (PCA). (2001). Concrete slab surfacedefects: Causes, prevention, and repair, Skokie, IL.

RIEGL. (2013). “RIEGL laser measurement systems.” ⟨http://www.riegl.com⟩ (Dec. 6, 2013).

Sahs, S. (2000). “ Climbing techniques for bridge inspection.” Proc.,8th Int. Bridge Management Conf., Transportation Research Board,Washington, DC.

© ASCE 04014086-11 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.

Shakarji, C. (1998). “Least-squares fitting algorithms of the NISTalgorithm testing system.” J. Res. Natl. Inst. Stand. Technol.,

103(6), 633–641.Son, S., Park, H., and Lee, K. H. (2002). “Automated laser scanning system

for reverse engineering and inspection.” Int. J. Mach. Tools Manuf.,42(8), 889–897.

Su, L. T. (1994). “The relevance of recall and precision in user evaluation.”J. Am. Soc. Inf. Sci., 45(3), 207–217.

Suwwanakarn, S., Zhu, Z., and Brilakis, I. (2007). “Automated air pocketsdetection for architectural concrete inspection.” Proc., ConstructionCongress I, American Society of Civil Engineers, Reston, VA.

Tang, P., Hubber, D., and Akinci, B. (2011). “Characterization of laser scan-ners and algorithms for detection flatness defects on concrete surfaces.” J.Comput. Civ. Eng., 10.1061/(ASCE)CP.1943-5487.0000073, 31–42.

Teza, G., Galgaro, A., and Moro, F. (2009). “Contactless recognition ofconcrete surface damage from laser scanning and curvature computa-tion.” NDT&E Int., 42(4), 240–249.

Ytterberg, C. N. (1996). “Flatness tolerances for random-traffic floors.”Concrete construction magazine, ⟨http://www.concreteconstruction.net⟩ (Dec. 6, 2013).

Zhu, Z., and Bilakis, I. (2008). “Detecting air pockets for architecturalconcrete quality assessment using visual sensing.” J. Inf. Technol.Constr., 13, 86–102.

© ASCE 04014086-12 J. Comput. Civ. Eng.

J. Comput. Civ. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

KO

RE

A A

DV

AN

CE

D I

NST

. OF

on 0

9/30

/14.

Cop

yrig

ht A

SCE

. For

per

sona

l use

onl

y; a

ll ri

ghts

res

erve

d.