Upload
ingrid-giacomeli
View
213
Download
0
Embed Size (px)
Citation preview
8/12/2019 1-s2.0-S0926580510001147-main
1/9
Detection of large-scale concrete columns for automated bridge inspection
Z. Zhu , S. German, I. Brilakis
School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA. 30332, USA
a b s t r a c ta r t i c l e i n f o
Article history:
Accepted 29 July 2010
Keywords:
Concrete columns
Automatic identication systems
Images
Information technology
Bridge inspection
There are over 600,000 bridges in the US, and not all of them can be inspected and maintained within the
specied time frame. This is because manually inspecting bridges is a time-consuming and costly task, and
some state Departments of Transportation (DOT) cannot afford the essential costs and manpower. In this
paper, a novel method that can detect large-scale bridge concrete columns is proposed for the purpose ofeventually creating an automated bridge condition assessment system. The method employs image stitching
techniques (feature detection and matching, image afne transformation and blending) to combine images
containing different segments of one column into a single image. Following that, bridge columns are detected
by locating their boundaries and classifying the material within each boundary in the stitched image.
Preliminary test results of 114 concrete bridge columns stitched from 373 close-up, partial images of the
columns indicate that the method can correctly detect 89.7% of these elements, and thus, the viability of the
application of this research.
Published by Elsevier B.V.
1. Introduction
According to data from the Bureau of Transportation Statistics up
to 2008, there arecurrently more than 600,000 highway bridges in theUSA, around 25% of which are rated as either structurally decient,
functionally obsolete or both[1]. At current spending levels (roughly
$10.5 billion per year),it is estimated that the annual gap of 1.9 billion
dollars is required to reduce this percentage to near zero by 2024 [2].
The problem could be alleviated with the introduction of automated
inspection to replace current manual practices, which costs local
Departments of Transportation (DOT) millions of dollars every year
[3]. Each of these bridges requires inspection at regular intervals
(usually not exceeding two years with few exceptions) to determine
their physical and functional conditions and ensure that they still
satisfy present service requirements [4]. Routine inspections are
widely adopted and are carried out manually by certied bridge
inspectors, following the National Bridge Inspection Standards by the
Federal Highway Administration (FHWA) [5] and the Manual on
Bridge Evaluation by the American Association of State Highway and
Transportation Ofcials (AASHTO)[4]. Inspection results are mainly
based on the inspectors' observations and visual assessment and
represent the condition state information on bridge elements (e.g.
columns, abutments, decks/slabs, and girders). This information is
input into a bridge management system, such as PONTIS[6], to help
transportation agencies make system-wide prioritization decisions in
allocating limited construction maintenance and rehabilitating
resources.
Several limitations of manual inspections have been identied in
previous research studies. First, the results from manual inspectionsare subjective and not always reliable [7,8]. Second, manually
collecting bridge inspection data is costly and time-consuming
[9,10]. Third, a number of safety risks are associated with inspections
since inspectors often work at high heights or in heavy trafc zones
[11]. Also, the requirement of experienced inspectors in bridge
inspection poses a challenge for the construction industry, which is
now facing the pressing shortage of experienced and highly trained
inspection personnel[12].
In order to overcome these limitations, it is necessary to create
automated bridge inspection toolkits that can considerably reduce the
eld work required for inspectors and automatically produce bridge
inspection reports. Toward this objective, several methods have been
created. For example, Jauregui et al. utilized QuickTime Virtual Reality
and panoramic image-creation tools for recording observations and
measurements during bridge eld inspections [9]. Jaselskis et al.
investigated the use of laser scanning to acquire and process bridge
as-built data[13]. Moreover, Abdel-Qader et al. focused on detecting
the presence of bridge surface cracks [14], and they found that the fast
Haar transform had more accurate crack detection results than three
other edge detection techniques (Canny, Sobel and Fourier Trans-
form)[15]. Although these methods help inspectors nd defects on
bridges, the detected defects cannot be used directly to rate bridge
conditions. This is because the measurements of these defects are of
little value unless they can be spatially correlated with the members
on which they lay. For example, a diagonal (shear) crack versus a
horizontal (exural) crack indicates a completely different type of
Automation in Construction 19 (2010) 10471055
Corresponding author. Tel.: +1 404 385 1276.
E-mail addresses: [email protected](Z. Zhu),[email protected](S. German),
[email protected](I. Brilakis).
0926-5805/$ see front matter. Published by Elsevier B.V.
doi:10.1016/j.autcon.2010.07.016
Contents lists available at ScienceDirect
Automation in Construction
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / a u t c o n
http://dx.doi.org/10.1016/j.autcon.2010.07.016http://dx.doi.org/10.1016/j.autcon.2010.07.016http://dx.doi.org/10.1016/j.autcon.2010.07.016mailto:[email protected]:[email protected]:[email protected]://dx.doi.org/10.1016/j.autcon.2010.07.016http://www.sciencedirect.com/science/journal/09265805http://www.sciencedirect.com/science/journal/09265805http://dx.doi.org/10.1016/j.autcon.2010.07.016mailto:[email protected]:[email protected]:[email protected]://dx.doi.org/10.1016/j.autcon.2010.07.0168/12/2019 1-s2.0-S0926580510001147-main
2/9
damage on a column. Therefore, bridge elements also have to be
detected, so that the defects can then be spatially correlated with the
corresponding elements to produce relative measurements for
automated bridge inspection. The detection of bridge elements is
always the rst step for automated bridge assessment. This will be
followed by defects detection (e.g. crack detection) and defects
impact assessment.
Existing work in structural element detection cannot capture both
the bridge columns and the detailed defect information on theirsurfaces from a single image, which is necessary in order to accurately
correlate the defect and member spatial properties. This is because
bridge elements are usually large-scale. For the purpose of this study,
a large-scale structural element is dened as the element that cannot
be completely t into an image when the smallest defects of interest
occupy at least 200 pixels on the element surfaces [16]. In order to
detect these elements and retrieve their dimensions, existing
methods require structural elements to be entirely visible in the
eld of view (FOV). This forces users to fully zoom-out with their
cameras or stand far away when capturing large-scale columns,
considering that the FOV of a typical compact camera is limited to
50(H)35(V), when the human's FOV can reach 200(H)135(V)
[17]. On the other hand, capturing defect data often requires close up
images as to allow for detailed information retrieval concerning the
extent and impact of each defect. As a result, when columns are fully
visible, the defect image data on these elements is invisible ( Fig. 1).
This paper presents an automated method for the detection of
large-scale concrete bridge columns captured in multiple close-range
images taken for the purpose of defects detection and assessment. To
the authors' knowledge, there are no existing computer vision
techniques that can be used directly for concrete column recognition
other than those presented in the paper. As a result, current existing
defect detection methods can only be used to detect surface defects
after structural element surfaces are manually extracted from image
background[18].
The proposed method starts with the SIFT detector to nd the
features that are invariant to afne transformation in each image. The
SIFT-based descriptors exhibit high matching accuracy, which makes
them out-perform other local descriptors [19,20]. Although thecomputational complexity is associated with the SIFT detector, it is
not a critical issue in this study, since no real-time requirements are
imposed for routine bridge inspection. The detected features are then
matched to calculateimage transformationmatrices, which areused to
combine the imagescontaining different segments of one columninto
one single image. Following that, the concrete column surfaces in the
image are identied.An edge mapis produced from thestitched image
using the Canny edge detector, and the long vertical lines in the map
are retrieved using the Hough Transform. The bounding rectangle for
each pair of long vertical lines is created. If the resulting rectangle
resembles the shape of a column (i.e. the rectangle's width is smaller
than its length) and the color/texture in the rectangle are matched to
concrete based on the classication result of an articial neural
network, the region between the pair of lines is assumed to be a
concrete column surface, belonging to a concrete bridge column.
The method was implemented using Microsoft Visual C++. Adatabase of close-range concrete bridge images (373 total) was used
to test the method. The test results were compared with manual
detection results to nd the detection precision and recall of the
method. Precision measures how many detected columns are
correctly identied, while recall measures how many actual columns
are correctly identied. Based on the test results, the average
precision ratio and recall of the method are 89.7% and 84.3%,
respectively, which validate that most large-scale concrete bridge
columns can be correctly detected using the method proposed in this
paper.
2. Background
2.1. Manual bridge routine inspection
Highway bridge inspections are carried out manually by certied
inspectors following the established standards and manuals (i.e.
National Bridge Inspection Standards and AASHTO Manual on Bridge
Evaluation). Before going onto a bridge, the inspectors prepare
sketches and note templates for references throughout the inspection
[21]. During inspection, they record the actual bridge conditions by
observing existing defects that lay on primary bridge components,
such as decks, exterior beams and piers. The defects to be inspected
include different types of cracks (e.g. exural, shear, vertical and bond
cracks), loss of cover and spalling, and corrosion and eforescence.
This information is used to rate bridgeconditions. So far, there aretwo
rating systems that can be adopted. The FHWA Recording and Coding
Guide [22] denes one system which uses the National Bridge
Inventory (NBI) zero to nine scale for rating a bridge decksuperstructure and substructure. The second system, the PONTIS
rating system, uses a set of three to ve condition states to describe
the conditionof approximately 160 bridge elements, such as columns,
girders, slabs and trusses [23]. Both rating systems are built on
qualitative denitions. For example, in the NBI rating system, a bridge
is rated at Fair Condition (5) when all of its primary components are
sound but may have minor cracking, spalling, or scour [22].
Fig. 1.The loss of detailed defect information when tting a large-scale column in the view.
1048 Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
8/12/2019 1-s2.0-S0926580510001147-main
3/9
Such manual inspections have several inherent limitations. First,
they are inefcient. Reviewing previous reports, preparing inspection
plans, and collecting and analyzing eld data requires a large amount
of inspector hours [9,20], which has proven costly over time. In
Tennessee, in 2006 and 2007, the annual cost of bridgeinspection was
$7.6 million; considering the Tennessee Department of Transporta-
tion lost $22million in Federal bridgefundsin thelast twoyears, some
bridge inspections and maintenance projects have to be pushed back
[24]. In addition, the results from manual inspections are subjectiveand highly variable. An on-site survey conducted by FHWA Non-
Destructive Evaluation Center (NDEVC) indicated that manual
inspections were completed with signicant variability[7,8]. Finally,
the current guidelines need considerable updating with regard to
health monitoring and non-destructive testing techniques, as well as
in regards to theproper integration of that data into the ratingprocess
[25].
2.2. Recent research efforts towards automated bridge routine inspection
Researchers have proposed many methods regarding the aim of
automated bridge inspection, most of which focus on as-built bridge
data collection and defect detection. Laser scanning and photogram-
metry are two remote sensing techniques recently used to collect
bridge as-built data. Laser scanning can create accurate and detailed
3D range images. Most laser scanners used in construction adopt the
time-of-ight principle. Other scanners adopting triangulation or
structured-light are seldom applied for scanning large structures due
to their limited scanning range of meters and/or safety constraints
[26,27]. In the time-of-ight principle, a pulse of laser light is emitted
to probe the object in the scene, and a time-of-ight laser range nder
at the heart of this type of scanner, nds the distance of a surface by
timing the round-trip time of the pulse of light. So far, this technique
has been used in several bridge survey applications [13,28]. Aside
from laser scanning, photogrammetry has also been used for this
purpose, due to the affordable and portable equipment involved[29].
It relies solely on the present light energy to retrieve bridge geometry
data, which is implicitly contained in images captured from different
perspectives. Examples of using photogrammetry to document bridgeas-built conditions can be found in the work of Jauregui et al. [9,30].
As for defect detection, many methods have been created to locate
cracks in images. They use image processing techniques, such as
wavelet transforms and segmentation, to extract crack points from
the image background. The effectiveness of these methods has been
veried in real concrete structures. For example, Abdel-Qader et al.
proposed a principal component analysis (PCA) based algorithm for
detecting unsupervised bridge cracks [14]. Also, they compared the
effectiveness of four edge detection techniques (Canny edge detector,
Sobel edge detector, Fourier transform and fast Haar transform) in
detecting cracks on concrete bridges[15]. According to the detection
results for 50 images, they found that the fast Haar transform was the
most reliable in crack detection. Sinha and Fieguth took advantage of
the linear property of crack features and introduced two crackdetectors for identifying crack pieces [31]. Yu et al. used a graphical
search to retrieve an integral crack based on manually provided
endpoints[32].
2.3. Structural element detection
Automating the collection of bridge as-built data and the detection
of surface defects is not entirely sufcient to provide a comprehensive
bridge inspection. In order to produce bridge ratings, the impact of
existing defects on bridge elements needs to be further evaluated,
which rst requires the detected defects to be spatially correlated
with the member on which they lay in order that then, the
interpretation of the specic defects according to the specic member
can take place. For this reason, bridge elements need to be detected
(Fig. 2). Although element detection is only one step in solving the
whole defects interpretation problem, it is always the rst step, and
critical to evaluate existing defects on bridge elements later. The
purpose of this concentrates on the automated detection of bridge
elements while keeping as much of thedetaileddefects information as
possible.As for the subsequent steps precise crack detection and the
structural analysis of existing cracks at the level of structural
engineering domains they are beyond the discussion of this paper.
In general, object detection is de
ned as locating a desired objectin a scene[33]. It can be performed in range data (3D) or visual data
(2D). The methods of detecting objects in range data rely on the
object's shape features, such as surfaces or contours, since range data
are insensitive to illuminations [34]. The shape information can be
described globally or locally. The global shape descriptors [35,36]
capture all shape characteristics of a desired object in range data.
Therefore, they are more discriminative, but less robust, than the local
shape descriptors[37,38]in cluttered scenes, where the range data of
the desired object is easily occluded [39]. Both types of methods
(global descriptors and local descriptors) are computationally
complex, especially when the number of search objects and the
number of scanned points is increased[40]. In order to overcome this
limitation, Bosche and Haas suggested representing a 3D CAD model
using range points so that the matching can be performed at the point
level. However, this method requires the model to be registered rst
in the laser scanner's spherical coordinate frame [41].
Objectdetection methods in 2D visualdata canbe further classied
as: 1) scale/afne-invariant feature-based, 2) color/texture-based and
3) shape-based. 2D Scale/afne-invariant feature-based methods use
image scale/afne-invariant features extracted by feature detectors,
such as SIFT[42], to locate an object in the image. Such methods have
shownto be powerful indetecting a specic object. Recently,the useof
3D local shape descriptors was investigated for object class detection.
For example, Ruiz-Correa et al. created numeric signature to form a
constellation of components to quantify 3D object surface geometry,
and then relied on a set of symbolic signatures to encode the spatial
conguration of these component for object category detection [43].
However, the concept of these 3D local shape descriptors cannot be
applied in object category detection in 2D visual data, especially instructural element detection. This is because structural elements
always have a uniform material on their surfaces, which makes it
difcult to retrieve enough 2D scale-invariant features to characterize
the geometry of whole 2D element surfaces. Moreover, structural
Fig. 2.Automated bridge inspection.
1049Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
http://localhost/var/www/apps/conversion/tmp/scratch_5/image%20of%20Fig.%E0%B2%808/12/2019 1-s2.0-S0926580510001147-main
4/9
elements, although geometrically simple, are characterized by large
topographical variations (e.g. aspect ratio) and therefore no simple
scale/afne transformation can be used to characterize them[44].
Color/texture based methods rely on the objects' interior color and
texturevalues to perform detection.Neto et al.observed that thecolor and
texture values for most materials (e.g. concrete and steel) in an image do
not change signicantly[45].Based on this observation, material regions
in an image can be identied by checking their color and texture values
[16,46]. Moreover, two dimensions of the detected regions (MCD andPMCD) for concrete and steel can be used to determine whether the type
of structural element of the region is a beam or a column[47]. However,
when one element (e.g. a concrete column) is connected to another
structural element with the same material (e.g. a concrete beam) this
method regards them as one single element instead of two separate
elements and produces erroneous results[44].
Edge information is another indicator for structural element
detection. Edges are dened as the areas where image intensities
change signicantly (e.g. discontinuities in object surface orienta-
tion). Shape-based methods make use of this information to identify
object boundaries by analyzing the distribution of edge points using
the Hough transform [48], covariance matrices [49] or principle
component analysis[50]. Such methods can detect thin and stick-like
objects whose color and texture are easily corrupted by image
background[51], but the sole reliance on edge information renders
these methods inadequate in complex scenes[52].
The authors have created a building column detection method using
both boundary and material information in an image/video [44]. The
method considered that each building concrete column has a pair of long
vertical lines and the textureand color patterns on its surface are uniform.
Based on these two visual characteristics, the method located building
concrete columns in images or videos by identifying their boundaries
using edge detection and Hough transform techniques and recognizing
concrete material using the concept ofmaterial signatures[46].
The proposed method, like other column detection methods,
requires concrete columns to be entirely visible in the FOV of a
camera. This limitation has no signicant adverse effects in most
construction applications. Take automated project progress monitor-
ing for an example. A user can always adjust a camera (changingposition, shooting direction, etc.) to an appropriate view, so that
concrete columns in the view can be separated each other, and the
concrete construction project progress can be automated by auto-
matically counting the number of concrete columns that have been
built at the site. However, in automated inspection, the method is not
always applicable, since it is difcult to capture the detailed defect
information on column surfaces in one image, when a user has to
adjust the FOV of a camera to t an entire column. It is common that
the user takes multiple pictures to capture the defects that are located
at different areas of one column and then, combine all defect
information contained in these pictures for the sake of rating the
column.
2.4. Image stitching
Image stitching is the combination of multiple images into one high-
resolution image based on their overlapping areas. It is extensively
investigated in the eld of computer vision, and so far there are several
commercial applications available, such as PhotoStitch[53]and REALVIZ
Stitcher[54]. In general, the process of stitching any two images can be
divided into two steps: 1) image alignment, the analysis of image
orientation information (e.g. translation and rotation), and 2) image
blending, the smooth combination of the color values of two images in
their overlapping areas[55].
Image alignment is performed using direct-based methods or feature-
based methods. Direct-based methods are also called pixel-based
methods [56]. These methods search for a transformation matrix that
can minimize the sum of absolute differences between the overlapping
pixels of two images. For example, Baker and Matthews generalized the
LucasKanade algorithm to implement the work of image alignment [57].
Similar work can also be found in the work of Agarwala et al. [58]and
Shum and Szeliski [59]. All of these methods provide accurate orientation
information since they use all available image data; however, a close
initialization is always required[60]. Also, they have a limited range of
convergence; thus,they work in matchingsequentialframes of a video but
fail too often to be useful in photo matching [59]. Feature-based methods
overcome these limitations. They determine image orientations byautomatically detecting and matching the same features that appear in
multiple images. Features can be points, line segments, or regions.
Matching the features in images is solved through the use of multiple
randomized kd trees, a spill tree or a hierarchical k-means tree[61].
Feature-basedmethodsare remarkably robust, since they caneven detect
and match the features in images that differ in scale [56].
The process of blending two images is nothing more than
calculating their weighted average. The weights can be simply xed
at 0.5 for both images or determined based on their distance
maps[62]. The combination of pixel values using a weighted average
always produces inaccuracies, such as double exposure, visible seams,
color distortion and contrast loss [63]. In order to address these
problems, image blending is performed in multiple resolutions or
bands. For example, Grundland et al. represented images with
Laplacian pyramids and then the blending was performed at each
level of the pyramids independently [63]. Allen et al. convolved an
image with a xed Gaussian kernel multiple times to generate
increasingly smoothed versions[64]. The differences between these
versions represent the image at different frequency levels. The
blending of two images was performed at the frequency level of
each image rst, and the ultimate result was collected from the
summation of the blending results at each level.
Brown and Lowe presented a novel system that can fully automate
these two steps[60]. In their work, the feature detector [42]is rst
used to extract and match the features that are invariant to image
scale and afne distortion in each image. Once these features are
matched, the locations ( x1 ;y1 T
; x2 ;y2 T
:::::: xn ;yn T) of these fea-
tures are then retrieved. For any two images, the afne transform
matrix between the two images can be calculated using Equation 1where x11 ;y
11
T; x12 ;y
12
T::: x1n ;y
1n
Tare feature locations in the rst
image, x21 ;y21
T; x22 ;y
22
T::: x2n ;y
2n
Tare the matched feature locations
in the second image and matrix A12 = a11 ; a12 ; a13
a21 ; a22 ; a23
is the afne
transform matrix that can transfer the locations of any image pixel in
the rst image to the second one. Thus, these two images can be
combined into one image with the afne transform matrix, and the
combination can be performed iteratively until one composite image
is created from all individual images. Their approach is robust for
automatically stitching most bridge column images based on the
following test results. For this reason, the method proposed in this
paper directly adopts their approach for image stitching. The main
contribution of this paper is to create a framework that combines the
method of image stitching with the method of normal concretecolumn detection for detecting large-scale columns captured in
multiple images.
x11 ;y11 ; 1
x12 ;y
22 ; 1
::::::::::::
x1n ;y1n ; 1
0BBBBB@
1CCCCCA
a11 ; a21a12 ; a22a13 ; a23
0@
1A =
x21 ;y21
x22 ;y
22
:::; :::
x2n ;y2n
0BBBBB@
1CCCCCA
1
3. Proposed method
A novel method of detecting a large-scale bridge column captured
in multiple close range images is proposed in order to overcome the
1050 Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
8/12/2019 1-s2.0-S0926580510001147-main
5/9
aforementioned limitations characteristic of existing automated
bridge inspection procedures (Fig. 3). First, the pictures depicting
different segments of a single bridge column are stitched together.
Then, using the Canny operator and the Hough Transform, the column
boundaries (i.e. a pair of long vertical lines) are located. At this point,
the visual features of the image region within the column boundaries
are input into an articial neural network which performs the
concrete material recognition and results in the complete detection
of the concrete column.For the purpose of the method discussed in this paper, the rst
step, image stitching, is performed based on the work of Brown and
Lowe[60].Prior to performing any image stitching, the image noise,
which is always produced duringthe process of capturing an image, is
reduced to minimize its effect on all subsequent image processing
operations. A 5 pixel by 5 pixel median lter is selected for this
purpose, due to its ability to signicantly improve the image signal to
noise ratio [65]. Then, by matching theresulting detected features,the
image transformation matrices can be calculated. These matrices are
used to form one single image from the multiple images containing
different segments of the column. Specically, the RANSAC algorithm
[66] is used to nd feature matching outliers and inliers, while the
bundle adjustment algorithm [67] is adopted to estimate the
transformation matrix elements for all images. Image stitching errors
such as visible image edges and radial distortions are removed
through a multi-band blending algorithm.
Once a stitched image is retrieved, the column needs to be located
within this image. The work of concrete column detection consists of
identifying long near-vertical lines for each column surface as well as
determining whether or not the material contained between two lines
consists of concrete. In order to detect the column boundaries, the
Canny operator is used to produce an edge map, a binary image
composed of edge points (marked as white) and non-edge points
(marked as black), distinguishing the edge of the column from its
surroundings in the rest of the image. The Hough transform is then
used to analyze the distribution of edge points and retrieve long
vertical line information in the edge map. Each retrieved vertical line
is compared to its neighboring vertical lines. If two vertical lines have
similar size, they are regarded as a pair. This comparison is performed
iteratively, till no further pair can be found. In order to eliminate line
pairs not corresponding to actual columns, the following two visual
characteristics of columns are used to narrow the retrieved line pairs.
First, the aspect ratio (length/width) of the rectangle formed by each
long vertical line pair should be at least smaller than one. Second,concrete column surface is supposed to have the uniform texture and
color patterns. Specically, a bounding rectangle for each pair of
vertical lines is rst constructed based on their Cartesian coordinates
(x,y), and the aspect ratio (length/width) of each bounding rectangle
is calculated in order to determineif theboundingrectangle is in fact a
column. Considering the large-scale bridge columns investigated in
this research, the ratio can be assumed to be on the order of six or
seven.
Following the column boundary identication, it is necessary to
determine the properties of the material within the boundary. Thus,
theimageregion containedin theidentied boundary is retrieved and
its visual features (e.g. normalized red, normalized green and
normalized blue) are calculated. The columns visible in the stitched
images can be detected using the authors' previous work in detecting
concrete columns [39]; however, in order to further reduce human
intervention in dening appropriate threshold values for concrete
material recognition, an articial neuralnetwork is used to classify the
type of material within the boundary. The network design of the
network is composed of three layers (input, hidden and output). In
the input layer, the network receives the visual features and transfers
them to the hidden layer. For each neuron node in the hidden layer,
individual material classication boundary lines in the feature space
are created. Then, these material classication boundary lines are
combined together using an ANDoperation in theoutput layer to form
a concrete material classication model to determine whether the
region is composed of concrete or not. The classier mainly has two
important parameters. One is the number of neuron-nodes in the
hidden layer. Technically, the larger the number of neuron nodes in
the hidden layer, the larger the number of classication boundariescan be found, which results in a more precise classication
approximation. Another important parameter is related to the
classier training. The network is trained using supervised learning
(back-propagation). The output values of the network are manually
specied as 1 for concrete samples and1 for non-concrete samples.
During the training, the network automatically adjusts the neurons'
weights to minimize identication errors. Although it is possible to
repeat the training process millions of time to retrieve minimum
identication errors, long training will result in an overtraining
problem, which makes the network only memorize the peculiar
patterns of the training samples but fail to generalize them. The exact
training repetition number is determined experimentally from the
authors' previous work [68]. Specically, 160 concrete and non-
concrete samples are retrieved from real construction site images andthey are divided into two sets: one for the classier training and the
other for the classier testing. The training is repeated from 100 to
400,000 times, and the corresponding training and testing accuracy
rates (i.e. the number of corrected classied samples over the number
of samples) are recorded. It is found that the accuracy for training
samples gradually increases with the increase of repetition training
times, however, the accuracy for test samples increases after reaching
a certain limit (9000) and then drops signicantly due to the over-
tting problem[68].
When the material contained in the region is identied as
concrete, one column surface is assumed to be detected. Thus, the
detection of the concrete column is performed with the combination
of the identication of the column's boundaries and the classication
of the concrete material.Fig. 3. Method overview.
1051Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
http://localhost/var/www/apps/conversion/tmp/scratch_5/image%20of%20Fig.%E0%B3%808/12/2019 1-s2.0-S0926580510001147-main
6/9
8/12/2019 1-s2.0-S0926580510001147-main
7/9
detected concrete column surfaces (TP+ FP), and the number of
actual concrete column surfaces (TP+ FN) in each image. This
information is used to calculate the detection precision and recall.
Table 1 illustrates the average detection precision (89.7%) and
recall (84.3%) for 114 concrete bridge column images stitched from
373 original images. The precision and recall for each single stitched
example are not used because they do not truly reect the overall
performance of the method. Large-scale concrete column images are
usually taken close to the columns in order to capture the detailed
defect information. Only one column (one or two column surfaces) isvisible in each example of the stitched images. As a result, the
prevision andrecall ratios vary signicantly case by case. Forexample,
suppose there aretwo visible columnsurfacesin onestitchedimage. If
both surfaces can be detected, the precision and recall for the image
are 100% and 100%. However, if one column surface is not correctly
detected, the precision will drop to 50%. Similarly, if one column
surface fails to be detected at all, the recall will drop to 50%.
5. Conclusion and future research
The limited FOV of a camera makes it not always possible to have a
large-scale bridge element t entirely in a view, and at the same time
capture all its surface defects of interest in detail. This limitation can
be overcome with the recent development of image stitchingtechniques in the eld of computer vision. The paper presents a
novel method of detecting large-scale concrete bridge columns from
digital images. The method starts with image stitching techniques to
combine multiple images containing different FOVs for one bridge
column into one stitched image. Then, the Canny operator and the
Hough Transform are used to recognize the pair of long vertical lines
of a concrete column in the stitched image. The visual features of the
region of the line pair are then presented to an articial neural
network to indicate whether the region contains concrete or not, so
that a large-scale concrete column surface can be detected.
The work presented in this paper provides a quantitative, quick,
and thus less costly means to assist in current bridge inspection
practices. Real concrete bridge images were used to validate the work
presented in this paper. The test results indicate that a large-scale
Fig. 5.Examples of image stitching and column detection: (a) the succeeded detection; (b) the failed detection.
Table 1
Precision and recall for 114 concrete bridge column images.
Summarization
Total TP 166
Total FP 19
Total FN 31
Average precision (TP/(TP +FP)) 89.7%
Average recall (TP/(TP + FN)) 84.3%
1053Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
http://localhost/var/www/apps/conversion/tmp/scratch_5/image%20of%20Fig.%E0%B5%808/12/2019 1-s2.0-S0926580510001147-main
8/9
bridge concrete column captured in multiple images can be detected
in the stitched image with high detection precision (89.7%) and recall
(84.3%). Future work will focus on two aspects: the validity of the
alteration to the proposed method which addresses curve edges, and
the detection and spatial correlation of the defects on the column
surfaces with the surfaces themselves.
Acknowledgement
Theauthors would like to thank theNational Science Foundationfor its
generous support of this project through Grants #0933931 and 0904109.
References
[1] Bureau of Transportation Statistics (BTS), Condition of U.S. highway bridges:19902008, Sept. 10, 2009 bhttp://www.bts.govN.
[2] R. Kirk, W. Mallett, Highway bridges: conditions and the federal/state role, 2007CRS report for congress, b http://www.fas.org/sgp/crs/homesec/RL34127.pdfN(April, 2010).
[3] USA Today, State-by-state update on bridge conditions, 2008 bhttp://www.usatoday.com/news/nation/2008-07-25-bridge-chart_N.htmN(April, 2010).
[4] American Association of State and Highway Transportation Ofcials (AASHTO),Manual for condition evaluation of bridges, 2nd Ed, 2001 Washington, D.C.
[5] Federal Highway Administration (FHWA), National bridge inspection standards,Federal Register 69 (239) (2004) 7441974439.
[6] PONTIS, Pontis Product Support Website, 2009 bhttp://www.pontis.bakerpro-jects.com/N(Sept. 9, 2009).
[7] B. Phares, G. Washer, D. Rolander, B. Graybeal, M. Moore, Routine highway bridgeinspection condition documentation accuracy and reliability, Journal of BridgeEngineering 9 (4) (2004) 403413.
[8] M. Moore, B. Phares, B. Graybeal, D. Rolander, G.A. Washer, Reliability of visualinspection for highway bridges, Volume I: Final report and, Volume II:Appendices, U.S. Department Of Transportation, Washington, D.C, 2001 FHWA-RD-01-020(021),.
[9] D. Jauregui, K. White, Implementation of virtual reality in routine bridgeinspection, Journal of Transportation Research Record 1827 (2003) 2935.
[10] T. Koglin, Movable bridgeengineering,John Wiley& Sons,Inc978-0-471-41960-0, 2007.[11] NJDOT, Bridge Inspection Work Zone, Setup Guide, 2009 bhttp://www.state.nj.us/
transportation/eng/structeval/pdf/InspectionWorkZoneSetup.pdfN(Sept. 19, 2009).[12] 12, Transportation Pooled Fund Program (TPFP, TPF Studies: Support of the
Transportation Curriculum Coordination Council, Federal Highway Agencies,2009 last visit: http://www.pooledfund.org/projectdetails.asp?id=435&sta-tus=4, (Sept. 3, 2009).
[13] E.J. Jaselskis, Z. Gao, R. Walters, Improving transportation projects using laserscanning, Journal of Construction Engineering and Management, ASCE 131 (3)(2005) 377384.
[14] I. Abdel-Qader, S. Pashaie-Rad, O. Abudayyeh, S. Yehia, PCA-based algorithm forunsupervised bridge crack detection, Advances in Engineering Software 37 (12)(2006) 771778.
[15] I. Abdel-Qader, O. Abudayyeh, M. Kelly, Analysis of edge-detection techniques forcrack identication in bridges, Journal of Computing in Civil Engineering 17 (4)(2003) 255263.
[16] I. Brilakis, L. Soibelman, Y. Shinagawa, Construction site image retrieval basedon material cluster recognition, Journal of Advanced Engineering Informatics,Volume 20, Issue 4, Elsevier Science, 2006, pp. 443 452, October 2006.
[17] B. Freeman, F. Durand, Panoramas, 2006 Class lecture notes, bhttp://www.groups.csail.mit.edu/graphics/classes/CompPhoto06/html/lecturenotes/15_pano_6.pdfN(Sept. 22, 2009).
[18] Z. Zhu, I. Brilakis, Detecting air pockets for architectural concrete qualityassessment using visual sensing special issue in sensors in construction andinfrastructure management, Journal of Information Technology in Construction13 (2008) 86102.
[19] K.Kolajczyk,C. Schmid,A performanceevaluation oflocaldescriptors,IEEETransactionson Pattern Analysis and Machine Intelligence 27 (2005) 16151630.
[20] J. Bauer, N. Sunderhauf, P. Protzel, Comparing several implementations of tworecently published feature detectors, Proc. of the International Conference onIntelligent and Autonomous Systems, Toulouse, France, 2007.
[21] [21] Sunkpho, J. (2001). A Framework for Developing Field Inspection SupportSystems. Ph.D Disseration, Department of Civil and Environmental Engineering,Carnegie Mellon University.
[22] Federal Highway Administration (FWHA), Recording and coding guide for thestructure inventory and appraisal of the nation's bridges, U.S. Department ofTransportation, Washington, D.C., 1995
[23] P. Thompson, R. Shepard, AASHTO Commonly Recognized Bridge Elements,Successful Applications and Lessons Learned National Workshop on Com-monly Recognized Measures for Maintenance, 2000 bhttp://www.pdth.com/images/coreelem.pdfN(Sept. 21, 2009).
[24] E. Wasserman, Are Tennessee Bridges Safe?, 2007 bhttp://www.tdot.state.tn.us/documents/Wasserman_editorial.pdfN(Sep. 21, 2009).
[25] Federal Highway Administration (FHWA), Questions and Answers on theNational Bridge Inspection Standards 23 CFR 650 subpart C, U.S. Department
of Transportation, 2009 bhttp://www.fhwa.dot.gov/Bridge/nbis/N(Sept. 29,2009).
[26] T. Ditto, Acquisition of three-dimensional surfaces using holographic primaryobjectives, 2007 last visit: http://www.drillamerica.com/PDF/3D.pdf.
[27] G. Sansoni, M. Trebeschi, F. Docchio, State-of-the-art and applications of 3Dimaging sensors in industry, cultural heritage, medicine, and criminal investiga-tion, Sensors 9 (2009) 568601.
[28] P. Tang, B. Akinci, Extracting surveying goals from point clouds to supportconstruction and infrastructure inspection, Construction Research Congress 2009,Seattle, USA, 2009.
[29] Z. Zhu, I. Brilakis, Comparison of civil infrastructure optical-based spatial data
acquisition techniques, Journal of Computing in Civil Engineering 23 (3) (2009)170177.[30] D. Jauregui, Y. Tian, R. Jiang, Photogrammetry Applications in Routine Bridge
Inspection and Historic Bridge Documentation, 2006 Report NM04STR-01,bh t t p : / / w w w . nm s h t d . s t a t e . n m . us / u p l o a d / i m a g e s / Re s e a r c h / NM04STR01PhotogrammetryApplicationsHistoricBridges_2006.pdf N (Sept.18, 2009).
[31] S. Sinha, P. Fieguth,Automated detectionof cracks in buried concretepipe images,Automation In Construction 15 (1) (2006) 5872.
[32] S.-N. Yu, J.-H. Jang, C.-S. Han, Auto inspection system using a mobile robot fordetecting concrete cracks in a tunnel, Automation in Construction 16 (3) (2007)255261.
[33] F. Ge, T. Liu, S. Wang, J. Stahl, Template-based object detection through partialshape matching and boundary verication, International Journal of SignalProcessing 4 (1) (2008) 148157.
[34] S. Stiene, K. Lingemann, A. Nuchter, J. Hertzberg, Contour-based object detection inrange images, Data Processing, Visualization, and Transmission, 2006, pp. 168175.
[35] D. Wang, J. Zhang, H. Wong, Y. Li, 3D model retrieval based on multi-shellextended Gaussian image, Lecture Notes in Computer Science: Advances in Visual
Information Systems, 2007, pp. 426437.[36] M. Kazhdan, T. Funkhouser, S. Rusinkiewicz, Rotation invariant spherical
harmonic representation of 3D Shape descriptors, Symposium on GeometryProcessing, 2003, pp. 156164.
[37] Y. Shan, H. Sawhney, B. Matei, R. Kumar, Shapeme histogram projection andmatching for partial object recognition, IEEE Transactions on Pattern Analysis andMachine Intelligence 28 (4) (2006) 568577.
[38] D. Huber, A. Kapuria, R. Donamukkala, M. Hebert, Parts-based 3D objectrecognition, Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, vol.2, 2004, pp. 8289.
[39] A. Patterson IV, P. Mordohai, K. Daniilidis, Object detection from large-scale 3Ddatasets using bottom-up and top-down descriptors, European Conference onComputer Vision, 2008, pp. 553566.
[40] F. Bosche, C. Haas, Towards automated retrieval of 3D designed data in 3D senseddata,Proceedings of the2007 ASCEInternational Workshopon Computing in CivilEngineering. Pittsburg, PA, 2007.
[41] F. Bosche, C. Haas, Automated retrieval of 3D CAD model objects in constructionrange images, Automation in Construction 17 (2008) 499512.
[42] D. Lowe, Distinctive image features from scale-invariant keypoints, InternationalJournal of Computer Vision 60 (2) (2004) 91110.
[43] S. Ruiz-Correa, L.G. Shapiro, M. Meila, G. Berson, M.L. Cunningham, R.W. Sze,Symbolic signatures for deformable shapes, IEEE Transactions on Pattern Analysisand Machine Intelligence 28 (1) (2006) 7590.
[44] Zhu, Z. and Brilakis, I. (2009) "Concrete Column Recognition in Images andVideos", Journal of Computing in Civil Engineering, American Society of CivilEngineers (in review).
[45] J. Neto, D. Arditi, M. Evens, Using colors to detect structural components indigital pictures, Computer Aided Civil and Infrastructure Engineering 17(2002) 6176.
[46] I. Brilakis, L. Soibelman, Y. Shinagawa, Material-based construction site imageretrieval, Journal of Computing in Civil Engineering, Volume 19, Issue 4, AmericanSociety of Civil Engineers, October 2005, pp. 341355.
[47] I. Brilakis, L. Soibelman, Shape-based retrieval of construction site photographs,Journal of Computing in Civil Engineering, Volume 22, Issue 1, American SocietyofCivil Engineers, January/February 2008, pp. 1420.
[48] L.A. Fernandes, M.M. Oliveira, Real-time line detection through an improvedhough transform voting scheme, Pattern Recognition 41 (1) (2008) 299314.
[49] D. Guru, B. Shekar, P. Nagabhushan, A simple and robust line detectionalgorithm based on small eigenvalue analysis, Pattern Recognition Letters 25(2004) 113.
[50] Y. Lee, H. Koo, C. Jeong, A straight line detection using principle componentanalysis, Pattern Recognition Letter 27 (2006) 17441754.
[51] P. David, D. DeMenthone, Object recognition in high clutter images using linefeatures, Proceedings of the 10th IEEE International Conference on ComputerVision, Vol. 2, Issue 17-21, 2005, pp. 15811588.
[52] C. Nikolaos, E. Anagnostopoulos, I. Anagnostopoulos, V. Loumos, E. Kayafas, Alicense plate-recognition algorithm for intelligent transportation systemapplications, IEEE Transactions on Intelligent Transportation Systems 7 (3)(2006) 377392.
[53] Canon PhotoStitch Sofware, bhttp://www.kenrockwell.com/canon/photostitch.htmN(Nov. 02, 2009).
[54] Realviz, bhttp://www.realviz.comN(Nov. 02, 2009).[55] Y. Chuang, Image Stitching,Nov. 01, 2009 LectureNotes, bhttp://www.csie.ntu.edu.tw/
~cyy/courses/vfx/05spring/lectures/handouts/lec06_stitching_4up.pdfN.[56] R. Szeliski, Image Alignment and Stitching: A Tutorial, 2005 Technical Report,
MSR-TR-2004-92.
1054 Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
http://www.bts.gov/http://www.fas.org/sgp/crs/homesec/RL34127.pdfhttp://www.usatoday.com/news/nation/2008-07-25-bridge-chart_N.htmhttp://www.usatoday.com/news/nation/2008-07-25-bridge-chart_N.htmhttp://www.pontis.bakerprojects.com/http://www.pontis.bakerprojects.com/http://www.state.nj.us/transportation/eng/structeval/pdf/InspectionWorkZoneSetup.pdfhttp://www.state.nj.us/transportation/eng/structeval/pdf/InspectionWorkZoneSetup.pdfhttp://www.pooledfund.org/projectdetails.asp?id=435&status=4http://www.pooledfund.org/projectdetails.asp?id=435&status=4http://www.groups.csail.mit.edu/graphics/classes/CompPhoto06/html/lecturenotes/15_pano_6.pdfhttp://www.groups.csail.mit.edu/graphics/classes/CompPhoto06/html/lecturenotes/15_pano_6.pdfhttp://www.pdth.com/images/coreelem.pdfhttp://www.pdth.com/images/coreelem.pdfhttp://www.tdot.state.tn.us/documents/Wasserman_editorial.pdfhttp://www.tdot.state.tn.us/documents/Wasserman_editorial.pdfhttp://www.fhwa.dot.gov/Bridge/nbis/http://www.drillamerica.com/PDF/3D.pdfhttp://www.nmshtd.state.nm.us/upload/images/Research/NM04STR01PhotogrammetryApplicationsHistoricBridges_2006.pdfhttp://www.nmshtd.state.nm.us/upload/images/Research/NM04STR01PhotogrammetryApplicationsHistoricBridges_2006.pdfhttp://www.kenrockwell.com/canon/photostitch.htmhttp://www.kenrockwell.com/canon/photostitch.htmhttp://www.realviz.com/http://www.csie.ntu.edu.tw/~cyy/courses/vfx/05spring/lectures/handouts/lec06_stitching_4up.pdfhttp://www.csie.ntu.edu.tw/~cyy/courses/vfx/05spring/lectures/handouts/lec06_stitching_4up.pdfhttp://www.csie.ntu.edu.tw/~cyy/courses/vfx/05spring/lectures/handouts/lec06_stitching_4up.pdfhttp://www.csie.ntu.edu.tw/~cyy/courses/vfx/05spring/lectures/handouts/lec06_stitching_4up.pdfhttp://www.realviz.com/http://www.kenrockwell.com/canon/photostitch.htmhttp://www.kenrockwell.com/canon/photostitch.htmhttp://www.nmshtd.state.nm.us/upload/images/Research/NM04STR01PhotogrammetryApplicationsHistoricBridges_2006.pdfhttp://www.nmshtd.state.nm.us/upload/images/Research/NM04STR01PhotogrammetryApplicationsHistoricBridges_2006.pdfhttp://www.drillamerica.com/PDF/3D.pdfhttp://www.fhwa.dot.gov/Bridge/nbis/http://www.tdot.state.tn.us/documents/Wasserman_editorial.pdfhttp://www.tdot.state.tn.us/documents/Wasserman_editorial.pdfhttp://www.pdth.com/images/coreelem.pdfhttp://www.pdth.com/images/coreelem.pdfhttp://www.groups.csail.mit.edu/graphics/classes/CompPhoto06/html/lecturenotes/15_pano_6.pdfhttp://www.groups.csail.mit.edu/graphics/classes/CompPhoto06/html/lecturenotes/15_pano_6.pdfhttp://www.pooledfund.org/projectdetails.asp?id=435&status=4http://www.pooledfund.org/projectdetails.asp?id=435&status=4http://www.state.nj.us/transportation/eng/structeval/pdf/InspectionWorkZoneSetup.pdfhttp://www.state.nj.us/transportation/eng/structeval/pdf/InspectionWorkZoneSetup.pdfhttp://www.pontis.bakerprojects.com/http://www.pontis.bakerprojects.com/http://www.usatoday.com/news/nation/2008-07-25-bridge-chart_N.htmhttp://www.usatoday.com/news/nation/2008-07-25-bridge-chart_N.htmhttp://www.fas.org/sgp/crs/homesec/RL34127.pdfhttp://www.bts.gov/8/12/2019 1-s2.0-S0926580510001147-main
9/9
[57] S.Baker, I. Matthews, Lucas-Kanade 20 years on:a unifyingframework:Part 1: thequantity approximated, the warp update rule, and the gradient descentapproximation, International Journal of Computer Vision 56 (3) (2004) 221255.
[58] A. Agarwala, et al., Interactive digital photomontage, ACM Transactions onGraphics 23 (3) (2004) 292300.
[59] H.-Y. Shum, R. Szeliski, Construction of panoramic mosaics with global and localalignment, International Journal of Computer Vision 36 (2) (2000) 101130.
[60] M. Brown, D. Lowe, Automaticpanoramic image stitchingusing invariant features,International Journal of Computer Vision 74 (1) (2007) 5973.
[61] M. Golparvar-Fard, F. Pea-Mora, S. Savarese, Application of D4AR a 4-dimensional augmented reality model for automating construction progress
monitoring data collection, processing and communication, ITcon Vol. 14, SpecialIssue Next Generation Construction IT: Technology Foresight, Future Studies,Roadmapping, and Scenario Planning, 2009, pp. 129153.
[62] A. Efros, Image Blending, 2008 Lecture Notes, bhttp://www.graphics.cs.cmu.edu/courses/15-463/2008_fall/Lectures/blending.pdfN(Nov. 05, 2009).
[63] M. Grundland, R. Vohra, G. Williams, N. Dodgson, Nonlinear multi-resolutionimage blending, MG&V 15 (3) (2006) 381390 (Feb. 2006).
[64] C. Allene, J. Pons, R. Keriven, Seamless image-based texture atlases using multi-band blending, Pattern Recognition, 2008. ICPR 2008. 19th InternationalConference on, 2008.
[65] M. Juneja, R. Mohana, An improved adaptive medialtering method for impulsenoise detection, International Journal of Recent Trends in Engineering 1 (1)(2009) 274278.
[66] M. Fischler, R. Bolles, Random sample consensus: a paradigm for model ttingwith application to image analysis and automated cartography, Communicationsof the ACM 24 (1981) 381395.
[67] W. Triggs, P. McLauchlan, R. Hartley, A. Fitzgibbon, Bundle adjustment: a modernsynthesis, Vision Algorithms: Theory and Practice, number 1883 in LNCS, 1999,pp. 298373.
[68] Zhu, Z. and Brilakis, I. (accepted for publication) "Intelligent Learning andDetection of Concrete Regions in Construction SiteImages"Proceedingsof the 6th
International Conference on Innovation in Architecture, Engineering andConstruction (AEC), June 911, 2010, University Park, PA.[69] G. Bradski, A. Kaehler, Learning OpenCV: Computer Vision with the OpenCV
Library, O'Reilly Media, Inc0596516134, 2008.[70] FANN, Fast Articial Neural Network Library, 2009 b http://www.leenissen.dk/
fann/N(April 10, 2009).[71] T. Wickens, Elementary Signal Detection Theory, Oxford University Press, New
York0195092503, 2002.
1055Z. Zhu et al. / Automation in Construction 19 (2010) 10471055
http://www.graphics.cs.cmu.edu/courses/15-463/2008_fall/Lectures/blending.pdfhttp://www.graphics.cs.cmu.edu/courses/15-463/2008_fall/Lectures/blending.pdfhttp://www.leenissen.dk/fann/http://www.leenissen.dk/fann/http://www.leenissen.dk/fann/http://www.leenissen.dk/fann/http://www.graphics.cs.cmu.edu/courses/15-463/2008_fall/Lectures/blending.pdfhttp://www.graphics.cs.cmu.edu/courses/15-463/2008_fall/Lectures/blending.pdf