Hindawi

Journal of Advanced Transportation

Volume 2019, Article ID 1813763, 9 pages
https://doi.org/10.1155/2019/1813763

Research Article
Automatic Segmentation and Enhancement of Pavement Cracks
Based on 3D Pavement Images

Baoxian Li ,1 Kelvin C. P. Wang,1,2 Allen Zhang ,2 Yue Fei,2 and Giuseppe Sollazzo3
1
School of Civil and Environmental Engineering, Southwest Jiaotong University, Chengdu 610031, China
2
School of Civil and Environmental Engineering, Oklahoma State University, Stillwater, OK 74078, USA
3
Department of Engineering, University of Messina Vill. S. Agata, C.da di Dio, 98166 Messina, Italy

Correspondence should be addressed to Allen Zhang; [email protected]

Received 19 January 2018; Revised 7 November 2018; Accepted 15 January 2019; Published 18 February 2019

Academic Editor: Aboelmaged Noureldin

Copyright © 2019 Baoxian Li et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Pavement cracking is a significant symptom of pavement deterioration and deficiency. Conventional manual inspections of
road condition are gradually replaced by novel automated inspection systems. As a result, a great amount of pavement surface
information is digitized by these systems with a high resolution. With pavement surface data, pavement cracks can be detected
using crack detection algorithms. In this paper, a fully automated algorithm for segmenting and enhancing pavement crack is
proposed, which consists of four major procedures. First, a preprocessing procedure is employed to remove spurious noise and
rectify the original 3D pavement data. Second, crack saliency maps are segmented from 3D pavement data using steerable matched
filter bank. Third, 2D tensor voting is applied to crack saliency maps to achieve better curve continuity of crack structure and higher
accuracy. Finally, postprocessing procedures are used to remove redundant noises. The proposed procedures were evaluated over
200 asphalt pavement images with diverse cracks. The experimental results demonstrated that the proposed method showed a high
performance and could achieve average precision of 88.38%, recall of 93.15%, and F-measure of 90.68%, respectively. Accordingly,
the proposed approach can be helpful in automated pavement condition assessment.

1. Introduction three-dimension (3D) imaging technologies. The early 2D

imaging based pavement detection systems [6, 7] were devel-
Effective and efficient pavement condition assessment is oped by integrating hardware such as line-scan cameras, laser
crucial for determining pavement maintenance schedules, illumination systems and other auxiliary equipment. With
evaluating performance, planning rehabilitation, etc. Because the emergence of advanced technologies such as high-speed
pavement cracking is an important indicator of pavement and high-resolution 3D industry cameras, the pavement
deterioration and deficiency, it is widely considered as an inspection methods based on 3D scanning have attracted
integral part of regional pavement distress surveys [1]. Many more and more interests for the following reasons:
studies show that timely and accurately inspected pavement (1) The surface information in 3D images collected by
cracks can help transportation agencies reduce road mainte- advanced data acquisition systems is more accurate than
nance cost and extend pavement service life [2, 3]. those in 2D images. Figure 1 shows a comparison between
In some developing countries, pavements are mainly 2D and 3D pavement images. The 2D images collected in
investigated by human inspectors [4]. The traditional man- gray-scale formats have limited data range, while the 3D
ual pavement inspection is unsafe, time-consuming, expen- images are able to represent the actual depths of pavement
sive, and subjective. Hence, the automation in pavement surfaces.
inspection and evaluation has become increasingly popular (2) 3D pavement images have higher quality than 2D
and dominant [5]. There are two types of imaging tech- pavement images. Due to their different imaging principles,
niques extensively adopted in automated pavement data 3D pavement images are less prone to noise related to oil
collection: two-dimensional (2D) imaging technologies and stains, dark shadows, tire marks, etc. [8, 9].
2 Journal of Advanced Transportation

Figure 1: The 2D pavement image (left) and 3D pavement image (right): the up green line is the transverse profile of a patch of pavement
surface.

Figure 2: PaveVision3D Ultra System (left) and representative 3D pavement data (right).

(3) The depth information collected by 3D techniques is frequency subbands. Unfortunately, those approaches have
more helpful in analyzing cracks, textures, rutting, etc. [10– limitations in detecting discontinuous or high-curvature
12]. cracks. Currently, there are some successful applications
Due to recent developments and innovations in hardware of machine learning techniques, such as Artificial Neural
devices, laser line-scanning based techniques tend to become Network (ANN) and Support Vector Machine (SVM), in
mature enough for high-resolution 3D pavement data collec- classifying cracks on pavement surface [24].
tion. Laurent et al. [13] developed a Laser Crack Measurement In rest of this paper, firstly, the proposed method is
System (LCMS) composed of two laser profilers to acquire explained in detail. Secondly, an image library of 200 pave-
high-resolution 3D road surface data. Moreno et al. [14] ment 3D data verifies the accuracy and effectiveness of the
proposed an electric vehicle equipped with a laser scanner proposed method. Lastly, discussion and conclusions are
to achieve high density of surveyed points. Furthermore, given, respectively.
the PaveVision3D System mounted on Digital Highway Data
Vehicle (DHDV) (Figure 2) is able to obtain full-lane-scale 2. Methodology
3D data in 1-mm resolution at a highway speed up to 100 km/h
no matter during night- or day-time [15, 16]. In this paper, all the testing and validation data are 3D
Although automation in pavement data collection has pavement images collected by PaveVision3D System. Each 3D
achieved remarkable progress, automated distress detection image in size of 2048 × 4096 is able to cover roughly 2 ×
still faces great challenge due to the complexity and diversity 4 m2 surface area with 1 mm resolution. As shown in Figure 3,
of pavement surfaces [17, 18]. As a major task of distress the proposed method represents the following procedures:
survey, automated crack detection has been studied for a long (1) Preprocessing techniques are utilized to remove noises
time. Intensity-thresholding methods have been proposed and to rectify pavement data. (2) Steerable Matched Filter
to transform the pavement images into a binary domain Bank (SMFB) is applied on 3D pavement data for segmenting
such that the pavement distresses are easier to be recognized crack saliency maps. (3) 2D Tensor Voting is used to enhance
[19, 20]. However, those methods fail to handle images the crack continuity based on the crack saliency maps. (4)
with unevenly distributed illuminance. Edge detection based Postprocessing is conducted to remove false-positive errors.
methods, such as morphological filters [21] and BEMD
[22], are also introduced for pavement crack detection. 2.1. Spurious Noise Removal and 3D Pavement Data Rectifica-
Nevertheless, those methods tend to generate discontinuous tion. 3D pavement data may have noises caused by invalid
or nonintegral cracks. Wavelet-based approaches [23] have laser points and vehicle vibration or movement. Therefore,
been utilized to decompose the original data into different spurious noises removal and pavement 3D data rectification
Journal of Advanced Transportation 3

1mm-resolution Spurious noise removal

3D pavement data and data rectification

SMFB application Crack saliency map

 SMFB preparation TV application

 Calculate SMFB responses

 Adaptive thresholding Post-processing Crack map

Figure 3: Flowchart of the proposed method.

are needed at the first stage. A typical 2D Gaussian filter with In this case, the size of the filter is 3 × 3, and 𝜎 is
standard deviation 𝜎 is used for noise removal. Equation (1) equal to 2. In order to determine the presence of a spurious
gives each value of the 2D Gaussian filter at position 𝑝0 = noise at each point (𝑥, 𝑦), the following criterion is conduct-
(𝑥, 𝑦). ed:
1 −(x2 +𝑦2 )/2𝜎2
𝑔 (p0 ; 𝜎) = 𝑒 (1)
√2𝜋𝜎

{𝐹𝑉𝑎𝑙𝑢𝑒 (𝑥, 𝑦) 𝑖𝑓 𝑎𝑏𝑠 [𝑉𝑎𝑙𝑢𝑒 (𝑥, 𝑦) − 𝐹𝑉𝑎𝑙𝑢𝑒 (𝑥, 𝑦)] > 𝑡ℎ𝑟𝑒𝑠
𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦) = { (2)
𝑉𝑎𝑙𝑢𝑒 (𝑥, 𝑦) 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
{

where 𝑉𝑎𝑙𝑢𝑒(𝑥, 𝑦) is the original pixel value at the point 𝑓(𝜎, 𝜃), where 𝜃 ∈ [−𝜋/2, 𝜋/2] is the orientation of the
(𝑥, 𝑦), 𝐹𝑉𝑎𝑙𝑢𝑒(𝑥, 𝑦) is the filtered pixel value at the point filter:
(𝑥, 𝑦), and thres is a given threshold. 2 2 2

After obtaining 𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦), another big-size 2D Gaus- (𝑥2 − 𝜎2 ) 𝑒−(𝑥 +𝑦 )/2𝜎
𝑔𝑥𝑥 (𝑥, 𝑦) =
sian filter is applied to smooth the entire image. In this √2𝜋𝜎5
case, the filter size is 101 × 101, and 𝜎 is equal to 80. Let 2 2 2
𝐹𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦) be the convolved images based on 𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦); (𝑦2 − 𝜎2 ) 𝑒−(𝑥 +𝑦 )/2𝜎
𝑔𝑦𝑦 (𝑥, 𝑦) = (4)
then the rectified image will be √2𝜋𝜎5
2 2 2
𝑅𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦) = 𝐹𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦) − 𝑉𝑎𝑙𝑢𝑒󸀠 (𝑥, 𝑦) (3) 𝑥𝑦𝑒−(𝑥 +𝑦 )/2𝜎
𝑔𝑥𝑦 (𝑥, 𝑦) = 𝑔𝑦𝑥 (𝑥, 𝑦) =
√2𝜋𝜎5
where 𝑅𝑉𝑎𝑙𝑢𝑒󸀠 is the rectified pixel value at the point 𝑓 (𝜎, 𝜃) = 𝑔𝑥𝑥 cos2 𝜃 + 2𝑔𝑥𝑦 cos 𝜃 sin 𝜃 + 𝑔𝑦𝑦 sin2 𝜃 (5)
(x, y); 𝐹𝑉𝑎𝑙𝑢𝑒󸀠 is the convolved pixel value at the point
(x, y). A filter bank is generated by using Steerable Matched
Figure 4 shows sample profiles in both transverse and Filter, namely SMFB. Table 1 lists 52 components of SMFB
longitudinal directions. The top images (a) and (b) show the with different parameters, filter size, 𝜎, and 𝜃. Four different
original pavement profile. The bottom images (c) and (d) 𝜎 are assigned to consider the varying widths of cracks. The
illustrate rectified profiles based on (3). The red lines are their orientations are incremented with a fix angle interval 15∘ to
smoothed profile. capture crack segments in varying orientations. In order to
yield nearly zero responses within noncrack area, the filters
are shifted to have a zero mean. All filters in SMFB are
2.2. Steerable Matched Filter Bank (SMFB). The steerable illustrated in Figure 5.
filter introduced by Freeman and Adelson [25] is a linear Each preprocessed 3D pavement image is convolved with
combination of a few basic filters. Particularly, steerable all 52 filters in SMFB. At each pixel, only the maximum
filter is popular in crack and ridge detection due to high convolutional response is preserved as a result of SMFB
efficiency [26–28]. In this study, the SMFB method uses operation. Mathematically, (6)∼(8) give the specific calcula-
second-derivative Gaussians as basic filters. Equation (1) gives tion procedures.
the 2D Gaussian with variance 𝜎, and (4) gives its second
derivatives. Equation (5) shows the formulation of the filter 𝑟𝑖 (𝑝; 𝜎, 𝜃) = 𝑘𝑖 (𝜎, 𝜃) ⊗ 𝑉 (𝑝) (6)
4 Journal of Advanced Transportation

450 460

440 440

430 420

420 400
0 200 400 600 800 1000 0 200 400 600 800 1000
(a) (b)
20 20

10 10

0 0

−10 −10
0 200 400 600 800 1000 0 200 400 600 800 1000
(c) (d)

Figure 4: Examples of pavement 3D data rectification: (a) and (b) the original transverse profile and longitudinal profile; (c) and (d) the
corresponding transverse profile and longitudinal profile.

Table 1: Composition of the SMFB.

Filter No. Filter size 𝜎 𝜃

1-13 21×21 3 -90 ,-75 ,-60∘ ,. . ., 90∘
∘ ∘

14-26 21×21 5 -90∘ ,-75∘ ,-60∘ ,. . ., 90∘

27-39 31×31 7 -90∘ ,-75∘ ,-60∘ ,. . ., 90∘
40-52 31×31 9 -90∘ ,-75∘ ,-60∘ ,. . ., 90∘

𝑟∗ (𝑝) = max 𝑟𝑖 (𝑝; 𝜎, 𝜃) (7) as discontinued fragments. In the paper, TV is adopted to

enhance connections between crack fragments.
∗
{1 𝑖𝑓 𝑟 (𝑝) > 𝑡ℎ𝑟𝑒𝑠 A second-order symmetric positive semidefinite tensor 𝑇
𝑅∗ (𝑝) = { (8) is associated with each pixel in the crack saliency maps. T is
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 mapped to a matrix (𝑎𝑖𝑗 )2×2 , whose eigenvalues are 𝜆 1 ≥ 𝜆 2 ≥
{
where p denotes a pixel located at (𝑥, 𝑦); 𝑉(𝑝) denotes 0, and corresponding eigenvectors are ⇀ 󳨀𝑒 and ⇀
1
󳨀𝑒 . Thereby
2
the preprocessed 3D pavement data; 𝑘𝑖 (𝜎, 𝜃) denotes the the tensor can be deposed as follows:
𝑖th steerable filter in SMFB with parameters 𝜎, 𝜃; 𝑟𝑖 (𝑝; 𝜎, 𝜃)
denotes the response based on convolutional output over 𝑘𝑖 𝑇 = 𝜆 1⇀
󳨀𝑒 ⇀
󳨀 𝑇 ⇀
󳨀 ⇀ 󳨀 𝑇
1 𝑒 1 + 𝜆2 𝑒 2 𝑒 2
and 𝑉(𝑝). 𝑟∗ (𝑝) denotes the maximum response; and 𝑅∗ (𝑝) (9)
denotes the binary crack map by thresholding implement. = (𝜆 1 − 𝜆 2 )⇀
󳨀𝑒 ⇀
󳨀 𝑇 ⇀
󳨀 ⇀ 󳨀 𝑇 ⇀ 󳨀 ⇀󳨀 𝑇
1 𝑒 1 + 𝜆2 ( 𝑒 1 𝑒 1 + 𝑒 2 𝑒 2 )
As illustrated in Figure 6, crack saliency maps are gen-
The first term (𝜆 1 − 𝜆 2 )⇀󳨀𝑒 ⇀
󳨀 𝑇
erated after implementing SMFB. Due to pavement texture, 1 𝑒 1 in (9) represents a stick
some noncrack pixels have high responses, resulting in false- tensor as an elongated ellipsoid and the second is called
positive errors. In addition, some crack pixels have low or “ball tensor” as a circular disk. First, a ball voting is used
even zero response, resulting in crack discontinuity and false- to estimate the crack-curve orientation at each crack pixel
negative errors. Thus, additional procedure is needed to from the crack saliency maps, that is, each detected crack
improve the detection accuracy. pixel is initialized as a ball tensor ( 10 01 ), and noncrack pixels
do not participate in voting. The ball voting is conducted
2.3. Tensor Voting. Tensor voting (TV) is a perceptual group- by adding the fields generated by the stick tokens spanning
ing method proposed by Guy and Medioni [29]. In computer 360∘ at regular intervals. In this way, the principal direction
vision, TV is widely utilized to infer curvilinear structures at each crack pixel is found, which is set as the orientation
[30], locally link the corrupted data [31], and extract the of the stick token. Then a stick voting is applied by means of
lines and curves from noisy images [32]. It is highly possible casting the votes from each stick token to all the pixels (crack
that some cracks have weak responses to the SMFB due to pixels and noncrack pixels). For stick voting, assuming that
various reasons. Consequently, some cracks may be detected 𝑂 is the origin location and 𝑃 is the voting location as shown
Journal of Advanced Transportation 5

=3 =7

=5 =9

Filter size: 21 × 21 Filter size: 31 × 31

Figure 5: Steerable filters in SMFB.

1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Original pavement 3D image Crack saliency map

Figure 6: Examples of generated crack saliency maps.

in Figure 7, the voting field can be defined by using a decay

function:
2 2 2
𝐷𝐹 (𝑠, 𝑘, 𝜎) = 𝑒−((𝑠 +𝑐𝑘 )/𝜎 ) (10)
P
where s is the arc length from this token to a target point in the
voting field, k is the curvature, 𝜎 is the scale of voting, and 𝑐 l
is a parameter controlling the degree of decay with curvature s
defined in (11) as O 

−16 (𝜎 − 1) log (0.1) Figure 7: Votes cast by a stick tensor at the origin O.
𝑐= (11)
𝜋2
As TV is used to enhance connections between crack
fragments in this paper, after the stick voting stage, the dense [33]. Lastly, the OR operation is executed on the dense tensor
tensor map is extracted as tensor voting result, which is map and the crack saliency map. An overall illustration of our
different from the original tensor voting method presented in method is shown in Figure 8.
6 Journal of Advanced Transportation

Input tokens (sparse)

Encode

Tensor tokens (sparse)

Tensor voting

Tensor tokens (refined)

Tensor voting

Saliency tensor map Crack saliency map

(dense)
Figure 9: Examples of TV approach: the size of each patch is 700 ×
300 pixel2 . The patches in the left column are cropped from the crack
OR operation saliency maps; the patches in the right column are cropped from the
results of TV operation.

Crack map with noise

Figure 8: Overall TV approach.

In (12) and (13), TP denotes true positives; that is, pixels
labeled as crack pixels in the ground truth are correctly
recognized as crack pixels; FP denotes false positives; that
After TV operation, missing parts of the detected cracks is, pixels labeled as noncrack pixels in the ground truth are
could be retrieved, resulting in enhanced continuity of cracks. incorrectly recognized as crack pixels; FN represents false
Figure 9 provides typical examples of connecting discontin- negatives; that is, pixels labeled as crack pixels in the ground
ued parts using TV. In Figure 9, small fragments close to each truth are incorrectly detected as noncrack pixels. The ground
other are linked together as a whole part, as highlighted by the truths of cracks were obtained by two steps: at the first step,
dashed circles.. crack maps were generated automatically by applying method
proposed by Zhang [34]; at the second step, manual labeling
2.4. Postprocessing. After the TV operation, some noise pixels was used to refine the crack maps provided by the first
may still exist. The remaining noises can seriously affect the step. A pixel-to-pixel comparison is conducted during the
precision of crack detection. Hence, postprocessing is needed evaluation.
to further remove noises and refine the final detection output. A test data set consisting of two hundred 3D pavement
In this paper, all connected components less than 1000 pixels images is selected to evaluate the proposed method. This
are removed. test set covers images from different road sections, various
lighting conditions (i.e., daylight and nightlight), and diverse
3. Experimental Results and Comparison severities of cracks (i.e., low-level or no crack, medium-
level crack, and high-level crack). The computer hardware
In this paper, precision, recall, and F-measure are used to used for experiments is summarized as follows: Intel Core
evaluate the performance of the proposed method. Precision i7-6700T, 3.00 GHz CPU, and 32 GB RAM. All algorithms
measures the exactness or fidelity of detection and segmenta- are implemented in MATLAB platform. Figure 10 presents
tion, while recall describes the completeness of detection and some typical detection results using the testing images. As
segmentation. F-measure is the harmonic mean of precision shown in Figure 10, the proposed method can detect cracks
and recall, where an F-measure reaches its best value at 1 and with varying widths, severity levels and contexts. In addition,
worst at 0. The definitions of precision, recall, and F-measure typical false-positive and false-negative errors are shown
are shown in within the dashed rectangles in Figure 10(b) and within the
dashed circles in Figure 10(d), respectively.
𝑇𝑃 Accordingly, the precisions and recalls for all the selected
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (12)
𝑇𝑃 + 𝐹𝑃 images are illustrated in Figure 11. The precision fluctuates
between 84.00% and 97.00%, and the average precision is
𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = (13) 88.38%, while the recalls range from 85.00% to 99.00%, and
𝑇𝑃 + 𝐹𝑁 the average recall is 93.15%. In addition, the F-measure is
2 × 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙 between 85.00% and 97.00%, and the average F-measure is
𝐹-𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = (14) 90.68%.
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙
Journal of Advanced Transportation 7

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0

(a) (b)
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2

0 0

(c) (d)

Figure 10: Experimentation results.

Table 2: Performance comparison of different methods.

Authors Method Precision Recall F-measure

Ouma and Hahn [3] A triple-Transform 91.25% 80.42% 85.49%
Zou and Cao [18] CrackTree 79% 92% 85%
Shuai and Yang [35] Steerable Matched Filter 92.6% 85.1% 88.7%
- Proposed method 88.38% 93.15% 90.68%

100 This article proposed a novel method for segmenting crack

98 maps based on 3D pavement images. The proposed method
96 implements the SMFB operation, Tensor Voting, and pre-
Percentage (%)

94 processing as well as postprocessing procedures in a specific

92 order to detect cracks from 3D pavement images.
90 The experiment using 200 testing images demonstrated
88 that the proposed method can achieve a high level of
86 detection efficiency, quantified as average precision 88.38%,
84 average recall 93.15%, and average F-measure 90.68%. The
20 40 60 80 100 120 140 160 180 200
3D Image NO. average precision was slightly lower than the average recall,
Precision
implying that some noncrack pixels were incorrectly detected
Recall as crack pixels. A possible reason is that the edges of some
Overall Precision pavement markings present height differences similar to
Overall Recall those occurred at cracking area. The proposed method used
the same set of fixed parameters to yield similar detection
Figure 11: Precisions and recalls for 200 selected 3D images. accuracies for all 200 testing images, implying that the
proposed method has achieved an efficient generalization
over varying cracks on 3D pavement surfaces.
Many research works reported the performance of their Although the proposed method is efficient in detecting
methods for crack segmentation, as listed in Table 2, which pavement cracks, it still needs to consume roughly 10.3s per
demonstrates that the proposed method in this paper has image due to expensive computations primarily introduced
a higher performance than those using other methods. It by TV. In the future, parallel computing techniques may be
is acknowledged that the same set of images/data should considered to optimize the processing speed.
be used to compare all methods, judge their performance,
and estimate their potential. That requires direct access to Data Availability
both dataset and programs/codes/algorithms. Further efforts
taken by different related research agencies are needed to All of the data related to this paper is available for peer
create a benchmark dataset and form a comparison protocol. researchers to validate.

4. Conclusions and Future Work Conflicts of Interest

Automated pavement crack survey has drawn more and more The authors declare that there are no conflicts of interest
attentions from both researchers and transportation agencies. regarding the publication of this paper.
8 Journal of Advanced Transportation

Acknowledgments [14] F.-A. Moreno, J. Gonzalez-Jimenez, J.-L. Blanco, and A. Esteban,

“An instrumented vehicle for efficient and accurate 3d mapping
The authors would like to thank Dr. Kelvin C. P. Wang’ s team of roads,” Computer-Aided Civil and Infrastructure Engineering,
for providing pavement 3D data. Funding for this research vol. 28, no. 6, pp. 403–419, 2013.
was supported by the National Natural Science Foundation [15] K. C. P. Wang, L. LI, W. Luo, and A. Larkin, “Potential
of China (Grants nos. U1534203 and 51478398). The authors measurement of pavement surface texture based on three-
gratefully acknowledge the support. dimensional image data,” in Proceedings of the Transportation
Research Board 91st Annual Meeting, 2012.
[16] W. Luo, K. C. R. Wang, L. Li, Q. J. Li, and M. Moravec, “Surface
References drainage evaluation for rigid pavements using an inertial
[1] K. Gopalakrishnan, S. K. Khaitan, A. Choudhary, and A. measurement unit and 1-mm three-dimensional texture data,”
Agrawal, “Deep convolutional neural networks with transfer Transportation Research Record Journal of the Transportation
learning for computer vision-based data-driven pavement dis- Research Board, vol. 2457, pp. 121–128, 2014.
tress detection,” Construction and Building Materials, vol. 157, [17] A. Zhang, K. C. P. Wang, B. Li et al., “Automated pixel-
pp. 322–330, 2017. level pavement crack detection on 3D asphalt surfaces using a
deep-learning network: pixel-level pavement crack detection on
[2] N.-D. Hoang and Q.-L. Nguyen, “A novel method for asphalt
3D asphalt surfaces,” Computer-Aided Civil and Infrastructure
pavement crack classification based on image processing and
Engineering, vol. 32, no. 10, pp. 805–819, 2017.
machine learning,” Engineering with Computers, 2018.
[18] Q. Zou, Y. Cao, Q. Li, Q. Mao, and S. Wang, “CrackTree:
[3] Y. O. Ouma and M. Hahn, “Wavelet-morphology based detec-
automatic crack detection from pavement images,” Pattern
tion of incipient linear cracks in asphalt pavements from
Recognition Letters, vol. 33, no. 3, pp. 227–238, 2012.
RGB camera imagery and classification using circular Radon
transform,” Advanced Engineering Informatics, vol. 30, no. 3, pp. [19] H. D. Cheng, X. J. Shi, and C. Glazier, “Real-time image thresh-
481–499, 2016. olding based on sample space reduction and interpolation
approach,” Journal of Computing in Civil Engineering, vol. 17, no.
[4] N.-D. Hoang and Q.-L. Nguyen, “Automatic recognition of 4, pp. 264–272, 2003.
asphalt pavement cracks based on image processing and
machine learning approaches: a comparative study on classifier [20] H. Oliveira and P. L. Correia, “Automatic road crack seg-
performance,” Mathematical Problems in Engineering, vol. 2018, mentation using entropy and image dynamic thresholding,” in
Article ID 290498, 16 pages, 2018. Proceedings of the 17th European Signal Processing Conference,
pp. 622–626, August 2009.
[5] M. Gavilán, D. Balcones, O. Marcos et al., “Adaptive road crack
[21] Y. Maode, B. Shaobo, X. Kun, and H. Yuyao, “Pavement crack
detection system by pavement classification,” Sensors, vol. 11, p.
detection and analysis for high-grade highway,” in Proceedings
9628, 2011.
of the 8th International Conference on Electronic Measurement
[6] C. Koch and I. Brilakis, “Pothole detection in asphalt pavement and Instruments, 4-552, 4-548 pages, 2007.
images,” Advanced Engineering Informatics, vol. 25, no. 3, pp.
[22] A. Ayenu-Prah and N. Attoh-Okine, “Evaluating pavement
507–515, 2011.
cracks with bidimensional empirical mode decomposition,”
[7] A. Mancini, E. S. Malinverni, E. Frontoni, and P. Zingaretti, EURASIP Journal on Advances in Signal Processing, vol. 2008,
“Road pavement crack automatic detection by MMS images,” in Article ID 861701, 2008.
Proceedings of the 2013 21st Mediterranean Conference on Control [23] J. Zhou, P. S. Huang, and F. P. Chiang, “Wavelet-based pavement
and Automation, MED 2013, pp. 1589–1596, June 2013. distress detection and evaluation,” Optical Engineering, vol. 45,
[8] Y. Tsai, C. Jiang, and Z. Wang, “Pavement crack detection using pp. 409–411, 2003.
high-resolution 3D line laser imaging technology,” in Proceed- [24] M. S. Kaseko, Z.-P. Lo, and S. G. Ritchie, “Comparison of
ings of the 7th RILEM International Conference on Cracking traditional and neural classifiers for pavement-crack detection,”
in Pavements: Mechanisms, Modeling, Testing, Detection and Journal of Transportation Engineering, vol. 120, no. 4, pp. 552–
Prevention Case Histories, A. Scarpas, N. Kringos, and I. Al- 569, 1994.
Qadi, Eds., Springer, Dordrecht, Netherlands, 2012.
[25] W. T. Freeman and E. H. Adelson, “The design and use of
[9] A. Zhang, K. C. P. Wang, B. Li et al., “Automated pixel-level steerable filters,” IEEE Transactions on Pattern Analysis and
pavement crack detection on 3D asphalt surfaces using a Machine Intelligence, vol. 13, no. 9, pp. 891–906, 1991.
deep-learning network,” Computer–aided Civil & Infrastructure
[26] M. Jacob and M. Unser, “Design of steerable filters for feature
Engineering, 2017.
detection using Canny-like criteria,” IEEE Transactions on
[10] J. Chang, K. Chang, and D. Chen, Application of 3D Laser Pattern Analysis and Machine Intelligence, vol. 26, no. 8, pp.
Scanning on Measuring Pavement Roughness, 2006. 1007–1019, 2004.
[11] J. L. Vilaça, J. C. Fonseca, A. C. M. Pinho, and E. Freitas, [27] S. Li, Y. Cao, and H. Cai, “Automatic pavement-crack detection
“3D surface profile equipment for the characterization of the and segmentation based on steerable matched filtering and an
pavement texture – TexScan,” Mechatronics, vol. 20, no. 6, pp. active contour model,” Journal of Computing in Civil Engineer-
674–685, 2010. ing, vol. 31, Article ID 04017045, 2017.
[12] Q. Li, M. Yao, X. Yao et al., “A real-time 3D scanning system [28] S. Mukherjee and S. T. Acton, “Oriented filters for vessel
for pavement rutting and pothole detections,” SPIE Optical contrast enhancement with local directional evidence,” in Pro-
Engineering + Applications, 2009. ceedings of the 12th IEEE International Symposium on Biomedical
[13] J. Laurent, D. Lefebvre, and E. Samson, Development of a Imaging, ISBI 2015, pp. 503–506, April 2015.
New 3D Transverse Laser Profiling System for the Automatic [29] G. Guy and G. Medioni, “Inference of surfaces, 3D curves, and
Measurement of Road Cracks, 2008. junctions from sparse, noisy, 3D data,” in Proceedings of the
Journal of Advanced Transportation 9

International Symposium on Computer Vision, 1995, pp. 599–

604, 1997.
[30] N. Strokina, T. Kurakina, T. Eerola et al., Detection of Curvilinear
Structures by Tensor Voting Applied to Fiber Characterization,
Springer, Berlin, Germany, 2013.
[31] L. Risser, F. Plouraboué, and X. Descombes, “Gap filling of 3-
D microvascular networks by tensor voting,” IEEE Transactions
on Medical Imaging, vol. 27, no. 5, pp. 674–687, 2008.
[32] R. Moreno, L. Pizarro, B. Burgeth, J. Weickert, M. A. Garcia,
and D. Puig, “Adaptation of tensor voting to image structure
estimation,” Mathematics and Visualization, no. 202519, pp. 29–
50, 2012.
[33] G. Medioni, C. K. Tang, and M. S. Lee, “Tensor voting: theory
and applications,” in Proceedings of The Right Format of Internet
Advertising (Rfia), vol. 34, pp. 1482–1495, 2000.
[34] A. Zhang, K. C. P. Wang, R. Ji, and Q. J. Li, “Efficient system
of cracking-detection algorithms with 1-mm 3D-surface models
and performance measures,” Journal of Computing in Civil
Engineering, vol. 30, Article ID 04016020, 2016.
[35] S. Li, Y. Cao, and H. Cai, “Automatic pavement-crack detection
and segmentation based on steerable matched filtering and an
active contour model,” Journal of Computing in Civil Engineer-
ing, vol. 31, no. 5, 2017.