Professional Documents
Culture Documents
A Novel Computer Vision-Based Monitoring Methodology For
A Novel Computer Vision-Based Monitoring Methodology For
DOI: 10.1002/stc.2271
RESEARCH ARTICLE
Yue Pan1 Dalei Wang1 Xiang Shen1 Yusheng Xu2 Zichao Pan1
1
College of Civil Engineering, Tongji
University, Shanghai, China Summary
2
Department of Surveying and Due to the absence of antifatigue design under vehicle-induced aerodynamic
Geo-informatics, Tongji University,
load (VIAL) during the process of structural design, structural failures of noise
Shanghai, China
barriers erected on urban highway viaducts have become a common issue in
Correspondence China. In this paper, a novel VIAL monitoring methodology is proposed, which
Dalei Wang, College of Civil Engineering,
can achieve a remote VIAL measuring only by analysis and application of the
Tongji University. Shanghai, China.
Email: wangdalei@tongji.edu.cn data from a traffic-monitoring camera. To establish this methodology, a VIAL
determinative model and a computer vision system for measuring the vehi-
Funding information
Fundamental Research Funds for the cle running characteristics are developed. The accuracy and reliability of this
Central Universities, Grant/Award methodology have been validated by a field experiment. By comparing with the
Number: 20140794 ; National Natural
results of vehicle-induced aerodynamic pressure (VIAP) from a sensor-based
Science Foundation of China,
Grant/Award Number: 51778472 wind pressure acquisition system, it was found that the systemic root mean
square deviations of maximum and minimum values of VIAP were 12.13% and
10.10%, respectively.
K E Y WO R D S
computer vision, noise barrier, numerical simulation, vehicle identification, vehicle-induced
aerodynamic load (VIAL)
1 I N T RO DU CT ION
Recently, more and more urban highway viaducts have been constructed in China, which result in the traffic noise prob-
lems. Thus, noise barriers are widely used to reduce the traffic noise pollution. The structure of noise barrier is commonly
composed of H-shaped steel columns and plug-in screen boards, which can facilitate the installation and replacement
and lead to the problems of vehicle-induced vibration and early fatigue failure at the same time. An important reason
for this is that the vehicle-induced aerodynamic load (VIAL) on a screen board is not considered in the structural design.
For example, in Chinese specification published by Ministry of Communications of PRC,1 only the natural wind load
on a screen board is considered.2 The standard in United Kingdom to determine wind load on temporary road signs is
also “generally perceived by the industry to be unrealistic, so that no account is taken of vehicle-induced aerodynamic
forces.”3 In AASHTO, although the structural design of noise barrier is well specified, the VIAL is still not considered in
the load list.
Actually, with respect to noise barriers erected along railway lines, train-induced aerodynamic load (TIAL) has been
already fully considered. But the determinative factors of the TIAL on a noise barrier is significantly different from those
of the VIAL. A significant characteristic of TIAL is that stochastic issues almost do not exist due to the identical shape of a
train, fixed train-barrier separation distance, and constant running speed. However, for the VIAL, the traffic condition is
Struct Control Health Monit. 2018;e2271. wileyonlinelibrary.com/journal/stc © 2018 John Wiley & Sons, Ltd. 1 of 19
https://doi.org/10.1002/stc.2271
2 of 19 PAN ET AL.
very complicated because of the various vehicle type and random running location and speed of vehicle. Thus, the VIAL
in the case of random traffic flow is very difficult to be determined.
In order to study the effect of the VIAL and fatigue issues of noise barriers, it is essential to accurately determine VIAL
and obtain its characteristics via a long-term monitoring. At present, there are three major methods that can be used to
obtain VIAL. The first one is to use mathematical models to predict the VIAL on roadside objects, for example, traffic
signs4 and pedestrian barriers.5 But its applicability is limited by the hypothesis of a potential flow due to the ignorance
of turbulent effects. The second one is to use sensor-based system to acquire the wind pressure on road-placed plates,6
highway traffic signs overhead,7 or roadside noise barriers.2 At the same time, the corresponding traffic conditions are
recorded separately to establish the relationship with VIAL. This method is unsuitable for long-term VIAL monitoring
due to two main disadvantages: (a) inconvenient to implement and (b) difficult to determine stochastic traffic conditions.
The third one is numerical simulation method, which is widely used in researches on VIAL8-12 and TIAL.13-17 It can fully
simulate a specified vehicle running process. And the certain VIAL can be obtained according to the traffic conditions
such as vehicle type, running speed, and vehicle-barrier separation distance. But it is a nonadaptive method to reflect the
stochastic characteristics of VIAL in an actual traffic flow.
In order to address the disadvantages of the abovementioned methods, in this paper, a novel monitoring methodology
(shown in Figure 1) for VIAL on noise barriers is proposed, based on a computer vision system integrated with VIAL
determinative models. First, a formula (so-called VIAL determinative model) about vehicle-induced aerodynamic pres-
sure (VIAP) coefficient is developed, which can be regarded as a “database” in this methodology. It can be established
by the method of numerical simulation or on-site acquisition. Second, a computer vision system is integrated with a
traffic-monitoring camera and established advanced computer vision algorithms, for example, vehicle identification and
image processing, which is used to capture and measure the key parameters for “database” querying and VIAL calculation.
Finally, a remotely VIAL acquisition is achieved by performing the following steps:
– Use the computer vision system to identify the moving vehicle type and measure its running speed and
vehicle-barrier separation distance.
– Feed all obtained parameters to a VIAL determinative model to calculate VIAP coefficients for a noise barrier screen
board.
– Use the obtained VIAP coefficients and measured running speed to calculate the wind pressure.
Comparing with the traditional sensor-based methods, the proposed methodology can achieve a remotely VIAL
acquisition with lower cost, less maintenance, and more feasible implementation.
Based on this, in order to specify the form of this expression, a process of function fitting based on a lot of Cp data is
necessary. Usually, Cp can be obtained via an on-site test. However, due to its disadvantages, involving limited number of
measuring points, high cost, insufficiency to obtain big data, and difficulties in driving and safety problems, a numerical
simulation technique is commonly used. In this paper, based on the fluid analysis software FLUENT, a numerical sim-
ulation method is used to obtain the Cp data for the VIAL determinative model fitting. Additionally, due to the vehicle
shape is one of the sensitive parameters in VIAL calculation, a stereovision method is also integrated to generate more
refined 3D vehicle model to improve the precision of numerical simulation.
• Semiglobal matching
Semiglobal matching (SGM)19 is used to obtain dense points cloud from pairs of calibrated images for detailed surface
modeling; see the third row in Figure 2. SGM tries to find correspondences for every pixel by combining the concepts
4 of 19 PAN ET AL.
FIGURE 2 Process of vehicle mesh modeling. (a) Toyota COASTER; (b) Ford Transit; (c) Buick Lacrosse
of global and local stereo methods to achieve an accurate and efficient matching result, which is supported by a
global cost function optimized in eight path directions across the whole image. The overall algorithm considers two
main steps: matching cost calculation and pathwise aggregation. The matching cost is used to measure the similarity
between corresponding pixels, which is usually conducted by means of mutual information20 and census.21 The
pixel-wise mutual information matching cost is proved to be very suitable for the rotated or scaled images, but it does
not scale well with increasing radiometric depth because of the sparser joint histogram. By considering the local
PAN ET AL. 5 of 19
neighborhood only, census can handle the local radiometric changes well. Nevertheless, for the unrectified image
pairs, it lacks robustness due to the affine transformation of the local neighborhood. Here, the final cost function
used is 6. The optimization of this cost function is performed in 1D only, which can be regarded as a NP complete
problem and solved by dynamic programming.
∑ ∑ ∑
E(D) = (C(p, Dp ) + P1 T[|Dp − Dq ] = 1] + P2 T[|Dp − Dq | > 1]), (6)
p q∈Np q∈Np
where p and q are pixel points, Np indicates the neighbor points of p, and D is the disparity map. P1 and P2 are cost
coefficients. With respect to the aggregation of matching cost, it is computed along eight paths from all directions
and met at each pixel. Along each path, the cost information is aggregated until reaching a pixel with a certain
disparity. Then the aggregated costs of all the paths are summed together. For each pixel, the disparity with lowest
cost is selected as the final disparity, namely, the winner-take-all strategy.
• Triangulated irregular network (TIN) mesh
A TIN22 is a digital data structure used for the representation of a surface on the basis of points cloud (see the fourth
row in Figure 2), which is a vector-based representation and made up of irregularly distributed nodes and lines with
three-dimensional coordinates (x, y, and z) that arranged in a network of nonoverlapping triangles. A TIN consists of
a triangular network of vertices, known as mass points, with associated coordinates in three dimensions connected
by edges to form a triangular tessellation. Three-dimensional visualizations are readily generated by the rendering
of the triangular facets. In regions where there is little variation in surface height, the points may be widely spaced,
whereas in areas of more intense variation in 3D coordinates, the point density is increased. A TIN is typically based
on a Delaunay triangulation, which maximize the minimum angle of all the angles of the triangles in the triangu-
lation. Nevertheless, it is not possible to avoid errors (such as surface holes and mixtures with other environmental
features), due to light conditions and camera angles. Therefore, to enclose the vehicle surface, a boundary filter and
hole repairing semiautomatic algorithm are also applied for a smooth model by ignoring windscreen wipers, door
handles, and tires; see the fifth row in Figure 2.
In this paper, VisualSFM23,24 is adopted for sparse points cloud acquisition, and SURE25 combined with SGM is adopted
for dense points cloud generation. In practice, more than 100 high-resolution images with reasonable overlaps are col-
lected for each vehicle. All images are taken in a uniform illumination environment, by considering a reduction of the
impact of shadows.
FIGURE 3 Calculation domain and meshed model. (a) Calculation domain layout; (b) the mesh
FIGURE 4 Comparison of simulation results. (a) Comparison of two turbulence models. (b) Comparison of two scaled meshes
of Cp is less than 0.2%. Additionally, when the mesh is independently considered, the overall mesh scale is halved and
the Y+ values of the vehicle, the noise barrier, and the ground are adjusted in the range of 30 to 100, compared with the
sparse mesh scenario (where Y+ values range between 70 and 200). The computational results in Figure 4b show reliable
grid-independent performance by the RNG 𝜅 − 𝜀 model used in the simulation of VIAL on a noise barrier.
FIGURE 5 Comparison of simulation results with respect to field measurements. (a) Distance = 1.1 m, Toyota COASTER. (b)
Distance = 1.1 m, Ford Transit. (c) Distance = 1.1 m, Buick Lacrosse. (d) Distance = 1.6 m, Toyota COASTER. (e) Distance = 1.6 m, Ford
Transit. (f) Distance = 1.6 m, Buick Lacrosse
3 1.1 2.506 0.128 −0.188 0.105 −0.209 −0.023 −0.021 −18.1 10.9
4 1.6 1.248 0.117 −0.179 0.099 −0.195 −0.019 −0.016 −16.0 8.7
4 1.6 2.506 0.09 −0.149 0.074 −0.162 −0.015 −0.013 −17.3 8.4
5 1.1 1.248 0.094 −0.181 0.087 −0.188 −0.006 −0.007 −6.9 3.9 17.96 7.72
Car III
5 1.1 2.506 0.067 −0.157 0.053 −0.171 −0.014 −0.013 −21.4 8.6
6 1.6 1.248 0.065 −0.124 0.055 −0.134 −0.01 −0.01 −15.5 7.8
6 1.6 2.506 0.047 −0.104 0.036 −0.114 −0.011 −0.01 −23.4 9.4
Note. Car I: Toyota COASTER; Car II: Ford Transit; Car III: Buick Lacrosse. d denotes the vehicle-barrier separation distance; h denotes the height of
wind pressure measuring point.
To obtain the 1D VIAL determinative model for these three vehicles, 21 simulation cases with different running param-
eters were performed totally. Seven values of vehicle-barrier separation distance were selected, including 0.5, 1.0, 1.5, 2.0,
3.0, 4.0, and 5.0m. The corresponding Cp value of each case is plotted in Figure 6, grouped by vehicle type. Meanwhile, to
obtain the curves, MATLAB Curve Fitting Toolbox is used.
PAN ET AL. 9 of 19
FIGURE 6 Extreme values of Cp dependency curves with typical vehicle-barrier separation distances (height = 1.785 m). (a) Logistic,
Toyota COASTER. (b) Logistic, Ford Transit. (c) Logistic, Buick Lacrosse. (d) Allometric, Toyota COASTER. (e) Allometric, Ford Transit.
(f) Allometric, Buick Lacrosse
Following this, to determine the form of the VIAL determinative model, two equations (general allometric and logistic)
were selected for comparison:
Logistic:
A −A
Cpm = A2 + 1 ( )2 p , d ∈ [0.5, 5] . (10)
1 + xd
0
General allometric:
Cpm = A × X B , d ∈ [0.5, 5] . (11)
Table 2 shows values of fitting R-square for two equations. It indicates that the logistic is better than general allometric.
Furthermore, when the design code of a Chinese highway is considered, the width of a road lane typically corresponds
to 3.5m in China, and the lane closest to the noise barrier should keep at least a 0.5-m separation distance with the noise
barrier. Therefore, the valid range of the vehicle-barrier separation distance for the model should be set from 0.5 to 5m.
Thus, the form of VIAL determinative model in this paper is finally established as Equation 10.
10 of 19 PAN ET AL.
The determination of a vehicle's running characteristics involves the vehicle identification and the measurement of driv-
ing characteristics parameters. The identification approach mainly includes a machine learning algorithm and a vehicle
type identification method. The machine learning algorithm is used to establish a standard vehicle geometric feature
database, which is used as a reference for vehicle classification. This is followed by analyzing a real-time image captured
by a traffic-monitoring camera to rapidly and accurately recognize the type of vehicle. This study adopts histogram of
oriented gradients (HOG)26 for vehicle description and a random forest (RF) classifier27 for fast vehicle classification.
FIGURE 8 Process of vehicle-barrier separation distance measurement. (a) Input image. (b) Detected vehicle. (c) Distance measurement
gradients, seen in Equations 14 and 15, respectively. In practice, two filters [ − 1, 0, 1] and [1, 0, − 1]T are used to convolute
the original image.
Gx (x, 𝑦) = I(x + 1, 𝑦) − I(x − 1, 𝑦), (12)
G𝑦 (x, 𝑦) = I(x, 𝑦 + 1) − I(x, 𝑦 − 1), (13)
√
G(x, 𝑦) = Gx (x, 𝑦)2 − G𝑦 (x, 𝑦)2 , (14)
G 𝑦 (x, 𝑦)
𝛼(x, 𝑦) = tan−1 ( , (15)
Gx (x, 𝑦)
where x and y is the coordination of point. I(x, y) means the intensity of point (x, y) on image.
Second, the histogram of gradient for each cell, which contains n×n pixels, is computed by a spatial binning processing.
k bins, which means k orientations of gradient from 0◦ to 360◦ , are used to vote the gradients in a specified cell. The value
of each bin in a histogram of gradient is the summary of weighted projection gradient values of pixels. In practice, the bin
size usually is nine and the size of a cell is eight.
Third, the HOGs in a block, which contains m × m cells, are concatenated and normalized. The normalization would
not only reduce the impact of illumination and shadows but also increase the robust of HOG descriptor. In practice, the
size of block is four, and the L − 2 normalization as Equation 16 is applied.
√
v ← v∕ ||v||2 + 𝜀2 , (16)
where 𝜀 is a minimal constant coefficient to specify that the divisor is a nonzero value.
Finally, the HOG descriptor for an image is obtained by concatenating the HOG descriptors of blocks.
3.1.2 RF classifier
The RF classifier is a combination of N tree-structured classifiers, where each is created by vectors independently random-
izing sampled from the input vectors, and each decision tree votes uniformly to select the most popular class to classify
input vectors.29 The RF classifier employed in this paper consists a combination of geometric features at each node to grow
a tree. In the training, a bagging method was used for each feature combination to generate a training set by randomly
drawing N replacement examples with the same size of the original training set.30 In the classification process, if there
are M classes cn (n ∈ 1, 2, · · ·, M) of elements needed to be classified, a sample p achieved M confidence degrees after the
classification, and each confidence degree 𝑓(p,cn ) represents the possibility of the sample p belonging to a certain class cn .
As shown in Figure 7 and Equation 17, the final output fd of the random forest corresponds to the average of the results
from all the decision trees. [ ]
1 ∑
𝑓d = Arg max 𝑓i (p, cn ) . (17)
N i=1,2,· · ·,N
With respect to vehicle-barrier separation distance measurements, a measuring algorithm is developed (illustrated in
Figure 8).
Specifically, at the beginning of all, a virtual vector line (the yellow line in Figure 8) perpendicular to noise barriers'
line is asked to be set manually to confirm the measuring direction. Thus, a 2D coordinate system is established. For the
convenience of the algorithm description, there are three key points, seen in Figure 8c, in this coordinate system that
are defined:
• PO : the original point of this coordinate system and the intersection point of virtual line (the measuring direction)
and noise barriers' line;
• PS : the intersection point of virtual line and the left edge of the detected vehicle bounding box;
• PE : the first intersection point of virtual line and the vehicle edge (white pixels in binarized image).
Based on these, the algorithm can be described in four steps: A, converting the input colorful image to a gray image,
(b) using the gray image to obtain the vehicle edges by Canny edge detector,31 (c) calculating the coordination of PS and
using an iterator to look for PE starting from PS and following the direction of the virtual line, and (d) calculating the
vehicle-barrier separation distance d by Equation 18:
S √
d= · (xe − xo )2 + (𝑦e − 𝑦o )2 , (18)
N
where xo , yo is the coordination of PO ; xe , ye is the coordination of PE ; and S is the ground real length and N is the length
in pixels of the same line.
In this section, a series of field tests are reported. These tests were performed to verify the proposed methodology on the
basis of the previously mentioned computer vision methods and the VIAL determinative model of Cp extreme values,
which is formalized via the CFD numerical simulation and discussed in Section 2.
In the field experiments, a total of 23 cases were implemented and three typical vehicles were employed, namely, the
Toyota COASTER, Ford Transit, and Buick Lacrosse as shown in Figure 9. Finally, the acquisition of VIAP extreme values
on a noise barrier was achieved only by analysis and application of the data from a traffic-monitoring camera.
FIGURE 9 Vehicles used for field experiment. (a) Toyota COASTER. (b) Ford Transit. (c) Buick Lacrosse
PAN ET AL. 13 of 19
FIGURE 10 View of test road. (a) Overall view: shows the full view of the test road, and the specific noise barrier used for wind pressure
measurement is determined. (b) Monitoring view: shows the monitoring view including the GCPs, distance bars, region of interest (ROI),
and the location of specific noise barriers as well as the speed information collected by the video-monitoring system. (c) Noise barrier
structure: shows an illustration of the structure and the positions of sensors
zone. The camera was used to record the video of a vehicle's passing process and its parameters is listed in Table 3. The
electromagnetic blanket was assigned to act as the trigger to capture the frame while the vehicle is passing by the test zone.
Meanwhile, it could capture an accurate running speed of vehicles (similar to many automobile over speed monitoring
systems). Thus, the speed of a running vehicle is measured by this system in this experiment. Additionally, three colorful
straight bars were stuck on the road pavement, which can assist a driver, as much as possible, to control a vehicle passing
with a given vehicle-barrier separation distance of 0.5, 1.0, or 1.5m.
FIGURE 11 Samples of training dataset. (a) Positive data. (b) Negative data
These videos were recorded prior to the formal field tests, whereas the drivers were familiarizing themselves with the
test environment and testing the traffic-monitoring system. All the positive samples were manually labeled, and negative
samples were randomly generated at multiple scales from background portions.
Additionally, 500 and 300 samples were used for training and testing, respectively. Eighty trees were used to build the
RF classifier. Finally, the average accuracy of classification for Toyota COASTER, Ford Transit, and Buick Lacrosse were
achieved 76.2%, 72.5%, and 83.2%, respectively.
10 40 1.5 49.9 1.60 0.104 −0.203 12.226 −23.855 13.019 −22.438 −0.793 1.417 6.09% 6.32%
11 60 0.5 58.3 1.71 0.097 −0.193 15.586 −31.079 18.007 −27.939 −2.421 3.140 13.44% 11.24%
12 60 1.0 60.2 1.00 0.166 −0.279 28.350 −47.705 30.401 −46.308 −2.052 1.397 6.75% 3.02%
13 60 1.5 59.4 1.66 0.100 −0.198 16.682 −32.942 18.178 −29.562 −1.496 3.380 8.23% 11.43%
14 80 1.5 60.4 1.80 0.092 −0.187 15.867 −32.178 18.193 −29.266 −2.326 2.912 12.79% 9.95%
15 40 1.5 44.7 1.60 0.083 −0.145 7.801 −13.718 6.141 −12.813 1.660 0.905 27.03% 7.06%
Buick Lacrosse Transit
16 40 1.0 52.6 0.65 0.138 −0.255 18.050 −33.373 16.040 −33.388 2.010 −0.015 12.53% 0.04%
17 40 0.5 52.5 0.60 0.149 −0.266 19.442 −34.647 19.222 −42.199 0.220 −7.552 1.15% 17.90%
18 60 1.0 64.7 1.18 0.092 −0.179 18.218 −35.319 16.641 −37.085 1.578 −1.766 9.48% 4.76%
19 60 0.5 69.4 0.72 0.126 −0.241 28.685 −54.971 29.509 −63.667 −0.825 −8.696 2.80% 13.66%
20 60 1.5 66.1 1.77 0.081 −0.135 16.626 −27.975 16.134 −30.278 0.492 −2.303 3.05% 7.61%
21 80 1.5 75.4 1.72 0.081 −0.138 21.782 −37.130 15.837 −32.992 5.945 4.138 37.54% 12.54%
22 80 1.0 75.1 1.10 0.095 −0.187 25.370 −49.821 24.924 −52.825 0.446 −3.004 1.79% 5.69%
23 80 0.5 76.4 0.65 0.138 −0.255 38.081 −70.407 36.626 −80.609 1.455 −10.202 3.97% 12.66%
Root mean square deviation 2.03 4.36 12.13% 10.10%
Note. v denotes running speed, unit: km/h; d denotes vehicle-barrier separation distance, unit: m. Computed VIAP denotes the results of VIAP obtained by the
proposed method, and Measured VIAP denotes the results of VIAP obtained directly by on-site wind-pressure sensors
FIGURE 12 The deviation of Cp (h = 1.785m). (a) Toyota COASTER. (b) Ford Transit. (c) Buick Lacrosse
In addition, the processing time is around 210ms in vehicle identification and 78ms in distance calculation on a laptop
®
with a configuration of Intel Core(TM) i7-7700 @ 2.8 GHz and 16GB memory.
FIGURE 13 A comparison of computed vehicle-induced aerodynamic pressure (VIAP) and measured VIAP versus running speed
(h = 1.785m). (a) Toyota COASTER. (b) Ford Transit. (c) Buick Lacrosse
FIGURE 14 A comparison of computed vehicle-induced aerodynamic pressure (VIAP) and measured VIAP versus vehicle-barrier
separation distance (h = 1.785m). (a) Toyota COASTER. (b) Ford Transit. (c) Buick Lacrosse
for Case 23, it is that the vehicle moves at a the high running speed and a small vehicle-barrier separation distance, which
leads to the VIAP errors more sensitive to the measuring deviations. Thus, at the same level systemic precision, a bigger
VIAP deviation is obtained in this case. In terms of Cases 21 and 15, the main reason there is that the shape of vehicle is
small and its VIAP is also small, especially the vehicle that ran with a great vehicle-barrier separation distance or small
running speed. It leads to the denominator that is not large in VIAP calculation. Thus, the error ratio becomes very sen-
sitive to a small VIAP deviation. In summary, to obtain the better VIAP measurements by this proposed methodology, it
is very important to improve the precision of the system as much as possible; thus, there are several error resources that
should be noted.
5 CO N C LUSION S
In this paper, a novel computer vision-based monitoring methodology is proposed to obtain a VIAL on a noise barrier. As
previously mentioned, three key parameters of a vehicle (type or shape, vehicle-barrier separation distance, and running
speed) dominate the VIAL value. The computer vision technique developed in this paper was successfully used to obtain
these key parameters. This technique was then integrated with a VIAL numerical simulation on the basis of CFD method
to collect VIAP on noise barrier screens. A specific VIAL determinative model was proposed to compute extreme VIAL
values. A stereo vision technique was applied to build a detailed vehicle surface model. Vehicle detection and classification
based on a HOG descriptor and RF classifier were used to recognize running vehicles. The feasibility and accuracy of the
methodology was verified through a field experiment.
Comparing with the traditional sensor-based methods, the proposed methodology can achieve a remote VIAL acquisi-
tion only by analysis and application of the data from a traffic-monitoring camera. Under the large-scale traffic-monitoring
network in China, the characteristics of its hardware and software configurations allow this methodology to be applied
with lower cost, less maintenance, and more feasible implementation. Although this method has great application poten-
tial and social value, for the establishment of VIAL probabilistic models, long-term VIAL monitoring data are necessary,
which are also relevant to the antifatigue design of noise barrier. Thus, this method still needs to be improved to be more
general and robust.
In the future, more VIAL determinative models for various types of vehicles and noise barriers, as well as applying the
proposed methodology to a real highway bridge or viaducts, will be studied.
18 of 19 PAN ET AL.
ACKNOWLEDGMENTS
The methodology described in this paper was financially supported by the National Natural Science Foundation of China
(Grant 51778472) and the Fundamental Research Funds for the Central Universities (Grant 20140794). The authors also
would like to express their gratitude to Shanghai Zhongchi Group Co., Ltd., for their support and assistance during the
experiments.
ORCID
REFERENCES
1. Ministry of Communications of PRC. Wind-resistent design specification for highway bridges; 2004.
2. Wang D, Wang B, Chen A. Vehicle-induced aerodynamic loads on highway sound barriers part1: Field experiment. Wind Struct Int J.
2013;17(4):435-449.
3. Quinn AD, Baker CJ, Wright NG. Wind and vehicle induced forces on flat plates—Part 2: Vehicle induced force. J Wind Eng Ind Aerodyn.
2001;89(9):831-847.
4. Sanz-Andrés A, Santiago-Prowald J, Baker C, Quinn A. Vehicle-induced loads on traffic sign panels. J Wind Eng Ind Aerodyn.
2003;91(7):925-942.
5. Sanz-Andrés A, Laverón A, Baker C, Quinn A. Vehicle induced loads on pedestrian barriers. J Wind Eng Ind Aerodyn. 2004;92(5):413-426.
6. Lichtneger P, Ruck B. Full scale experiments on vehicle induced transient loads on roadside plates. J Wind Eng Ind Aerodyn.
2015;136:73-81.
7. Cali Philip M, Covert Eugene E. Experimental measurements of the loads induced on an overhead highway sign structure by
vehicle-induced gusts. J Wind Eng Ind Aerodyn. 2000;84(1):87-100.
8. Wang DL, Zheng L, Chen AR. Running cars induced wind loads on sound barrier of elevated roads. Adv Mater Res. 2011;378-379:137-142.
9. Waymel F, Monnoyer F, William-Louis MJP. Numerical simulation of the unsteady three-dimensional flow in confined domains crossed
by moving bodies. Comput Fluids. 2006;35(5):525-543.
10. Corin RJ, He L, Dominy RG. A CFD investigation into the transient aerodynamic forces on overtaking road vehicle models. J Wind Eng
Ind Aerodyn. 2008;96(8-9):1390-1411.
11. Hemida H, Baker C. Large-eddy simulation of the flow around a freight wagon subjected to a crosswind. Comput Fluids.
2010;39(10):1944-1956.
12. Wang D, Wang B, Chen A. Vehicle-induced aerodynamic loads on highway sound barriers part 2: Numerical and theoretical investigation.
Wind Struct Int J. 2013;17(5):479-494.
13. Zhang J, Lu J, Tu Y, Kong C, Ge W. Numerical analysis of high-speed train induced impulsive pressure on railway bridge's noise barrier
and its flange plate with collision-preventing wall. China Railw Sci. 2009;30(3):28-32.
14. Long L, Zhao L, Liu L. Research on the air turbulent force loaded on noise barrier caused by train. Eng Mech. 2010;27(3):246-250.
15. Zhao L, Long L, Qingyun C. Dynamic properties of noise barrier structure subjected to train-induced impulsive wind pressure. J Beijing
Univ Aeronaut Astronaut. 2009;35(4):505-508.
16. Deng L, Shi Z, Liu Z. Research on dynamic characteristic of high-speed train sound barrier. Chinese J Railw Eng. 2009;49(11):101-104.
17. Zhang L, Zhang J, Zhang W. Analysis on fluid-structure interaction vibration of high-speed train passing by sound barrier. Chinese J Dyn
Control. 2014;12(2):153-159.
18. Nistér D. Preemptive RANSAC for live structure and motion estimation. Mach Vis Appl. 2005;16(5):321-329.
19. Hirschmuller H. Accurate and efficient stereo processing by semi-global matching and mutual information. In: IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, 2005 CVPR 2005, Vol. 2; San Diego, CA, USA; 2005:807-814.
20. Ernst I, Hirschmuller H. Mutual information based semi-global stereo matching on the GPU. In: 4th International Symposium on Visual
Computing (ISVC08); 2008; Las Vegas, NV, USA:228-239.
21. Humenberger M, Engelke T, Kubinger W. A census-based stereo vision algorithm using modified semi-global matching and plane fitting
to improve matching quality. In: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops. San
Francisco, CA, USA; 2010:77-84.
22. Peucker TK, Fowler RJ, Little JJ, Mark DM. Digital Representation of Three-Dimensional Surfaces by Triangulated Irregular Networks
(TIN). REVISED; 1976.
23. Wu C. Towards linear-time incremental structure from motion. In: 2013 International Conference on 3D Vision. Seattle, Washington,
USA; 2013:127-134.
24. Wu C, Agarwal S, Curless B, Seitz SM. Multicore bundle adjustment. In: 2011 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR). Colorado Springs, Colorado, USA; 2011:3057-3064.
25. Rothermel M, Wenzel K. SURE—Photogrammetric surface reconstruction from imagery. In: Proceedings LC3D Workshop. Berlin,
Germany; 2012:1-21.
PAN ET AL. 19 of 19
26. Dalal N, Triggs B. Histograms of oriented gradients for human detection. In: IEEE Computer Society Conference on Computer Vision and
Pattern Recognition, 2005. CVPR 2005. San Diego, CA, USA; 2005:886-893.
27. Breiman L. Random forests. Mach Learn. 2001;45(1):5-32.
28. Mao L, Xie M, Huang Y, Zhang Y. Preceding vehicle detection using histograms of oriented gradients. In: 2010 International Conference
on Communications, Circuits and Systems (ICCCAS). Chengdu, China; 2010:354-358.
29. Pal M. Random forest classifier for remote sensing classification. Int J Remote Sens. 2005;26(1):217-222.
30. Breiman L. Bagging predictors. Mach Learn. 1996;24(421):123-140.
31. Canny J. A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell. 1986;8(6):679-698.
32. Feng D, Feng MQ. Computer vision for SHM of civil infrastructure: From dynamic response measurement to damage detection—A review.
Eng Struct. 2018;156:105-117.
How to cite this article: Pan Y, Wang D, Shen X, Xu Y, Pan Z. A novel computer vision-based monitoring
methodology for vehicle-induced aerodynamic load on noise barrier. Struct Control Health Monit. 2018;e2271.
https://doi.org/10.1002/stc.2271