See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/325457128

A Vibration-Based Vehicle Classification System using Distributed Optical

Sensing Technology

Article in Transportation Research Record Journal of the Transportation Research Board · May 2018
DOI: 10.1177/0361198118775840

CITATIONS READS

6 432

4 authors, including:

Hongduo Zhao Difei Wu

Tongji University Tongji University
47 PUBLICATIONS 308 CITATIONS 12 PUBLICATIONS 94 CITATIONS

SEE PROFILE SEE PROFILE

Mengyuan Zeng
Tongji University
7 PUBLICATIONS 14 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Piezoelectric Energy Harvester for Roadway Applications View project

Mechanical behavior of HMA overlay in airfield composite pavement View project

All content following this page was uploaded by Difei Wu on 06 August 2018.

The user has requested enhancement of the downloaded file.

Article

Transportation Research Record

1–12
Ó National Academy of Sciences:
A Vibration-Based Vehicle Classification Transportation Research Board 2018
Reprints and permissions:
System using Distributed Optical Sensing DOI: 10.1177/0361198118775840
journals.sagepub.com/home/trr

Technology

Hongduo Zhao1, Difei Wu1, Mengyuan Zeng1, and Sheng Zhong1

Abstract
This paper presents a vibration-based vehicle classification system using distributed optical vibration sensing (DOVS) technol-
ogy and describes a comprehensive classification method including signal processing and feature extraction. With low mainte-
nance costs, this system can collect vehicle classification data in a larger scale. At first, it utilizes an embedded sensing fiber as
a distributed sensor to collect traffic-induced vibration signals, and then extracts several features from the raw signals to esti-
mate axle configurations and identify vehicle categories. At the same time, an empirical mode decomposition (EMD)-based
method is applied to reconstruct signals for features extraction, and then several extraction algorithms are proposed to
obtain the axle configuration, moving speed, and frequency-domain feature of each vehicle. When all features are extracted,
a multi-step classifier is designed to categorize vehicles into different classes. In addition, to evaluate the classification perfor-
mance of this system, a prototype system was installed on a relief road in Shanghai, China using precast concrete pavement
technology. With an overall accuracy of 89%, the test results show a good performance of this classification system.

Vehicle classification data plays a significant role in traf- of their advantages is that they are able not only to
fic and road management. Traffic management can use obtain traffic parameters from in-pavement detectors but
vehicle type information for road toll, traffic control, and also to monitor pavement damage and loading condi-
traffic guidance. Additionally, the information of trucks tions, which is beneficial to both traffic and road man-
can be used to assess the performance of past and current agement. In the field of vibration monitoring, the
pavement as well as to predict future performance, thus accelerometer is the most commonly used detector. Some
offering important assistance to road management. accelerometer-based systems and algorithms have been
Various technologies have been developed for vehicle proposed for axle detection and vehicle classification.
classification over recent decades and these are broadly Bajwa et al developed an accelerometer-based wireless
divided into two categories: non-embedded and sensor network for vehicle classification and also pro-
embedded detectors. The non-embedded detectors, posed a novel algorithm to count axles and estimate axle
including vision-based detectors (1–3), acoustic-based spacing (8). Using accelerometers and magnetometers,
detectors (4), and laser-based detectors, are installed Ma et al developed a wireless automatic vehicle classifi-
beside or over the road. With this type of detector, pave- cation prototype system and used a filter algorithm to
ment cuts or lane closures are not required during instal- identify axles (4). Additionally, Yeum et al used a Viola-
lation and precise information about vehicle types can be Jones algorithm to extract and classify the distinctive
obtained. However, the performance of the classification dynamic patterns of different vehicles by converting the
is sensitive to environmental conditions, such as visibi- measured acceleration signals to time-frequency images
lity, noise level, and temperature. By contrast, embedded (9). Markus and Hostettler also used the frequency-
detectors such as inductive loops (5), magnetometers (6), domain features to detect and classify vehicles (10).
and piezoelectric sensors (7) are installed beneath the
pavement surface. They are not sensitive to environmen- 1
tal conditions but there is often a high cost in the instal- Key Laboratory of Road and Traffic Engineering of the Ministry of
Education, Tongji University, Shanghai, P. R. China
lation and maintenance.
Recently, vibration-based methods for traffic moni- Corresponding Author:
toring have become a more efficient and less costly. One Address correspondence to Difei Wu: [email protected]
2 Transportation Research Record 00(0)

Although the accelerometer-based systems can offer System Description

high-precision information, they are not suitable for
large-scale measurement, and may require high-cost
Operating Principle of DOVS
maintenance. The distributed optical vibration sensing- DOVS system is a distributed sensing system that can
system (DOVS), which is an emerging technology in the obtain the vibration information along the sensing fiber.
sensing field, has attracted increasing attentions in The basic principle of vibration measurement relies on
laboratory studies and engineering applications. DOVS the phase optical time-domain reflectometer (f-OTDR)
has numerous advantages such as large-scale measuring, technology, which was first proposed in 1993 (12).
resistance to damage, anti-electromagnetic interference, Figure 1 shows a typical f-OTDR system architecture.
and highly reliable performance in severe environments The system uses a laser source with narrow linewidth and
(11). Moreover, DOVS is a distributed sensing system low frequency shift to form the interference of the back-
because every single unit in the fiber can be used as a scattered light waves. The measurement trace of the sys-
sensor. Therefore, DOVS is a low-cost and easy-to- tem shows the intensities of the Rayleigh Backscattered
implement system compared with single-unit sensors, light generated at different positions in the sensing fiber.
and there is a great potential for detecting and classifying The output intensity (I) of the backscattered light is
large-scale vehicles based on DOVS. related to the optical phase in the sensing fiber (13):
This paper presents a vibration-based vehicle classifi- pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cation system with the application of DOVS technology. I = I1 (t) + I2 (t) + 2 I1 (t)I2 (t) cos (2u + u0 ) ð1Þ
This system is capable of counting the number of axles,
estimating the axle spacing, extracting frequency-domain where
features, and classifying vehicles. The remainder of this I1(t) = the intensity of the first interference light,
paper is organized as follows: the operating principle of I2(t) = the intensity of the second interference light,
DOVS and the prototype system is introduced in Section u = the optical phase of backscattered light at a certain
2 along with the installation method of the classification position,
system; the signal processing methods are proposed in u0 = the phase difference between the two interference
Section 3; a series of feature extraction approaches are lights.
proposed and described in Section 4; after the extraction When external vibration is applied at certain position,
of axle features and frequency-domain feature, a multi- the corresponding phase of backscattered light will change,
step classifier is designed and presented in Section 5; the resulting in an oscillation of the output intensity.
conclusions are presented in the last section. Therefore, the DOVS system can convert the external

Figure 1. Architecture of DOVS.

Zhao et al 3

Figure 2. (a) Test site; (b) measurement devices; (c) system set-up.

vibration to this oscillation using a coupler and a photode- The sensing fiber is buried beneath the concrete pave-
tector, and then record them by an acquisition card and a ment surface, and coated with a steel wire and a polyethy-
processor. Moreover, the position of the external vibration lene sheath, which can protect the fiber from being
is represented by the distance from the start, which is damaged during installation and operation. The vibration
determined by calculating the time delay between the and position information are acquired through the interro-
injected pulsed light and backscattered light. The position- gator and the computer, with a sampling frequency of
ing accuracy of the system (Dz), namely spatial resolution, 2000 Hz and a spatial sampling interval of 2 m. Moreover,
relates to the pulse width (Tp) of the pulsed light: the embedded sensing fiber is firstly circled to sixteen fiber
coils (30 cm diameter, four rings) with adequate spaces,
cTp and then arranged into two rows with a 4.5 m space
Dz = ð2Þ
2n (Figure 2a and c). Through increasing the layout density,
where this layout improves the accuracy of spatial localization.
c = the velocity of light in vacuum,
n = the refractive index of the sensing fiber.
System Installation
To test and verify the system, a prototype system was
System Components installed on a relief road in Shanghai to collect the
Figure 2 illustrates the test site, measurement devices and traffic-induced vibration signals. The traffic composition
a scheme of the prototype system. There are three major was complex, including cars, buses, trucks, and multi-
components in the system including sensing fiber (3 mm- axle trailers.
diameter, single mode), interrogator (BAIANTEK, The installation of conventional in-pavement sensors
FT630) and a computer. such as induction loops and magnetic sensors is
4 Transportation Research Record 00(0)

Table 1. Vehicle Classification Categories

Class Vehicle type Axle composition Number

1 Passenger car/van 247

2 Two-axle bus 32

3 Two-axle truck 119

4 Three-axle truck 10

5 Three-axle truck 1

6 Four-axle truck 10

7 Four-axle trailer 50

8 Five-axle trailer 19

9 Five-axle trailer 1

10 Six-axle trailer 3

Total 482

destructive and may decrease pavement life. For the pro- (passenger or freight transport), the descriptions and
totype system, the precast concrete pavement technology number of each category are presented in Table 1.
is utilized to avoid these problems. Firstly, the sensing
fiber was buried into a precast concrete slab when cast- Signal Processing
ing and the end of the fiber was placed at the corner of
the slab. When the slab was assembled, the end was Empirical Mode Decomposition
taken out and connected to the interrogator, which com- The raw vibration signal was complicated as it consisted
pleted the system installation. The installation proce- of noisy signal and useful signal. For axle-induced vibra-
dures were carried out in four hours. tion signals, the useful signal mainly concentrates on spe-
Additionally, to improve the positioning performance cific frequency bands while noisy signals are evenly
of the system, a 10 km-length extended fiber is utilized to distributed over frequency. To eliminate noises and
connect the sensing fiber to the interrogator (Figure 2b). extract useful signals, the empirical mode decomposition
(EMD) method is utilized for signal reconstruction. It is
a decomposition method that can decompose raw signal
Data Acquisition into several intrinsic components, namely, intrinsic mode
Vibration signals induced by over 400 vehicles were col- functions (IMFs). Each IMF represents the detail com-
lected with this system. The speeds of the vehicles ranged ponents of the raw signal of different frequency bands.
between 8 and 25 m/s. The vehicles were classified into 10 An IMF is a function that satisfies the following two
categories according to their axle features and usage requirements: (1) the number of extrema and the number
Zhao et al 5

of zero crossings must either be equal or differ at most by The residue r1(t) is considered as the new signal and
one; (2) the running mean value of the envelope defined the sifting process is applied from the beginning. As a
by the local maxima and the envelope defined by the result, the signal x(t) will be decomposed into a finite
local minima is zero. The procedures to decompose the number of IMFs Ci(t). The sifting process ends when the
raw data (f(t)) are shown as follows (14): last residue rn(t) is a constant or a monotonic function.
Firstly, the local extrema x(t) are identified. The local By summing up all the IMFs and the final residue, the
maxima are connected forming the upper envelope u(t) x(t) is written as:
while the local minima are connected forming the lower
envelope l(t). This connection is implemented by a cubic X
n
x(t) = Ci (t) + rn (t) ð6Þ
spline interpolation scheme. Then, the running mean can
i=1
be defined as:

u(t) + l(t)
m1 (t) = ð3Þ Signal Denoising and Reconstruction
2
Figure 3 illustrates the decomposition of a vibration sig-
Then, the first component h1(t) is computed by sub- nal induced by a four-axle truck and two passenger cars.
tracting m1(t) from the signal x(t): The first time-amplitude curve shows the raw signal.
Obviously, IMFs in different frequency bands show dif-
h1 (t) = x(t)  m1 (t) ð4Þ
ferent patterns: in high-frequency IMFs (IMF1;2), the
The component will be checked if it meets the two vehicle-induced signals are not significant; in IMF 3;5,
IMF requirements. If not, h1(t) is treated as the new data, the axle configuration is easy to identify. In the low-
and the steps above will be repeated until the component frequency IMFs (IMF 7;9), the vehicle-induced vibra-
becomes an IMF. Then, the first IMF is obtained, which tion is significant. However, the axle configuration can
is denoted with C1(t). hardly be identified due to the signal superposition. This
Next, we define the residue: indicates that the superposed part mainly concentrates
on the low frequency band while less of them are super-
r1 (t) = x(t)  C1 (t) ð5Þ posed in high frequency bands.

Figure 3. Empirical mode decomposition.

6 Transportation Research Record 00(0)

Although the selected IMFs contain useful signal, the where

noisy signal cannot be neglected. We use a soft threshold E(n) = the short-time energy of the nth frame,
function to eliminate the noisy part of selected IMFs: w(n) = the window function.
 Moreover, the rectangular window function is
0 jx(i)j.th selected:
ht (x) = ð7Þ
sgn(x(i))(jx(i)j  th) otherwise 
1, 0  n  M  1
where w(n) = ð11Þ
0, otherwise
ht (x) = the denoised signal.
The threshold th is calculated by where
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi M = the width of the rectangular window.
th = s 2 log N ð8Þ For the vehicles at different speeds, the width can be
adjusted for feature extraction.
Then, the reconstructed signals are reconstructed by
superposing several denoised IMFs (IMF n;m):
Features Extraction
X
m
fr(t) = h(IMFi ) ð9Þ Event Extraction
i=n
The first step in feature extraction is to identify the mov-
In this paper, three signals (fr1(t), fr2(t), fr3(t)) are ing vehicles and extract the corresponding signals from a
reconstructed for three purposes: event extraction, axle long-time series and the reconstructed signal fr1(t) is uti-
feature extraction, and frequency-spectrum analysis. For lized in this step.
event extraction, IMF 8;9 are selected to reconstruct the In the prototype system, this step was conducted every
signal; for axle feature extraction, IMF 3;5 are selected; 10 or 20 seconds. Short-time energy analysis, whose win-
for frequency-spectrum analysis, IMF 1;9 are selected. dow width is 1,000 (equals to 0.5 s), was utilized to deter-
mine whether the vehicle was present or not. Each event
was identified by deciding if the energies of continuous
Short-Time Energy Analysis time segments were significantly high compared with the
energies in ‘‘no vehicle’’ segments. Then, the correspond-
The vehicle-induced vibration signal is time-varying and
ing time series were extracted.
nonstationary in a long segment. However, signal in a
short segment is nearly stationary. Therefore, the signal
can be split into many short segments to be processed. In Axle Features Extraction
this paper, a short-time energy method is used to split
Axle features include axle number and axle spaces.
the reconstructed signal into a series of short segments.
According to Figure 3, in IMF 4;6, axles will induce
The short-time energy of time series x(m) is defined by
several separated short-time vibrations in a long-time
X
n series. Therefore, we can use short-time energy analysis
En = ½x(m)  w(n  m)2 ð10Þ to separate each axle from others, and then extract axle
m = nN + 1 features from the time–energy curve. Figure 4 presents

Figure 4. Procedures of axle identification.

Zhao et al 7

Figure 5. Time–energy distribution of different points.

the procedures of axle feature extraction. In this proce- Speed Estimation

dure, the reconstructed signal fr2(t) is utilized. The speed of the passing vehicle is an important factor in
association with the duration of the traffic-induced vibra-
tion. It can be estimated by calculating the ratio of the
Multi-Point Combination
distance between adjacent measuring positions (DL) to
During the test, it was found that the positions of pass- the time difference (T) of the vehicle passing over the two
ing wheels have a certain randomness in cross section. If positions:
the sampling points are located in the middle of the two-
wheel trajectories, the collected signal will be a super- DL
v= ð13Þ
posed result of the signals induced by left and right T
wheels, resulting in difficulties in the axle counting. As a
The time difference was derived by calculating the differ-
result, a multi-point combination method is proposed to
ence between the times at which the front axle (most of
improve the identification accuracy.
the front axles are single-axle) moved across the two
As the wheels in a single axle are symmetric, half of
measuring positions. The two times for exacting were
the sampling points were selected (eight points) to make
obtained from the maximum time series (Em) through
the analysis. Figure 5 shows the short-time energies of
searching the first peaks.
the selected eight points (a four-axle trailer, the window
width is 80); P8 locates on the edge of the slab while P1
locates in the middle of the slab; the color of each grid Axle Identification
represents the corresponding short-time energy.
Obviously, the four dark bands represent the four axles Figure 6 shows the time–energy curves (induced by
of the trailer. However, the short-time energies of several a five-axle truck and a two-axle small truck) with differ-
points are inadequate for axle identification. For exam- ent window widths, suggesting that the window width
ple, the energy of the front axle is lower at P1, P2 and will affect the axle identification. With a width of 150,
P3, and the energy of the tandem-axle is lower at P4 and the two axles of the two-axle small truck are easily sepa-
P6. To ensure the reliability of identification, the sam- rated from the time–energy curve. However, the two
pling points that are closest to the wheel path should be axles in the tandem axle of the five-axle truck cannot be
selected for axle identification. Thus, the maximum separated while they are identified with a window width
instant energy for each sampling point is selected to of 100. Therefore, the window width needs to be ade-
rebuild a new time series Em(n): quately small to reliably identify the rear axles of trucks
(tandem axles or tridem axles, whose spacing is usually
Em (n) = max (E1 (n), E2 (n)    E8 (n)) ð12Þ 1;1.5 m). In this paper, the window width M is deter-
mined by
where
E1 (n), E2 (n)    E8 (n)= time–energy curves of P1, DD
M=  fs ð14Þ
P2..P8. 2v
8 Transportation Research Record 00(0)

Figure 6. Time–energy curves with different window widths.

where systems. As vehicles of different types have different sus-

DD = the spacing of the tandem axles or tridem axles, pension systems and weights, the frequency analysis can
v = the vehicle speed, also be used to identify the type of moving vehicles.
fs= the sample frequency. To transform the time-domain signal into frequency
Then, the axles are counted by searching the peaks in domain, a short-time Fourier transform (STFT) method
the Time–energy curve, and the axle spacing can be deter- is used to describe how the amplitude of a signal is dis-
mined by calculating the product of the time between tributed over frequency. Assuming the time-domain sig-
adjacent peaks and moving speed, nal is S(t), the equation of STFT is
X
d =v  t ð15Þ St (v) = S(t)h(t  t)ejtv ð16Þ
t
Generally, vehicle axles are classified into three cate-
gories: single axle, tandem axle (two rear axles), and tri- where
dem axle (three rear axles). In China, axle spacing of the St(v) = the spectrum function,
tandem axle and tridem axle ranges between 1.0 and 1.35 h(t) = the window function,
m, while the distance between two adjacent single-axle is v = the angular frequency.
greater than 1.8 m. Moreover, for trailers, the space The frequency spectrums of different vehicles (passen-
between first two axles ranges between 3.2 and 3.6 m, ger car, bus, truck, and trailer) are presented in Figure 7.
while the space is usually below 2.0 m in four-axle trucks The speeds of the four vehicles are 15, 18, 16, 14 m/s,
(the spacing information was obtained from http:// respectively. It shows that there are significant differ-
www.chinacar.com.cn/). Consequently, axle configura- ences between the four vehicles in the frequency domain.
tions of moving vehicles are easily determined depending For passenger car and truck, the dominant frequencies
on the space calculated between adjacent axles. were mostly concentrated on the low-frequency band
Subsequently, the multi-axle trucks and trailers are iden- (below 30 Hz), while the distributions of bus-induced
tified based on axle configurations. As shown in and trailer-induced signals were wider in frequency
Figure 6, it was calculated that the axle spacing of the domain.
four-axle truck was 1.9 + 3.9 + 1.3 m, so it was deter- As the induced signals of four types of vehicle are dis-
mined that the vehicle included two single-axles and a tributed differently in the frequency domain, we
tandem axle. extracted the ‘‘weighted frequency’’ as an index to char-
acterize the amplitude distribution in the frequency
domain, which is calculated by:
Frequency-Domain Feature Extraction
Moving vehicles will not only affect the vibration fea- Ðfh
S(f )  fdf
tures in time domain, but also cause a variation of fl
frequency-domain features, such as frequency spectrum. fw = ð17Þ
Ðfh
The frequency-domain features of the pavement are S(f )df
influenced by the speeds, weights, and suspension fl
Zhao et al 9

Figure 7. Frequency spectrums of different vehicles over time.

Figure 8. Classification procedures.

where (over 40 Hz, mainly concentrate in the fourth compo-

fw = the weighted frequency, nents), we selected the first three components to calculate
S(f) = the amplitude-frequency curve, the weighted frequency, and the fh and fl were set to be 0
fh = high limiting frequency, Hz and 40 Hz, respectively.
fl = low limiting frequency.
Based on J. Hajek’s study (15), the vehicle-induced
ground vibration comprises four components: (1) funda- Vehicle Classification and Validation
mental frequency (1;2 Hz), induced by suspended mass;
(2) unsuspended mass-induced vibration (10;15 Hz); (3)
Classification Procedures
pavement unevenness-caused vibration (0;40 Hz, related As shown in Figure 8, there are three steps in the
to moving speed); and (4) free vibration of pavement designed vehicle classifier. Step 1 and Step 2 are event
structure (usually over 20 Hz, related to the structural extraction and features extraction which are described in
properties). Considering the high-frequency components Section 3. The third step is vehicle classification based on
10 Transportation Research Record 00(0)

multiple features. Heavy vehicles (multi-axle trucks and Two-Axle Vehicles Classification
trailers, Class 4;10) are classified based on the axle con- As previously described, multi-axle vehicles are classified
figurations (number of axles and axle spacing). For the based on their axle configurations. However, the two-
vehicles that possess similar axle features (vans, two-axle axle vehicles (class 1;3) cannot be identified because
buses, and two-axle truck), a multi-parameter classifier is they have similar axle configurations. The vehicles of the
designed to identify those types of vehicles, the frequency-
three classes have the same numbers of axles, and there
domain features and moving speeds are introduced.
are overlaps between their axle spacing ranges (passenger
car/van: 1.87;3.85 m; two-axle bus: 3.5;6.1m; two-axle
Axle-Based Classification and Validation truck: 3.3;6.2 m).
In Step 3, the types of multi-axle vehicles (Class 6;10) In this classifier, a SVM (Support Vector Machine)
are classified based on axle features. Firstly, multi-axle was utilized for multidimensional classification, as it is
vehicles are identified if the number of axles is greater appropriate for small-sample clustering. Three features
than three. Then, the specific types of the vehicles are were selected for classification including axle spacing,
identified based on their axle configurations. The classifi- weighted frequency (fw), and moving speed. The 381
cation results are provided in Table 2. records of two-axle vehicles are divided into two sets: a
It can be seen that the overall accuracy is 89.36%, training set (60 samples, including 20 passenger cars, 20
indicating that this prototype system shows a good per- two-axle trucks, and 20 two-axle buses) and a testing set
formance in the classification of heavy vehicles. (321 samples). A radial basis function (RBF) is utilized
Obviously, the classification of single-unit trucks (class as the kernel function:
4;6) and four-axle trailers (class 7) is more accurate than  
five-axle trailers (class 8) and six-axle trailers (Class 10),  xi  xj  2
k(xi , xj ) = exp (  ) ð18Þ
especially for those vehicles at high speed. Additionally, 2s2
it is also noted that five five-axle trailers were misidenti-
fied as four-axle trailers, because the corresponding By using Matlab 2014b and LibSVM 3.2 (16), the clas-
tridem-axles were identified as tandem-axles. The reason sifier is designed based on the training set. The perfor-
is that the axles in tridem-axle, compared with tandem- mance of the classifier is tested and validated through the
axle, have more effect on each other as there is one more testing set. The testing results are listed in Table 3.
axle, causing difficulties in identifying the middle axle of Although the overall classification accuracy was
tridem-axle. 89.41%, the classification performance of two-axle trucks

Table 2. Classification Performance of Multi-Axle Vehicles

Class Description Actual number Classified Successful classified Accuracy

4 Three-axle truck 10 10 10 100.00%

5 Three-axle truck 1 1 1 100.00%
6 Four-axle truck 10 10 10 100.00%
7 Four-axle trailer 50 51 46 92.00%
8 Five-axle trailer 19 14 14 73.68%
9 Five-axle trailer 1 1 1 100.00%
10 Six-axle trailer 3 2 2 66.67%
Overall 94 89 84 89.36%

Table 3. Classification Performance of Two-Axle Vehicles

Class Description Actual number Classified Successful classified Accuracy

Training set 1 Passenger car/van 20 22 20 100%

2 Two-axle bus 20 21 20 100%
3 Two-axle truck 20 17 17 85%
Testing set 1 Passenger car/van 214 230 206 92.38%
2 Two-axle bus 12 12 11 91.67%
3 Two-axle truck 95 79 70 73.68%
Overall 321 321 287 89.41%
Zhao et al 11

(class 5) was unsatisfactory (73.68%). In the testing set, it Sheng Zhong. All authors reviewed the results and approved
is noted that over 20 two-axle trucks were misidentified the final version of the manuscript.
as passenger cars/vans. The reason is that the frequency
spectra of trucks are strongly related to the loading con-
References
ditions. If the weight of a truck is sufficiently light and
the axle spacing is sufficiently small, the induced vibra- 1. Lee, H., and B. Coifman. Side-Fire Lidar-Based Vehicle
tion signal will behave more like a car than a truck. Classification. Transportation Research Record: Journal of
the Transportation Research Board, 2012. 2308: 173–183.
2. Yao, Z., H. Wei, Z. Li, and J. Corey. Fuzzy C -Means
Image Segmentation Approach for Axle-Based Vehicle
Conclusion Classification. Transportation Research Record: Journal of
In conclusion, this paper presents a vibration-based vehi- the Transportation Research Board, 2016. 2595: 68–77.
cle classification system using DOVS, and a prototype 3. Mimbela, L. E., and L. A. Klein. Summary of Vehicle
Detection and Surveillance Technologies Used in Intelligent
system was successfully installed and tested. The system
Transportation Systems. U.S. Federal Highway Adminis-
utilizes in-pavement sensing fiber to monitor the traffic- tration, Intelligent Transportation System, 2007.
induced vibrations with an interrogator to record the sig- 4. Ma, W., D. Xing, A. McKee, R. Bajwa, C. Flores, B.
nals. A special layout is designed to improve the spatial Fuller, and P. Varaiya. A Wireless Accelerometer-Based
localization accuracy. Automatic Vehicle Classification Prototype System. IEEE
A prototype system was implemented and tested in Transactions on Intelligent Transportation Systems, Vol. 15,
Shanghai. Using precast concrete pavement technology, No. 1, 2017, pp. 104–111.
the system was successfully installed on a concrete road. 5. Coifman, B., and S. Kim. Speed Estimation and Length
An EMD method, short-time energy analysis and STFT Based Vehicle Classification from Freeway Single-Loop
are employed for feature extraction. Based on short-time Detectors. Transportation Research Part C, Vol. 17, 2009,
energy analysis, the algorithms that estimate axle config- pp. 349–364.
6. He, Y., Y. Du, L. Sun, and Y. Wang. Improved Wave-
urations are proposed. The frequency-domain feature
form-Feature-Based Vehicle Classification Using a Single-
extraction is conducted via STFT. Point Magnetic Sensor. Journal of Advanced Transporta-
Vibration signals induced by over 400 vehicles were tion, Vol. 49, No. 5, 2015, pp. 663–682.
collected using the prototype system, and then the classi- 7. Kim, S., and B. Coifman. Evaluation of Axle-Based and
fication performance was evaluated. Although results Length-Based Vehicle Classification Stations. Transporta-
show a good classification accuracy of heavy vehicles, tion Research Record: Journal of the Transportation
the identification accuracy of tridem axles is lower. For Research Board, 2013. 2339: 1–12.
two-axle vehicles, the classification performance is good 8. Bajwa, R., R. Rajagopal, P. Varaiya, R. Kavaler, and N.
on passenger cars and buses, while the accuracy for two- Street. In-Pavement Wireless Sensor Network for Vehicle
axle truck classification is lower. Classification Categories and Subject Descriptors. Proc.,
10th International Conference on Information Processing in
Future improvements to this system and classification
Sensor Networks (IPSN), IEEE, Chicago, Ill., 2011.
method will be made through system installation and the
9. Yeum, C. M., S. J. Dyke, and R. E. B. Rovira. Accelera-
multidimensional classification method. A more reliable tion-Based Automated Vehicle Classification on Mobile
method is needed to accelerate the system installation Bridges. Computer-Aided Civil and Infrastructure Engineer-
and prevent the system from being damaged. In addition, ing, Vol. 31, 2016, pp. 813–825.
in the multidimensional classification, more features in 10. Stocker, M., P. Silvonen, M. Rönkkö, and M. Kolehmainen.
the frequency domain will be considered to improve the Detection and Classification of Vehicles by Measurement of
classification accuracy, and more data is needed to vali- Road-Pavement Vibration and by Means of Supervised
date classification performance. Machine Learning Detection and Classification of Vehicles
by Measurement of Road-Pavement Vibration and by
Means of Supervised Machine. Journal of Intelligent Trans-
Acknowledgments portation Systems, Vol. 20, No. 2, 2016, pp. 125–137.
This work was supported by the National Nature Science 11. Bao, X., and L. Chen. Recent Progress in Distributed
Foundation of China (Grant NO: 51778480) and China Fiber Optic Sensors. Sensors, Vol. 12, No. 7, 2012, pp.
Scholarship Council. 8601–8639.
12. Taylor, H. F., and C. E. Lee. Apparatus and Method for
Fiber Optic Intrusion Sensing. U.S. Patent and Trademark
Author Contributions Office, Washington, D.C., 1993.
The authors confirm contribution to the paper as follows: study 13. Shatalin, S. V., V. N. Treschikov, and A. J. Rogers. Inter-
conception and design: Hongduo Zhao, Difei Wu; data analysis ferometric Optical Time-Domain Reflectometry for Dis-
and draft manuscript preparation: Difei Wu, Mengyuan Zeng; tributed Optical-fiber Sensing. Applied Optics, Vol. 37, No.
field tests and data collection: Difei Wu, Mengyuan Zeng and 24, 1998, pp. 5600–5604.
12 Transportation Research Record 00(0)

14. Huang, N. E., Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Transportation Association of Canada, Charlottetown,
Q. Zheng, N. C. Yen, C. C. Tung, and H. H. Liu. The Canada, 2006.
Empirical Mode Decomposition and the Hilbert Spectrum 16. Chang, C. -C., and C. -J. Lin. LIBSVM – A Library for
for Nonlinear and Non-Stationary Time Series Analysis. Support Vector Machines. https://www.csie.ntu.edu.tw/
Proceedings Mathematical Physical and Engineering ;cjlin/libsvm/. 2016. Retrieved July 20, 2017.
Sciences, Vol. 454, No. 1971, 1998, pp. 903–995.
15. Hajek, J., C. Blaney, and D. Hein. Mitigation of Highway The Standing Committee on Highway Traffic Monitoring
Traffic-Induced Vibration. Proc., Annual Conference of the (ABJ35) peer-reviewed this paper (18-00201).

View publication stats