Zurich Open Repository and

Archive
University of Zurich
Main Library
Strickhofstrasse 39
CH-8057 Zurich
www.zora.uzh.ch

Year: 2018

Assessing and mapping of road surface roughness based on GPS and

accelerometer sensors on bicycle-mounted smartphones

Zang, Kaiyue; Shen, Jie; Huang, Haosheng; Wan, Mi; Shi, Jiafeng

Abstract: The surface roughness of roads is an essential road characteristic. Due to the employed carrying
platforms (which are often cars), existing measuring methods can only be used for motorable roads. Until
now, there has been no effective method for measuring the surface roughness of un-motorable roads,
such as pedestrian and bicycle lanes. This hinders many applications related to pedestrians, cyclists
and wheelchair users. In recognizing these research gaps, this paper proposes a method for measuring
the surface roughness of pedestrian and bicycle lanes based on Global Positioning System (GPS) and
accelerometer sensors on bicycle-mounted smartphones. We focus on the International Roughness Index
(IRI), as it is the most widely used index for measuring road surface roughness. Specifically, we analyzed
a computing model of road surface roughness, derived its parameters with GPS and accelerometers on
bicycle-mounted smartphones, and proposed an algorithm to recognize potholes/humps on roads. As
a proof of concept, we implemented the proposed method in a mobile application. Three experiments
were designed to evaluate the proposed method. The results of the experiments show that the IRI
values measured by the proposed method were strongly and positively correlated with those measured by
professional instruments. Meanwhile, the proposed algorithm was able to recognize the potholes/humps
that the bicycle passed. The proposed method is useful for measuring the surface roughness of roads that
are not accessible for professional instruments, such as pedestrian and cycle lanes. This work enables us to
further study the feasibility of crowdsourcing road surface roughness with bicycle-mounted smartphones.

DOI: https://doi.org/10.3390/s18030914

Posted at the Zurich Open Repository and Archive, University of Zurich

ZORA URL: https://doi.org/10.5167/uzh-159758
Journal Article
Published Version

The following work is licensed under a Creative Commons: Attribution 4.0 International (CC BY 4.0)
License.

Originally published at:

Zang, Kaiyue; Shen, Jie; Huang, Haosheng; Wan, Mi; Shi, Jiafeng (2018). Assessing and mapping of road
surface roughness based on GPS and accelerometer sensors on bicycle-mounted smartphones. Sensors,
18(3):914.
DOI: https://doi.org/10.3390/s18030914
 sensors
Article
Assessing and Mapping of Road Surface Roughness
based on GPS and Accelerometer Sensors on
Bicycle-Mounted Smartphones
Kaiyue Zang 1,2 ID
, Jie Shen 1,2,3, *, Haosheng Huang 4, * ID
, Mi Wan 1,2 and Jiafeng Shi 1,2
1 Key Laboratory of Virtual Geographic Environment (Nanjing Normal University), Ministry of Education,
Nanjing 210023, China; kaiyuezang@163.com (K.Z.); onemefly@163.com (M.W.);
jiafeng.shi@outlook.com (J.S.)
2 School of Geography Sciences, Nanjing Normal University, Nanjing 210023, China
3 Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and
Application, Nanjing 210023, China
4 GIScience Center, University of Zurich, 8057 Zurich, Switzerland
* Correspondence: shenjie@njnu.edu.cn (J.S.); haosheng.huang@geo.uzh.ch (H.H.);
Tel.: +86-25-8589-1347 (J.S.); +41-44-63-56534 (H.H.)

Received: 7 February 2018; Accepted: 12 March 2018; Published: 19 March 2018

Abstract: The surface roughness of roads is an essential road characteristic. Due to the employed
carrying platforms (which are often cars), existing measuring methods can only be used for motorable
roads. Until now, there has been no effective method for measuring the surface roughness of
un-motorable roads, such as pedestrian and bicycle lanes. This hinders many applications related to
pedestrians, cyclists and wheelchair users. In recognizing these research gaps, this paper proposes
a method for measuring the surface roughness of pedestrian and bicycle lanes based on Global
Positioning System (GPS) and accelerometer sensors on bicycle-mounted smartphones. We focus
on the International Roughness Index (IRI), as it is the most widely used index for measuring
road surface roughness. Specifically, we analyzed a computing model of road surface roughness,
derived its parameters with GPS and accelerometers on bicycle-mounted smartphones, and proposed
an algorithm to recognize potholes/humps on roads. As a proof of concept, we implemented
the proposed method in a mobile application. Three experiments were designed to evaluate the
proposed method. The results of the experiments show that the IRI values measured by the proposed
method were strongly and positively correlated with those measured by professional instruments.
Meanwhile, the proposed algorithm was able to recognize the potholes/humps that the bicycle passed.
The proposed method is useful for measuring the surface roughness of roads that are not accessible
for professional instruments, such as pedestrian and cycle lanes. This work enables us to further
study the feasibility of crowdsourcing road surface roughness with bicycle-mounted smartphones.

Keywords: road surface roughness; IRI; bicycle-mounted smartphone; accelerometer; GPS

1. Introduction
The surface roughness of roads is an important road characteristic, and it is closely related to
driving safety, passenger comfort, road maintenance and so on. With the increase of road mileage
all over the world, the quality of roads has become an important concern. Among different road
quality indicators, road roughness is probably one of the most important ones [1,2]. No matter
which mode of transportation users take, bumpy roads will significantly increase the difficulty
of travel and the potholes/humps of a road will even hinder the progress of vehicles and road
users. Moreover, bumpy roads significantly reduce the lifespan of the roads themselves [1].

Sensors 2018, 18, 914; doi:10.3390/s18030914 www.mdpi.com/journal/sensors

Sensors 2018, 18, 914 2 of 17

Therefore, recently, lots of research interest has been centered on developing methods for collecting
information regarding pavement [3–5].
While existing measuring methods, such as vehicle-mounted 3D LiDAR measuring systems,
can provide very accurate information on road surface roughness, they are often very expensive
and restricted to professional applications [1]. There are also several studies that have used
low-cost accelerometers and other sensors available on smartphones to measure road surface
roughness, mainly using cars as the carrying platforms [6–8]. However, due to the employed vehicles,
these systems can only be used on motorable roads. Until now, there has been no effective method of
measuring the road surface roughness of pedestrian and bicycle lanes, which hinders many potential
applications related to pedestrians, wheelchair users and cyclists, such as navigation and routing [9–11].
At present, smartphones are integrating more and more sensors (e.g., Global Positioning System
(GPS) and accelerometers). These sensors can potentially provide some necessary data to estimate
road surface roughness. For example, the most commonly used measure of road surface roughness
is the International Roughness Index (IRI) [1,12]. To compute IRI values, the accumulative vertical
displacement value (resulting from potholes and humps on the road) and the overall travel distance
are needed. The accumulative vertical displacement value can be derived using accelerometer sensors,
while the overall travel distance can be obtained using GPS sensors. These two kinds of sensors
are available in most of the current smartphones. This makes smartphones promising devices for
measuring road surface roughness. Meanwhile, due to their rather rigid structure, bicycles are often
very sensitive to the ups and downs of roads. Mounting smartphones on bicycles would potentially
allow users to map road surface roughness when they travel around, especially on bicycle and
pedestrian lanes. This will become a more promising option as cycling in cities becomes more and
more popular, which will potentially allow large-scale crowdsourcing of road surface roughness,
similar to the famous volunteered geographic information (VGI) project OpenStreetMap (https://
www.openstreetmap.org/).
In recognizing the research gaps and the high potential of smartphones, this paper aims to propose
a method for measuring the road surface roughness of bicycle and pedestrian lanes based on GPS and
accelerometer sensors on bicycle-mounted smartphones. We mainly focus on IRI. Firstly, we analyzed
the computing model of road surface roughness and derived the parameters of the IRI model with
a bicycle-mounted GPS and accelerometer on a smartphone. An algorithm was then proposed to
identify potholes and humps on roads. As a proof of concept, an iOS mobile application “RoadSR”
was developed based on the proposed method. Furthermore, three experiments were designed to
evaluate the proposed method. We demonstrate that the IRI values measured by the proposed method
strongly and positively correlate with those measured by professional instruments.
The rest of this article is organized as follows: Section 2 presents related work. In Section 3,
we describe the proposed method. Section 4 reports the proof of concept and the evaluation,
and Section 5 discusses the results. We draw conclusions and future work in Section 6.

2. Related Work

2.1. Road Surface Roughness: Definitions and Evaluation Metrics

Road surface roughness can be understood as an expression of the deviation of the road surface
from a true planar surface [13]. It is an important road characteristic because it affects not only riding
quality, but also safety, road service life, fuel consumption and maintenance costs [1]. It is often used
interchangeably with smoothness. Road surface roughness cannot be measured directly; it requires
calculations and deductions based on measurements.
The statistical analysis of road roughness began in the fifties in the twentieth century.
In approximately 1950, the American Association of State Highway and Transportation Officials
(AASHTO) conducted a subjective evaluation of road roughness [14]. In 1966, Elson Spangler
and William Kelly from General Electric developed a first objective measurement method using
Sensors 2018, 18, 914 3 of 17

a longitudinal inertial measuring instrument. The underlying theory behind this method is still in use
today [15].
Different evaluation metrics have been proposed to indicate or report road surface roughness,
such as theSensors
International
2018, 18, x FORRoughness
PEER REVIEW Index (IRI), Present Serviceability Rating (PSR), Root Mean
3 of 17 Square
Vertical Acceleration (RMSVA), Mean Panel Rating (MPR), Ride Number (RN), Slope Variance (SV),
longitudinal inertial measuring instrument. The underlying theory behind this method is still in use
and Profile Index (PI) [1,16]. Among them, IRI, which was developed by the World Bank in the
today [15].
1980s [12], is currently the most
Different evaluation widely
metrics haveused
been indicator
proposed toof road or
indicate surface roughness,
report road because it can be
surface roughness,
transformedsuchtoasother metrics. It Roughness
the International is computed Indexusing
(IRI), “a mathematical
Present model
Serviceability Ratingknown as the
(PSR), Root Mean ‘quarter-car
Squarerepresents
model’, which Vertical Acceleration
the way (RMSVA), Meansystem
a single tire Panel Rating (MPR),ofRide
(a quarter Number
a car) (RN), Slope
is affected by the profile
Variance (SV), and Profile Index (PI) [1,16]. Among them, IRI, which was developed by the World
of the pavement [/road]” [1]. The commonly recommended units are meters per kilometer (m/km)
Bank in the 1980s [12], is currently the most widely used indicator of road surface roughness,
or millimeters
because per meter
it can (mm/m) [17].
be transformed Based
to other onItIRI
metrics. measurements,
is computed several other
using “a mathematical modelindices
knownhave been
proposed,assuch as the Mean
the ‘quarter-car Ride
model’, Index
which (MRI) the
represents andwaythea Half-Car Ride Index
single tire system (HRI),
(a quarter of a which
car) is have no
affected by the
essential differences to profile
IRI. Inofthis
the pavement
research,[/road]”
we used [1]. The
the commonly recommended
IRI to represent road units
surfaceare meters
roughness.
per kilometer (m/km) or millimeters per meter (mm/m) [17]. Based on IRI measurements, several
other indices
2.2. Measurement have been
Methods proposed,
of Road such
Surface as the Mean Ride Index (MRI) and the Half-Car Ride Index
Roughness
(HRI), which have no essential differences to IRI. In this research, we used the IRI to represent road
surface of
A number roughness.
different professional methods and devices have been used over the years for
measuring road surface roughness [1,3,4], including Rod and Level, Straight Edge, Profilograph,
2.2. Measurement Methods of Road Surface Roughness
Response-Type Instruments, Walking Profilers, and Inertial Profilers. Inertial Profilers are the most
A number of different professional methods and devices have been used over the years for
sophisticated roadway profiling system currently available. A typical example in regard to this aspect
measuring road surface roughness [1,3,4], including Rod and Level, Straight Edge, Profilograph,
is the vehicle-mounted laser crossWalking
Response-Type Instruments, section measuring
Profilers, system
and Inertial (Figure
Profilers. 1), Profilers
Inertial which uses
are theamost
laser scanner
and accelerometer
sophisticated to record
roadwaythe longitudinal
profiling displacement,
system currently available.and a horizontal
A typical exampledisplacement
in regard to thisrecorder to
aspect
measure the is the vehicle-mounted
distance travelled. These laserkinds
cross section measuring
of systems systema (Figure
provide relatively1), which uses aand
accurate laserrepeatable
scanner and accelerometer to record the longitudinal displacement, and a horizontal displacement
simulationrecorder
of thetoroad profile, which can then be analyzed to produce various roughness statistics,
measure the distance travelled. These kinds of systems provide a relatively accurate and
including repeatable
the IRI. These systems
simulation are often
of the road profile,very
whichexpensive
can then be and restricted
analyzed to produceto professional
various roughnessapplications
in highway engineering. More importantly, due to the employed vehicles, these systems can only be
statistics, including the IRI. These systems are often very expensive and restricted to professional
applicationsroads,
used on motorable in highway engineering.
and are More for
not suitable importantly,
measuring due theto the employed
surface vehicles, of
roughness these
other roads,
systems can only be used on motorable roads, and are not suitable for measuring the surface
such as cycle lanes and pedestrian lanes.
roughness of other roads, such as cycle lanes and pedestrian lanes.

(a) (b)
Figure 1. Vehicle-mounted laser cross section measuring system: (a) laser Scanner to record the
Vehicle-mounted
Figure 1. longitudinal laser cross section measuring system: (a) laser Scanner to record the
displacement (b) horizontal displacement recorder.
longitudinal displacement (b) horizontal displacement recorder.
With the rapid development in mobile sensor technology, the cost of accelerometers and other
measurement sensors (e.g., GPS) is getting lower and lower. These sensors as well as others have
With become
the rapid development in mobile sensor technology, the cost of accelerometers and
portable and have been integrated into most of the existing smartphones. Therefore, recent
other measurement sensors
research interest has been(e.g., GPS)
directed towardis using
getting
theselower
low-costand lower. These
accelerometers sensors
and other sensors toas well as
measure
others have road surface
become portableroughness [8]. In the
and have beenfollowing section, we
integrated intoreview
most these
of studies.
the existing smartphones.
Most of the existing studies related to this aspect used cars as a carrying platform. For example,
Therefore, recent research interest has been directed toward using these low-cost accelerometers and
Gonzalez et al., (2008) placed a tri-axial accelerometer on a car to obtain acceleration information and
other sensors to measure road surface
estimate the road surface roughness roughness [8].etInal.,the
[18]. Eriksson following
(2008) added a section,
GPS devicewetoreview
Gonzalezthese
et studies.
Most al.’s
of the existing studies related to this aspect used cars as a carrying platform.
solution to record trajectories and locate potholes [19]. Mednis et al., (2010) further added a For example,
Gonzalez et al. (2008) placed a tri-axial accelerometer on a car to obtain acceleration information
and estimate the road surface roughness [18]. Eriksson et al. (2008) added a GPS device to Gonzalez
et al.’s solution to record trajectories and locate potholes [19]. Mednis et al. (2010) further added
a microphone to identify potholes from the sound signal [20]. Vittorio Astarita and colleagues assessed
Sensors 2018, 18, 914 4 of 17

road surface quality based on GPS and accelerometer data on smartphones placed on cars [21–23].
Several recent studies also used cars as a carrying platform. Some of them further investigated the
relationships between
Sensors 2018, 18, x FOR sensor output and the roughness index [7,24,25], while the others proposed
PEER REVIEW 4 of 17

algorithms for pothole recognition [26]. Li and Goldberg (2018) proposed a mobile crowdsensing
microphone to identify potholes from the sound signal [20]. Vittorio Astarita and colleagues
system for road surface assessment based on GPS and accelerometer data from smartphones placed
assessed road surface quality based on GPS and accelerometer data on smartphones placed on cars
on cars [6]. Alessandroni et al. (2017) further investigated the influence of speed on road roughness
[21–23]. Several recent studies also used cars as a carrying platform. Some of them further
sensing using smartphones on cars [27]. Different from the above studies, Tai et al. (2010) used
investigated the relationships between sensor output and the roughness index [7,24,25], while the
a motorcycle as a carrying
others proposed tool, for
algorithms andpothole
used the smartphone’s
recognition accelerometer
[26]. Li and and GPS
Goldberg (2018) sensor
proposed to detect
a mobile
road damage location information [28]. Chang et al. (2009) used automatic
crowdsensing system for road surface assessment based on GPS and accelerometer data from robots with GPS equipment
to estimate road roughness
smartphones placed on[29].cars [6]. Alessandroni et al., (2017) further investigated the influence of
Due to on
speed theroad
fact roughness
that IRI and otherusing
sensing metrics of road surface
smartphones on carsroughness cannot
[27]. Different frombethemeasured directly,
above studies,
Tai et al., (2010)
a computational model used a motorcycle
is needed as a carrying
to convert sensortool, and (e.g.,
output used the smartphone’s
acceleration dataaccelerometer and
from accelerometer
GPS
sensors) to sensor to detectThe
these metrics. roadmost
damage location
direct method information [28]. Chang
is calculating et al., (2009)
the cumulative used displacement
vertical automatic
robots
value of roadwith GPS equipment
bumps by using the to estimate
quadratic road roughness
integral [29].
of acceleration data [25]. Other attempts have
Due to the fact that IRI and other metrics of road surface roughness cannot be measured
included the artificial neural network [30,31], spectrum analysis [32,33], frequency domain analysis [7],
directly, a computational model is needed to convert sensor output (e.g., acceleration data from
and the genetic algorithm [24]. However, it is also important to note that these studies mainly used
accelerometer sensors) to these metrics. The most direct method is calculating the cumulative
cars as the carrying platform.
vertical displacement value of road bumps by using the quadratic integral of acceleration data [25].
InOther
summary, several
attempts havemethods
includedhave been proposed
the artificial involving[30,31],
neural network the use of low-cost
spectrum accelerometers
analysis [32,33],
and other
frequency domain analysis [7], and the genetic algorithm [24]. However, it is also important to note to
sensors to measure road surface roughness and have shown promising results compared
professional
that theseequipment.
studies mainlyThese studies
used have
cars as mainly used
the carrying cars as the carrying platform, while several
platform.
others have In relied
summary, on motorcycles
several methods and automatic
have robots.involving
been proposed Due to the
theuse
carrying platforms
of low-cost employed,
accelerometers
and other are
these methods sensors
not to measure
suitable forroad surface roughness
measuring the surface androughness
have shown ofpromising
un-motorableresultsroads,
comparedsuch as
to professional
cycle lanes and pedestrianequipment.
lanes.These studies
Until now, have
theremainly
has beenusednocars as the method
effective carrying forplatform, while the
measuring
surfaceseveral
roughnessothersofhave
theserelied
roads.on motorcycles and automatic robots. Due to the carrying platforms
employed, these methods are not suitable for measuring the surface roughness of un-motorable
In recognizing these research gaps, this research aimed to propose a method for measuring the
roads, such as cycle lanes and pedestrian lanes. Until now, there has been no effective method for
road surface roughness of pedestrian and cycle lanes. Particularly, we studied the feasibility of using
measuring the surface roughness of these roads.
GPS and accelerometer sensors on bicycle-mounted smartphones for assessing road surface roughness.
In recognizing these research gaps, this research aimed to propose a method for measuring the
road surface roughness of pedestrian and cycle lanes. Particularly, we studied the feasibility of using
3. Methodology
GPS and accelerometer sensors on bicycle-mounted smartphones for assessing road surface roughness.
To comprehensively address the above research objectives, we applied the following methodology.
3. Methodology
3.1. Computational Model of Road
To comprehensively Surface
address Roughness:
the above IRI
research objectives, we applied the following methodology.
Road surface roughness does not have a unified general definition. Normally, computation of
3.1. Computational Model of Road Surface Roughness: IRI
road surface roughness includes three steps: (1) data acquisition; (2) pre-processing; and (3) calculation.
This paperRoad
chosesurface roughness
the IRI as the does not have
evaluating a unified of
indicator general
road definition. Normally, computation
surface roughness. of
IRI is calculated
road surface roughness includes three steps: (1) data acquisition, (2) pre-processing,
from a measured longitudinal road profile by accumulating the output from a quarter-car model and and (3)
calculation. This paper chose the IRI as the evaluating indicator of road surface roughness. IRI is
dividing by the profile length to yield a summary roughness index with units of slope [12].
calculated from a measured longitudinal road profile by accumulating the output from a quarter-car
For example, for a vehicle driving on a road like that shown in Figure 2, ti (i ∈ [1, n]) is the
model and dividing by the profile length to yield a summary roughness index with units of slope [12].
sampling time, and hi is for
For example, the alongitudinal offset
vehicle driving onof
a the
roadroad
likesurface at ti . in
that shown Through calculation,
Figure 2, ( ∈ 1, n the
) is vertical
the
displacement
sampling(Vh i ) of
time, each
and sampling
hi is intervaloffset
the longitudinal can beof obtained:
the road surface at . Through calculation, the
vertical displacement ( ℎ ) of each sampling interval can be obtained:
Vhi = |hi − hi−1 | (1)
ℎ = |ℎ − ℎ | (1)

Figure
Figure 2. Example
2. Example ofofa aroad
roadlongitudinal
longitudinal profile.
profile.
Sensors 2018, 18, 914 5 of 17
Sensors 2018, 18, x FOR PEER REVIEW 5 of 17

Based on
Based on the
the IRI
IRI definition,
definition, we
we can
can sum
sum up up the
the vertical
vertical displacement
displacement of all sampling
of all sampling intervals,
intervals,
and divide it with the travel distance(S) to compute
and divide it with the travel distance(S) to compute IRI: IRI:
∑ ℎ ∑ |ℎ − ℎ |
IRI = ∑n Vhi= ∑n |hi − hi−1 | (2)
i =2
IRI = = i =2 (2)
S S
Therefore, when computing IRS, we must know the total travel distance and the vertical
Therefore,value
displacement whenof computing
each sampling IRS,time.
we must
Travelknow
distancethe can
total
betravel distance
computed andGPS
via the the location.
vertical
displacement value
However, vertical of each sampling
displacement is not time.
a valueTravel distance
that can can bedirectly
be obtained computed
andviacanthe
be GPS location.
derived from
However,
the outputvertical displacementsensors.
of the accelerometer is not a value that can
In physics, we be
know obtained
that directly and can be derived from
the output of the accelerometer sensors. In physics, we know that
ℎ
ν =
νv = dVhdt (3)
νdνv = d2 Vh
ℎ (3)
α
α = v = dt= dt2

wherettisistime,
where time,νvνis vertical speed,
is vertical αv isαvertical
speed, acceleration,
is vertical and Vhand
acceleration, is the ℎvertical
is thedisplacement. Therefore
vertical displacement.
Therefore tstop
x
∑ℎ =Vh = |α |(dt)2
|α v| ( )
(4)
(4)
tstart

Now, we
Now, we can
can deduce
deduce that
that
s |α |( ) 2
∑∑n ℎVh ∬ tstop |αv |(dt) (5)
IRI =
IRI = i =2 =
i t
= start (5)
S S
Figure 3a shows the quarter-car model commonly used in computing IRI, in which the car can
divided into
be divided into three
three parts:
parts: body,
body, suspension
suspension and and wheel.
wheel. Several parameters are included when
computing the road road vertical
vertical displacement
displacement (Vh):
(Vh): the mass of the vehiclevehicle body
body (m (m11), the mass of wheel
and suspension
suspension (m (m22),),the
theelastic
elasticcoefficient
coefficientofofthe
thespring
spring (K(K
1),1the elastic
), the coefficient
elastic coefficientof the tiretire
of the (K2),(Kthe
2 ),
damping
the damping coefficient of theofsuspension
coefficient (C), and
the suspension two
(C), andbody
twovertical displacements
body vertical (x1 and x2(x
displacements ). 1However,
and x2 ).
in the common
However, in the bicycle
common structure, the frame’s
bicycle structure, theand the wheel’s
frame’s and the elastic
wheel’s capacities are very are
elastic capacities limited.
very
Therefore,
limited. in the bicycle
Therefore, in themodel
bicycle(Figure
model3b), the bicycle
(Figure 3b), thecan be regarded
bicycle as a rigidasstructure,
can be regarded so K1, C
a rigid structure,
and
so K1K, 2Ccan
andbeK2ignored; the model
can be ignored; theparameters are reduced
model parameters to m and
are reduced to mx. and
We x.assumed
We assumedthat the thatrider
the
keepskeeps
rider the bicycle posture
the bicycle as stable
posture as possible.
as stable In other
as possible. words,
In other he/she
words, triestries
he/she to avoid making
to avoid making the
bicycle
the swing.
bicycle To To
swing. further reduce
further reducethethe
impact
impact ofofpotential
potentialsmall
smallswings,
swings,we weused
used high-pass
high-pass filter to
clean the
thesensor
sensordata
data when
when we we computed
computed IRI Section
IRI (see (see Section
3.2.2 for 3.2.2
morefordetails).
more details).
In the rigidIn structure,
the rigid
structure,
there is no there
springisshock
no spring
effectshock effect
and the andbody’s
bicycle the bicycle body’s
vertical vertical displacement
displacement only makes
only makes contact with
contact
the with the
tire–road tire–road
surface, surface, and
and therefore, x istherefore,
equal to Vh. x is equal to Vh.

(a) (b)
Figure 3.
Figure 3. The
The original
original quarter-car model (a)
quarter-car model (a) and
and our
our bicycle
bicycle model
model (b)
(b) used
used in
in the
the computation
computation of
of the
the
International Roughness Index (IRI).
International Roughness Index (IRI).

Another important thing that needs to be considered is the installation position of smartphones
on bicycles. When the smartphone is fixed to different parts of a bicycle, different acceleration data
might be recorded by the smartphone. Among all other places, the center of the handlebar seems to
Sensors 2018, 18, 914 6 of 17

Another important thing that needs to be considered is the installation position of smartphones
on bicycles. When the smartphone is fixed to different parts of a bicycle, different acceleration
data might
Sensors 2018,be
18, xrecorded by the smartphone. Among all other places, the center of the6 ofhandlebar
FOR PEER REVIEW 17

seems to be a typical place for fixing smartphones on bicycles, according to cycling web platforms,
be a typical place for fixing smartphones on bicycles, according to cycling web platforms, like
like Cycling Weekly (http://www.cyclingweekly.com/) and ChinaBike (http://www.chinabike.net/).
Cycling Weekly (http://www.cyclingweekly.com/) and ChinaBike (http://www.chinabike.net/).
Therefore, in the current study, we fixed the smartphones to the center of the handlebar, as can be seen
Therefore, in the current study, we fixed the smartphones to the center of the handlebar, as can be
in Figure
seen in 4. We further
Figure discuss
4. We further this issue
discuss in Section
this issue 5. 5.
in Section

(a) (b)

Figure 4. The installation position of smartphones on bicycles: (a) a sketch; (b) an example.
Figure 4. The installation position of smartphones on bicycles: (a) a sketch; (b) an example.
After these considerations, we are now ready to calculate IRI. As can be seen from Equation (5),
weAfter
needthese
to have S, α , and t to compute
considerations, we are now
IRI. ready to calculate
In the next IRI.explore
section, we As canhow
be seen
GPS from Equation
sensors and (5),
weaccelerometer
need to havesensors
S, αv , on
andbicycle-mounted
t to compute smartphones can be
IRI. In the next used towe
section, derive thesehow
explore parameters, and
GPS sensors and
thus computesensors
accelerometer IRI. on bicycle-mounted smartphones can be used to derive these parameters,
and thus compute IRI.
3.2. Deriving IRI from GPS and Accelerometer Sensors on Bicycle-Mounted Smartphone
3.2. Deriving IRI from
According GPS
to the and Accelerometer
analysis Sensors
in the previous on Bicycle-Mounted
section, Smartphone
the basic elements of the IRI computing
model are the accumulative vertical displacement value from the longitudinal profile (i.e., the
According to the analysis in the previous section, the basic elements of the IRI computing model
numerator in Equation (5)) and the overall distance of the measurement (i.e., the denominator).
areTheir
the accumulative vertical displacement value from the longitudinal profile (i.e., the numerator
values can be derived from sensors on bicycle-mounted smartphones. For example, the travel
in Equation (5)) and the
distance (S) can be computedoverall distance
using of the measurement
data recorded by GPS sensors(i.e., the denominator).
on smartphones, and α can Their
be values
canobtained
be derived from sensors on bicycle-mounted smartphones. For example, the travel
from accelerometer sensors. These two kinds of sensors are available in most of the current distance (S)
cansmartphones.
be computed using
GPS dataon
sensors recorded by GPS
smartphones cansensors on smartphones,
return the current locationand αv can be obtained
(longitude/latitude) as from
accelerometer sensors.
well as information likeThese
traveltwo kinds
speed. of sensors on
Accelerometers arecurrent
available in most of
smartphones arethe current
often smartphones.
tri-axial, and
GPSprovide acceleration
sensors information
on smartphones on three
can return axes
the (x, y, and
current z) in the
location unit of g. In the following,
(longitude/latitude) as wellweas show
information
how these data can be used to compute IRI, as specified in Equation (5).
like travel speed. Accelerometers on current smartphones are often tri-axial, and provide acceleration
information on three axes (x, y, and z) in the unit of g. In the following, we show how these data can
3.2.1. Calculating the Travel Distance (S)
be used to compute IRI, as specified in Equation (5).
There are two ways to calculate the travel distance (S). Firstly, with GPS, we can get the
longitude and latitude
3.2.1. Calculating coordinates
the Travel Distanceof each
(S) sampling point. The travel distance can be computed by
summarizing the distances between every two adjacent points. The distance between two sampling
There
points are
can betwo ways to
calculated calculate theusing
approximately travel
thedistance (S).formula,
Haversine Firstly, specifically,
with GPS, we can get the longitude
and latitude coordinates of each sampling point. The travel distance can be computed by summarizing
the distances between every two adjacent−points. The distance between two − sampling points can be
d = 2 ∗ ∗ arcsin sin + cos( ) ∗ cos( ) ∗ sin( ) (6)
2
calculated approximately using the Haversine formula, specifically, 2

where and are the latitude and  longitude of point 1, and are the latitude and
s !
ϕ2 − ϕ1 λ − λ1
d=
longitude of 2 ∗ 2,
point ∗ arcsin
R and 2
sin radius (mean
R is the Earth’s +radius ∗ cos
cos( ϕ=1 )6371 ( ϕ2 ) ∗ sin ( 2
km). ) (6)
2 2
Alternatively, the travel distance can be computed using the travel speed measured at each
sampling point. Specifically,
where ϕ1 and λ1 are the latitude and longitude of point 1, ϕ2 and λ2 are the latitude and longitude of
point 2, and R is the Earth’s radius (mean radius
= = 6371 km). (7)

where is the travel speed measured at time t. It can be obtained directly from GPS sensors.
Sensors 2018, 18, 914 7 of 17

Alternatively, the travel distance can be computed using the travel speed measured at each
sampling point. Specifically,
Z t
S= Vt dt (7)
0
where Vt is the travel speed measured at time t. It can be obtained directly from GPS sensors.
For smartphones, the latitude and longitude provided by GPS often lead to an accuracy of 10 m,
while for instant travel speeds provided by GPS, the accuracies reported by manufacturers range from
0.1 m/s to 0.2 m/s [34]. Therefore, using travel speeds might lead to more accurate estimation of
travel distance than using longitude/latitude coordinates of each GPS point [35]. Considering this,
the second method was adopted in this paper to compute the travel distance.

3.2.2. Obtaining Vertical Displacement

In theory, we can use altitude information returned by GPS for computing vertical displacement.
However, the altitude accuracy of GPS sensors is often very poor, and their sampling frequency is often
rather low (typically 1 Hz). Therefore, using altitude data provided by GPS for calculating vertical
displacement is problematic. In this paper, we used vertical acceleration provided by accelerometers on
smartphones, which allows a much higher sampling rate. As the installation of the smartphone on the
bicycle as well as the orientation of the smartphone are unknown and variable, the vertical acceleration
(αv ) may appear in either dimension of the tri-axial acceleration data returned by the accelerometer.
In other words, the z-axis acceleration data from the accelerometer cannot be taken directly as the
vertical acceleration (αv ). We need a method to derive αv from the tri-axial acceleration values.
In the acquisition process, we set a requirement in advance: when the bicycle rider started the
recording, the bicycle had to be in the normal riding posture, and kept stationary for more than 5 s.
As the smartphone, which is mounted on the bicycle, is stationary, the only force it receives is the
gravitational one, and the direction is vertical and downward, with a value of 1 g. Therefore,

A x ∗ A x + Ay ∗ Ay + Az ∗ Az = 1 (8)

where A x , Ay , and Az are the average acceleration values of the x, y, and z axes in these 5 s,
obtained from the accelerometer sensor on the smartphone.


Deriving the vertical acceleration (αv ) from any tri-axial acceleration output A = A x , Ay , Az


can be considered as projecting the vector A onto the reference vector A = A x , Ay , Az , measured at
the beginning of the acquisition process. In other words, αv is the scalar projection of vector A and
A. Therefore,

A· A A x ∗ A x + Ay ∗ Ay + Az ∗ Az
αv =   = q = A x ∗ A x + Ay ∗ Ay + Az ∗ Az (9)
 A
A x ∗ A x + Ay ∗ Ay + Az ∗ Az

With these, the parameters required in Formula (5) have been fully prepared. We can then use it
to compute IRI values from the data obtained from GPS and accelerometer sensors on smartphones.
During the acquisition process, we asked bicycle riders to keep the speed and the bicycle posture
as stable as possible, and not to deliberately pass through or avoid bump areas on the road. To further
reduce the influence of the potential speed changes and small swings, we used a high-pass filter
(setting the value cutoff frequency as 0.5 Hz) to clean the sensor data when calculating vertical
acceleration. We will further discuss these issues in Section 5.

3.2.3. Sampling Frequency Selection

As showed in the above two sections, a GPS sensor and an accelerometer sensor are needed
to estimate travel distance and vertical displacement. In order to provide accurate estimations,
the sampling frequencies of both sensors should be carefully selected. For GPS sensors, we set the
Sensors 2018, 18, 914 8 of 17

sampling frequency as 1 Hz. This is because the maximum update rate of GPS sensors on existing
smartphones is 1 Hz (while professional GPS devices can offer higher frequencies). Setting the sampling
frequency as 1 Hz allowed our method to be feasible for existing smartphones.
Regarding accelerometer sensors on smartphones, the sampling rate can go up to 2000 Hz.
As mentioned in [12], the sample interval of IRI should be no larger than 300 mm for
accurate calculation. The typical speed of bicycles is about 15.5 km/h [36]. Therefore, to allow
accurate calculation of IRI, the sampling rate of accelerometers should not be less than
15.5 km/h ÷ 300 mm ≈ 15 Hz. In order balance the sampling accuracy and the computational power
of mobile phones, we chose 100 Hz.

3.3. Recognition of Potholes and Humps on Roads

In addition to measuring the road surface roughness (i.e., using IRI as in the above sections),
we needed an algorithm to recognize potholes and humps on roads, and identify their locations.
This kind of information is very useful for many applications, such as navigation and routing as well
as road maintenance.
The underlying considerations of the algorithm are

(1) when a bicycle drives through a pothole or hump of the road, the accelerometer sensor on the
bicycle-mounted smartphone will record some rather high instantaneous acceleration values.
In other words, some spikes can be observed from the accelerometer output.
(2) The bicycle will touch the pothole or hump twice, once with the front wheel and the other time
with the rear wheel. Therefore, at least two spikes will be observed. Furthermore, the duration
between the first spike and the last one is longer than L/ν, where L is the distance between the
centers of the front and rear wheels, and ν is the travel speed of the bicycle.
(3) If two adjacent potholes or humps are longer than L, we would consider that these two adjacent
potholes or humps are far away from each other and should be considered as two different
potholes or humps.

Figure 5 shows the algorithm in pseudo code. It takes a series of vertical acceleration
measurements, an acceleration threshold (δa ), and a time gap threshold (δt ) as inputs. The output of
the algorithm is a set of bumpy areas, containing the timestamps of the start and the end of each of the
bumpy areas. These timestamps can be synchronized with GPS measurements to identify the locations
of these bumpy areas. The acceleration threshold (δa ) is used to recognize spikes from the acceleration
measurements, while the time gap threshold (δt ) is used to check the time distance between spikes
(as mentioned before). The algorithm tries to find the first measurement with a vertical acceleration
bigger than δa (Line 8), i.e., the first spike. When the first spike is found, it continually searches for the
next one. If the time difference of a spike and its previous one is bigger than δt , a candidate pothole or
hump area with all the previous spikes is formed (Lines 10–15). If the duration of the candidate area is
bigger than δt , this candidate area is recognized as a pothole/hump area (Lines 17–21). Otherwise, it is
a noise. The algorithm continues this process until the last measurement. At the end, it returns a set of
bumpy areas.
The algorithm has two parameters to calibrate: the acceleration threshold (δa ), and the time gap
threshold (δt ). In our initial test, we found that when riding at normal speed, even on rough roads,
the instantaneous vertical acceleration value did not exceed 2 g. Therefore, we set the value of δa to
2 g. The time gap threshold (δt ) can be set as L/ν, where L is the distance between the centers of
the front and rear wheels, and ν is the travel speed of the bicycle. Therefore, it can be computed
automatically based on the travel speed of the bicycle, e.g., as derived from GPS sensors. In our
evaluation experiments, we set δt as 1 m ÷ 15.5 km/h ≈ 0.23 s, as the distance between the centers of
the front and rea wheels of adult bicycles is normally 1 m, and the typical speed of bicycles is about
15.5 km/h [36].
Sensors 2018, 18, 914 9 of 17
Sensors 2018, 18, x FOR PEER REVIEW 9 of 17

Figure
Figure 5. Algorithm
5. Algorithm forthe
for therecognition
recognition of
of potholes
potholesand
andhumps
humpson on
roads.
roads.

4. Implementation and Evaluation

4. Implementation and Evaluation
As a proof of concept, this section describes the implementation of a mobile application for
As a proof
mapping of surface
road concept, this section
roughness describespotholes
and identifying the implementation
and humps on of a mobile
roads, application
using the method for
mapping
and road surface
algorithm roughness
introduced and identifying
above. We furtherpotholes
evaluate and
the humps on roads,
proposed using the
methodology method
with three and
algorithm introduced above. We further evaluate the proposed methodology with three
experiments, aiming to answer the following questions: (1) Are the IRI values measured comparable experiments,
aiming to answer
to those thewith
reported following questions:
professional (1) Are
instruments? (2)the
HowIRIdovalues
the IRImeasured comparable
values measured to those
differ from
different
reported withbicycles, different
professional people and the
instruments? (2) fixing
How methods
do the IRIof smartphones?
values measured(3) How well
differ doesdifferent
from the
algorithm
bicycles, perform
different peoplein recognizing potholes
and the fixing and humps
methods on roads?
of smartphones? (3) How well does the algorithm
perform in recognizing potholes and humps on roads?
4.1. Implementation
4.1. Implementation
We developed a mobile phone application called “RoadSR” based on an iOS system. The
application uses GPS and accelerometer sensors on smartphones to record location and acceleration
We developed a mobile phone application called “RoadSR” based on an iOS system.
information. The sensor outputs are then further processed to calculate IRI values as well as to
The application uses GPS and accelerometer sensors on smartphones to record location and acceleration
identify potholes and humps on roads. The application finally visualizes the IRI values computed
information. The sensor
and the potholes andoutputs
humps are areidentified
then further
on aprocessed
map, usingto calculatemap
AutoNavi IRI SDK.
values as well as to identify
potholes and humps
To use on roads.
“RoadSR”, Theonly
the rider application
needs to fixfinally visualizesonto
the smartphone the the
IRIbicycle,
values click
computed and the
the “Start”
potholes and
button onhumps are identified
the bottom on a map,interface,
left of the application using AutoNavi map SDK.
and data about the road surface roughness will
be automatically
To use “RoadSR”, collected using
the rider GPSneeds
only and accelerometer sensors (Figure
to fix the smartphone onto6a).
theThe rider can
bicycle, clickclick
thethe
“Start”
“Finish” button on the bottom right to stop recording. When the recording
button on the bottom left of the application interface, and data about the road surface roughness is stopped, the
application
will be automatically
automatically collected calculates
using GPSthe and
IRI values of the road
accelerometer just (Figure
sensors traveled6a).
and The
identifies
rider can
potholes/humps (Figure 6c). The rider can then view the results on a map view. Figure 5d shows a
click the “Finish” button on the bottom right to stop recording. When the recording is stopped,
map view, visualizing the bicycle trajectory, the surface roughness (IRI) of the road in different
the application automatically calculates the IRI values of the road just traveled and identifies
colors, and potholes/humps along the route as black dots.
potholes/humps (Figure 6c). The rider can then view the results on a map view. Figure 5d shows
a map view, visualizing the bicycle trajectory, the surface roughness (IRI) of the road in different colors,
and potholes/humps along the route as black dots.
Sensors 2018, 18, 914 10 of 17
Sensors 2018, 18, x FOR PEER REVIEW 10 of 17

Figure 6. Measuring process based on the “RoadSR” app. (a) The rider fixes the smartphone onto the
Figure 6. Measuring process based on the “RoadSR” app. (a) The rider fixes the smartphone onto the
bicycle and clicks “Start” to start data recording. (b) The rider rides the bicycle. (c) When the
bicycle and clicks “Start” to start data recording. (b) The rider rides the bicycle. (c) When the recording
recording is stopped, the app automatically calculates the IRI values of the road just traveled and
is stopped, the app automatically calculates the IRI values of the road just traveled and identifies
identifies potholes/humps. (d) The computed IRI values and the potholes/humps recognized are
potholes/humps. (d) The computed IRI values and the potholes/humps recognized are visualized on
visualized on a map view.
a map view.

4.2. Experiment 1: Comparing the Method with the Professional System

4.2. Experiment 1: Comparing the Method with the Professional System
This experiment aimed to study how well the IRI values measured by the proposed method
This experiment aimed to study how well the IRI values measured by the proposed method
correlate with those measured by professional instruments. We selected 10 road sections in Banqiao
correlate with those measured by professional instruments. We selected 10 road sections in Banqiao
Town, southwest of Nanjing, China (Figure 7). These 10 road sections differed in pavement materials
Town, southwest of Nanjing, China (Figure 7). These 10 road sections differed in pavement materials
and conditions, road widths and lengths. We used the car mounted laser pavement scanner (see
and conditions, road widths and lengths. We used the car mounted laser pavement scanner
Figure 1) from the Nanjing Highway Scientific Research Institute, which is a professional instrument
(see Figure 1) from the Nanjing Highway Scientific Research Institute, which is a professional
and system for measuring IRI values on motorable roads. For each road section, we first drove the
instrument and system for measuring IRI values on motorable roads. For each road section, we first
car, followed by the bicycle. When riding the bicycle, we tried to keep the speed and the bicycle
drove the car, followed by the bicycle. When riding the bicycle, we tried to keep the speed and the
posture as stable as possible. At the end of the collection, we computed the IRI values of each road
bicycle posture as stable as possible. At the end of the collection, we computed the IRI values of each
section, taking 20 m as the computing unit.
road section, taking 20 m as the computing unit.
A Pearson correlation coefficient was computed to assess the relationship between the IRI
A Pearson correlation coefficient was computed to assess the relationship between the IRI values
values measured by the car with the laser scanner to those measured by our proposed method. The
measured by the car with the laser scanner to those measured by our proposed method. The results
results show that there was a very positive and significant correlation between the IRI values
show that there was a very positive and significant correlation between the IRI values measured by
measured by the proposed method and those measured by professional instruments (r = 0.893, n = 9,
the proposed method and those measured by professional instruments (r = 0.893, n = 9, p = 0.001).
p = 0.001).
In summary, the experiment shows that the IRI values measured by the proposed method strongly
In summary, the experiment shows that the IRI values measured by the proposed method
and positively correlate with those measured by professional instruments. In other words, the proposed
strongly and positively correlate with those measured by professional instruments. In other words,
bicycle-mounted smartphone based IRI collection method can well reflect the road roughness level.
the proposed bicycle-mounted smartphone based IRI collection method can well reflect the road
roughness level.
Sensors 2018,
Sensors 2018, 18,
18, 914
x FOR PEER REVIEW 11 of
11 of 17
17

Sensors 2018, 18, x FOR PEER REVIEW 11 of 17

Figure 7. The 10 test roads of Experiment 1 and their details, located in Banqiao Town, southwest of
Figure
Figure 7.
7. The
The 10 test
test roads
10China.roads of
of Experiment
Experiment 11 and
and their
their details,
details, located
located in
in Banqiao Town, southwest
Banqiao Town, southwest of
of
Nanjing,
Nanjing, China.
Nanjing, China.
4.3. Experiment 2: Method Stability
4.3. Experiment 2:
4.3. Experiment 2: Method
Method Stability
Stability
This experiment aimed to investigate how the IRI values measured by the proposed method
This
Thisdiffer for various
experiment
experiment people,
aimed
aimed to bicycles andhow
to investigate
investigate smartphone
how the
the IRI fixing methods.
IRI values
values measured
measured Particularly,
by
by the we usedmethod
the proposed
proposed two
method
differ
differ fordifferent
forvarious settings
variouspeople, (Figure
people, 8):
bicycles
bicyclesGroup
andand A (mountain
smartphone bicycle +
fixingfixing
smartphone fixing
methods. holder), and
Particularly,
methods. Group B (city
we usedwe
Particularly, bicycle
twoused +
different
two
settings directly fixed). We assigned three male graduate students to each group. We told them to ride on the
different(Figure
settings8):(Figure
Group8): A Group
(mountain bicycle +bicycle
A (mountain fixing holder),
+ fixing and Group
holder), andBGroup
(city bicycle
B (city+bicycle
directly+
right side of the road and to not deliberately pass through or avoid bumpy areas on the roads.
fixed).
directlyWe assigned
fixed). three male
We assigned graduate
three students to
male graduate each group.
students to each Wegroup.
told them to ride
We told themon to
theride
right
onside
the
of theside
right roadofandtheto notand
road deliberately pass through
to not deliberately orthrough
pass avoid bumpy
or avoidareas on the
bumpy roads.
areas on the roads.

(a) (b)
Figure 8. Photos of the two sets of equipment and smartphone fixing modes: (a) Group A: mountain
bicycle with iPhone 6S, with fixing holder; (b) Group B: “Bluegogo” sharing city bicycle with iPhone 6,
directly fixed. (a) (b)
Figure 8.
8. Photos
Photos of
of the
the two
two sets
sets of
of equipment
equipment and
and smartphone fixing
fixing modes: (a)
(a) Group
Group A:
A: mountain
Figure The experiment was carried out on eightsmartphone
road sections in modes:
the Xianlin campusmountain
of Nanjing
bicycle
bicycle with iPhone 6S,
withUniversity with
iPhone 6S, (Figurefixing
with fixing holder;
holder; (b) Group
(b)section B:
Group was “Bluegogo”
B: “Bluegogo” sharing city bicycle with
with iPhone 6,
Normal 9). Each road measured sharing city of
by two sets bicycle
equipment iPhone
(Group6,A,
directly
directly fixed.
fixed.
Group B), each with three users. In total, there were six data collections for each section. We also
ensured that each user only rode one time in each section.
The experiment wasthe
Figure 10 shows
carried
results
outtheon eight road sections in the Xianlin campus of Nanjing
The experiment was carried outofon experiment.
eight Several interesting
road sections resultscampus
in the Xianlin can be seen from Figure
of Nanjing 10.
Normal
NormalOverall,
University
the (Figure
IRI values 9). Each
measured road
by section
both groupswas measured
fell into the by
same two sets
range, of
with equipment
an average (Group A,
difference
University (Figure 9). Each road section was measured by two sets of equipment (Group A, Group B),
Group B), eachm/km;
of 1.80 with Meanwhile,
three users.theInIRI
total,
valuestheregroup
were Asixweredata collections for each section. We also
each with three users. In total, there were six of
data collections always lower
for each than
section. those of group
We also B. (2)
ensured that
ensuredWithin
that each user group,
the same only rode one time
IRI values in eachbysection.
measured different users seem rather consistent. The average
each user only rode one time in each section.
difference
Figure in Group
10 shows A was
the results of 0.76 m/km, whileSeveral
the experiment. in Group B it wasresults
interesting 0.66 m/km.
can beInseen
other words,
from the10.
Figure
Figure 10 shows the results of the experiment. Several interesting results can be seen from
Overall, the IRI values measured by both groups fell into the same range, with an average difference
Figure 10. Overall, the IRI values measured by both groups fell into the same range, with an average
of 1.80 m/km; Meanwhile, the IRI values of group A were always lower than those of group B. (2)
difference of 1.80 m/km; Meanwhile, the IRI values of group A were always lower than those of
Within the same group, IRI values measured by different users seem rather consistent. The average
group B. (2) Within the same group, IRI values measured by different users seem rather consistent.
difference in Group A was 0.76 m/km, while in Group B it was 0.66 m/km. In other words, the
Sensors 2018, 18, 914 12 of 17

Sensors 2018,
The average 18, x FOR PEER
difference REVIEWA was 0.76 m/km, while in Group B it was 0.66 m/km.12In
in Group of 17
other
Sensors 2018, 18, x FOR PEER REVIEW 12 of 17
words, the deviation over different users seems much smaller than the deviation over different groups
deviation over different users seems much smaller than the deviation over different groups (A and
(A and B).(3)(3)
B).
deviation TheThe
over
rougher
rougher the the
different
road,
road,
users the
themuch
greater
greater
seems
the
thesmaller
IRI IRI deviation
deviation between
than the
between
users
deviation
users
and
over
and
groups.
different
groups.
groups (A and
B). (3) The rougher the road, the greater the IRI deviation between users and groups.

Figure 9. The eight test roads of Experiment 2 and their details, located in the Xianlin campus of
Figure 9. The eight test roads of Experiment 2 and their details, located in the Xianlin campus of
Nanjing
Figure 9. Normal University,
The eight test roadsNanjing, China. 2 and their details, located in the Xianlin campus of
of Experiment
Nanjing Normal University, Nanjing, China.
Nanjing Normal University, Nanjing, China.

Figure 10. Comparison of IRI values measured by different groups and users.
Figure 10. Comparison of IRI values measured by different groups and users.
Figure 10.
In summary, Comparison
this experimentofshowsIRI values
thatmeasured
overall the byproposed
different groups and users. smartphone
bicycle-mounted
based IRI values measured
In summary, by different
this experiment shows sets
thatofoverall
equipment and different
the proposed users, to some
bicycle-mounted extent, are
smartphone
very
In
based consistent,
summary,
IRI values with
thismeasured a rather
experiment small that
by shows
different average difference
setsoverall of
the proposed
of equipment and1.80 m/km. users,
Meanwhile,
bicycle-mounted
different to some the proposed
smartphone
extent, arebased
method
very achieves
consistent, with a very
a good
rather consistency
small average over different
difference of users,
1.80 better
m/km.
IRI values measured by different sets of equipment and different users, to some extent, are very than
Meanwhile,that with
the different
proposed
equipment.
method
consistent, with aMore
achievesratheraresearch
very
small should
good
average be done
consistency toover
difference investigate
1.80 these
different
of results
users,
m/km. further.
better than that
Meanwhile, with different
the proposed method
equipment. More research should be done to investigate these results further.
achieves a very good consistency over different users, better than that with different equipment.
4.4. Experiment 3: Recognition of Potholes and Humps on Roads
More4.4.
research should be done to investigate these results further.
Experiment 3: Recognition of Potholes and Humps on Roads
This experiment aimed to evaluate the performance of the potholes/humps recognition
algorithm proposed
This experiment
4.4. Experiment 3: Recognitioninaimed
Section to3.3.evaluate
of Potholes Weand
chose seven
the
Humps road
performance
on Roadssectionsof with
the different types of potholes
potholes/humps recognitionand
humps (Figure
algorithm proposed11). inThe potholes
Section 3.3. and humps
We chose differed
seven roadinsections
type and longitudinal
with offset.
different types of We started
potholes andthe
This
humps experiment
application
(Figureand aimed
11).rode
Theon tothese
evaluate
potholes road the performance
sections.
and humps In order
differed in of
to the
type potholes/humps
record
and the actual riding
longitudinal recognition
offset. tracking
We started algorithm
and the
the
proposed in Section
potholes/humps
application 3.3.
and rode We
thaton the chose
bicycle
these seven
road road
passed
sections. sections
through,
In orderwe towith
useddifferent
record athe types
tracking
actual of potholes
camera
riding mounted
tracking and
and humps
onthe an
(Figure 11). The potholes
potholes/humps that andthe humps
bicycle differed in type and
passed through, we longitudinal
used a tracking offset. We started
camera mounted the application
on an
and rode on these road sections. In order to record the actual riding tracking and the potholes/humps
Sensors 2018,
Sensors 18, x914
2018, 18, FOR PEER REVIEW 13 of
13 of 17
17

unmanned aerial vehicle (UAV) to capture the image data at the moment that the bicycle passed a
that the bicycle passed through, we used a tracking camera mounted on an unmanned aerial vehicle
pothole/hump.
(UAV) to capture the image data at the moment that the bicycle passed a pothole/hump.

Sensors 2018, 18, x FOR PEER REVIEW 13 of 17

unmanned aerial vehicle (UAV) to capture the image data at the moment that the bicycle passed a
pothole/hump.

Figure11.
Figure Potholesand
11.Potholes andhumps
humpson
onthe
theseven
sevenroad
roadsections,
sections,asaswell
wellasastheir
theirtypes
typesand
andlongitudinal
longitudinaloffsets.
offsets.

Figure 11. Potholes and humps on the seven road sections, as well as their types and longitudinal offsets.
We compared
We compared thepotholes/humps
the potholes/humpsrecognized recognizedbyby thethe proposed
proposed algorithm
algorithm withwith the ground
the ground truth
truth obtained
obtained fromWe from
the UAV theimage
compared UAV imageThe
data.
the potholes/humpsdata. The of
results results
recognized of the
thebyexperiment
the experiment
proposed show thatshow
algorithm that
the proposed
with the the algorithm
ground proposed
algorithm was
was able totruth able to
obtained
correctly correctly
from
recognize allrecognize
the UAV ofimage allThe
data. of results
the potholes/humpsthe potholes/humps
of (and
the experiment (and
show that
their locations) theirthelocations)
that bicyclethat
proposed
the the
passed
bicycle algorithm
passed (see was able to
Figure 12 correctly
for an recognize all except
example), of the potholes/humps
for the crack (and
in theirthird
the locations)
road that the (Figure
section
(see Figure 12 for an example), except for the crack in the third road section (Figure 13). Having a closer
bicycle passed (see Figure 12 for an example), except for the crack in the third road section (Figure
13).
lookHaving
at the13). a closer
crack, wealook
found atthat
the itcrack, we small,
is rather found and that it longitudinal
is rather small, and its longitudinal
−0.7 offer is
Having closer look at the crack, we found that its
it is rather small, and offer is about
its longitudinal offercm,
is which is
about −0.7
the main reasoncm, which
about −0.7why is the main
the proposed
cm, which reason why
is the main algorithm the
reason why the proposed
failed algorithm
to detect
proposed failed
thisfailed
algorithm crack. to detect
Thisthis
to detect raises this
crack.an crack. This
Thisinteresting
raises
questionan regarding
interesting whatquestion
raises an interesting be regarding
canquestion regarding
considered what
aswhat can be considered
can be
a pothole considered
or hump, as as
or amore a pothole
pothole or hump,
precisely, ora or
hump,
more or
pothole more
or hump
precisely, aprecisely,
that influences pothole a or
pothole
peoples’ hump or hump
travel that that influences
influences
experiences. peoples’ travel
peoples’
More travelexperiences.
research experiences.
should beMoreMore
done research should be
onresearch
this done be done
should
aspect.
on this aspect.
on this aspect.

(a) (b)
Figure 12. Potholes and humps at Road Section 7 and the accelerometer data. (a) The test bicycle was
Potholes
Figure 12.ridden and humps at Road Section 7 and the accelerometer data. (a) The test bicycle
over the speed bump and the other two humps (i.e., manhole covers), all of which were
was ridden over the speed bump and the other two humps (i.e., manhole covers), all of which were
(a) (b)
correctly recognized by the proposed algorithm; (b) the accelerometer data recorded when the bicycle
Figure 12. Potholes
was ridden andareas.
over these humps at Road Section 7 and the accelerometer data. (a) The test bicycle was
ridden over the speed bump and the other two humps (i.e., manhole covers), all of which were
 Sensors 2018, 18, x FOR PEER REVIEW 14 of 17

Sensors 2018,correctly
18, 914 recognized by the proposed algorithm; (b) the accelerometer data recorded when the 14 of 17
bicycle was ridden over these areas.

(a) (b)
Figure 13. Crack at Road Section 3 and the accelerometer data. (a) The test bicycle was ridden over
Figure 13. Crack at Road Section 3 and the accelerometer data. (a) The test bicycle was ridden over the
the crack which the proposed algorithm failed to recognize; this is due to the fact it is rather small
crack which the proposed algorithm failed to recognize; this is due to the fact it is rather small and its
and its longitudinal offset is only about −0.7 cm; (b) the accelerometer data recorded when the bicycle
longitudinal offset is only about −0.7 cm; (b) the accelerometer data recorded when the bicycle was
was ridden over this area.
ridden over this area.
5. Discussion
5. Discussion
As shown in the above section, “Implementation and Evaluation”, the proposed method is
feasible
As shownforinmeasuring
the above road surface
section, roughness and identifying
“Implementation potholes/humps
and Evaluation”, on roads,
the proposed usingis
method data
feasible
recorded road
for measuring by GPS surfaceandroughness
accelerometer sensors on potholes/humps
and identifying bicycle-mounted on smartphones.
roads, usingThe datathree
recorded
experiments illustrated that the IRI values measured by the proposed method strongly and
by GPS and accelerometer sensors on bicycle-mounted smartphones. The three experiments illustrated
positively correlate with those measured by professional instruments. Meanwhile, the proposed
that the IRI values measured by the proposed method strongly and positively correlate with those
algorithm was able to recognize the potholes/humps (and their locations) that a bicycle passed. The
measured by professional
proposed method is very instruments. Meanwhile,
useful for measuring the roughness
surface proposedof algorithm
roads thatwas ableaccessible
are not to recognize
for the
potholes/humps
professional instruments (which are often installed on motor vehicles), such as pedestrian lanes and for
(and their locations) that a bicycle passed. The proposed method is very useful
measuring surface roughness of roads that are not accessible for professional instruments (which are
cycle lanes.
often installed
Whenon motor vehicles),
measuring the road suchsurfaceas roughness
pedestrianusing lanesthe and cycle lanes.
proposed method, we asked bicycle
riders to
When keep the speed
measuring the road andsurface
the bicycle postureusing
roughness as stable
the as possible,method,
proposed and to not wedeliberately
asked bicycle passriders
through
to keep or avoid
the speed and bump areas onposture
the bicycle the roads.asThis might
stable not be veryand
as possible, realistic
to not fordeliberately
real-world settings, as
pass through
users might change their posture during riding. Changing the bicycle
or avoid bump areas on the roads. This might not be very realistic for real-world settings, as users posture will influence the
acceleration signals of the accelerometer sensor, and thus have a negative impact on the
might change their posture during riding. Changing the bicycle posture will influence the acceleration
measurement results. Methods regarding signal processing and noise removal might help to address
signals of the accelerometer sensor, and thus have a negative impact on the measurement results.
this issue. We will address this in our follow-up studies.
Methods In regarding
this study, signal processing
we fixed and noise
smartphones on theremoval
center of might help to
the bicycle address this
handlebar. Whileissue.
this isWea will
address this in our follow-up studies.
typical place where cyclists fix their smartphones, it might be interesting to investigate whether
In this study,
different we fixedpositions
installation smartphones on theon the centerlead
bicycle of theto bicycle
differenthandlebar.
measurement Whileresults.
this is aTotypical
placecomprehensively
where cyclists fix their smartphones,
address this issue, theitinstallation
might be interesting
position of tothe investigate
smartphone whether
shoulddifferent
be
considered
installation together
positions withbicycle
on the the length
lead of the wheelbase
to different of the bicycle.
measurement In To
results. some extreme cases, the
comprehensively address
accelerometer sensor on the smartphone might be only subjected
this issue, the installation position of the smartphone should be considered together with to angular acceleration. Therefore,
the length
of theitwheelbase
makes senseoftothe investigate
bicycle.this issue in
In some detail, and
extreme identify
cases, an optional installation
the accelerometer sensor on position based
the smartphone
on the length of a wheelbase.
might be only subjected to angular acceleration. Therefore, it makes sense to investigate this issue in
In Experiment 1, we compared the IRI values measured by a professional vehicle and those by
detail, and identify an optional installation position based on the length of a wheelbase.
our proposed bicycle-mounted smartphone-based method. It is important to note that the vehicle
In Experiment
travelled 1, we of
in the center compared
the roads,the IRIfor
while values measured
security reasons, by the abicycles
professional vehicle
only used and
the side ofthose
the by
our proposed bicycle-mounted
road. Therefore, smartphone-based
the travel trajectories method. match.
do not completely It is important to noteled
This obviously thattothe
somevehicle
travelled in the in
differences center
the IRI ofvalues
the roads, while for security
they computed. To furtherreasons,
improvethe thisbicycles
experiment,onlyoneused theuse
could side
theof the
road.bicycle
Therefore,
on thetheother
travelsidetrajectories
of the roads dothat
not the
completely match. This
vehicle traveled, obviously
and check led the
whether to some
averagedifferences
IRI
in thevalues matchthey
IRI values thosecomputed.
from the professional
To furthervehicle.
improve this experiment, one could use the bicycle on the
other side of the roads that the vehicle traveled, and check whether the average IRI values match those
from the professional vehicle.
Experiment 2 showed that different bicycles (i.e., mountain bicycle and city bicycle) lead to
slight differences in the IRI values measured. This might be due to the different elasticity coefficients
of the bicycles and their wheels/tires. Further studies should be done to investigate this aspect
Sensors 2018, 18, 914 15 of 17

further. Meanwhile, it would also be interesting to see whether different riding styles lead to different
measurement results.
The proposed algorithm for identifying potholes/humps currently uses a fixed acceleration
threshold (δa ) (we set it as 2 g for the experiments) to identify sparks in the signal data. While the
algorithm and the threshold value worked very well, as shown in the experiment results, the algorithm
could be further improved to make it a self-adaptive algorithm.
An important implication of this work should be also mentioned. As shown in the experiments,
the proposed bicycle-mounted smartphone-based method can be used to map the road surface
roughness and identify potholes/humps on roads. Currently, cycling in cities is becoming more
and more popular. Therefore, potentially a huge number of contributors are available to map the
road surface roughness around the areas they travel, using their own smartphones. This would be
a nice add-on to the VGI project, OpenStreetMap. Having this kind of information available would
allow many interesting and innovative applications, such as navigation and routing for pedestrian,
bicycles and wheelchairs.

6. Conclusions and Future Work

Recent years have seen increasing interest in mapping the surface roughness of roads, as it is
closely related to driving safety, passenger comfort, road maintenance and so on. While several
professional systems, such as vehicle-mounted laser scanning, exist, they can only be used for
motorable roads. Until now, there has been no effective method for measuring the road surface
roughness of pedestrian and bicycle lanes. This article proposed a method for measuring the
road surface roughness of pedestrian and bicycle lanes based on GPS and accelerometer sensors
on bicycle-mounted smartphones. We mainly focused on the IRI metric.
A proof-of-concept and three experiments were designed to evaluate the proposed method.
Results of the implementation and the experiments showed that the IRI values measured by the
proposed method strongly and positively correlate with those measured by professional instruments,
and the proposed method was able to recognize potholes and humps (and their locations) passed by
the bicycles.
In conclusion, bicycle-mounted smartphones can be used to measure the surface roughness of
roads as well as to identify potholes and humps along the roads. This approach is especially useful for
those roads that are not accessible by professional instruments (which are often installed on motor
vehicles like cars), such as pedestrian lanes and cycle lanes. This work enables us to further study
the feasibility of large-scale crowdsourcing of road surface roughness (like OpenStreetMap) with
bicycle-mounted smartphones, as cycling has become more and more popular in recent years.
As a next step, we will further improve the proposed method by considering different
bike-riding styles, travel speeds, bicycle models, smartphone installation positions, and road conditions
(e.g., uphill and downhill). In the meantime, we would like to develop a self-adaptive algorithm to
identify potholes and humps on the roads. Furthermore, as mentioned previously, we would like to
expand this work towards large-scale crowdsourcing of road surface roughness. To achieve this aim,
we plan to implement a web platform allowing cyclists to upload their recorded GPS and accelerometer
data, as well as provide some other information, like bike type and relevant parameters (e.g., length of
wheelbase), and installation position of smartphones. The IRI values will then be derived, stored,
and visualized in a map view. Algorithms will then be developed to combine results from different
cyclists for the same road segment. These IRI values can be also integrated to enrich the VGI project,
OpenStreetMap, with surface roughness information on pedestrian and cycle lanes. This will allow
many interesting applications, such as navigation and routing for bicycles, pedestrian, and wheelchairs,
and road maintenance.

Acknowledgments: The research work presented in this paper is supported by National Key R&D Program
of China (2016YFE0131600) and NSFC (National Natural Science Foundation of China) project No. 41371433.
Sensors 2018, 18, 914 16 of 17

We thank Yiqiu Tan from the School of Transportation Science and Engineering, Harbin Institute of Technology
for her share on knowledge of road surface roughness. Thank Quanzhen Xu and Dayou Zhu from the Nanjing
Highway Scientific Research Institute for measuring the IRI in the test area with the car mounted laser pavement
scanner in our experiment. Both reviewers provided very constructive comments, which help to significantly
improve the paper. We appreciate their time and effort.
Author Contributions: K.Z., J.Shen, H.H., M.W. conceived and designed the study; J.Shen organized the overall
research; H.H. developed the recognition algorithm; M.W. developed the “RoadSR” mobile application; K.Z. and
M.W. performed the experiment and analyzed the data; all authors wrote the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Smith, K.; Ram, P. Measures and Specifying Pavement Smoothness; FHWA: Washington, DC, USA, 2016.
2. Chang, C.L.; He, S.; Qian, J.Z.; Mao, L.J. Pavement flatness test technology summary. J. Highw. Transp.
Res. Dev. 2006, 4, 6,7,13.
3. Perera, R.W.; Kohn, S.D.; Rada, G.R. LTPP Manual for Profile Measurements and Processing;
FHWA: McLean, VA, USA, 2008.
4. Sayers, M.W.; Karamihas, S.M. The Little Book of Profiling: Basic Information about Measuring and Interpreting
Road Profiles; University of Michigan Transportation Research Institute: Ann Arbor, MI, USA, 1998.
5. Coenen, T.B.J.; Golroo, A. A review on automated pavement distress detection methods. Cogent Eng.
2017, 4, 1374822. [CrossRef]
6. Li, X.; Goldberg, D.W. Toward a mobile crowdsensing system for road surface assessment. Comput. Environ.
Urban Syst. 2018. [CrossRef]
7. Douangphachanh, V.; Oneyama, H. A Model for the Estimation of Road Roughness Condition from Sensor
Data Collected by Android Smartphones. JJSCE SerD3 Infrastructure Plan. Manag. 2014, 70, I_103–I_111.
[CrossRef]
8. Harikrishnan, P.M.; Gopi, V.P. Vehicle Vibration Signal Processing for Road Surface Monitoring. IEEE Sens. J.
2017, 17, 5192–5197. [CrossRef]
9. Mobasheri, A. A Rule-Based Spatial Reasoning Approach for OpenStreetMap Data Quality Enrichment;
Case Study of Routing and Navigation. Sensors 2017, 17, 2498. [CrossRef] [PubMed]
10. Mobasheri, A.; Huang, H.; Degrossi, L.C.; Zipf, A. Enrichment of OpenStreetMap Data Completeness with
Sidewalk Geometries Using Data Mining Techniques. Sensors 2018, 18, 509. [CrossRef] [PubMed]
11. Gartner, G.; Huang, H.; Millonig, A.; Schmidt, M.; Ortag, F. Human-centred mobile pedestrian navigation
systems. Mitteilungen der Österreichischen Geographischen Gesellschaft 2011, 153, 237–250.
12. Sayers, M.W. On the calculation of international roughness index from longitudinal road profile.
Transp. Res. Rec. 1995, 1501, 1–12.
13. ASTM E867-06. Standard Terminology Relating to Vehicle-Pavement Systems; ASTM International:
West Conshohocken, PA, USA, 2012.
14. Haas, R.C.G.; Hudson, W.R.; Zaniewski, J.P. Modern Pavement Management; Krieger Publishing:
Malabar, FL, USA, 1994.
15. Shi, N.; Chen, Z.D.; Liu, F.M. Rapid detection method of asphalt pavement distress. J. Guangxi Univ.
2013, 4, 865–871.
16. Chen, Y. Fuzzy Skyhook Surface Control Using Micro-Genetic Algorithm for Vehicle Suspension
Ride Comfort. In Intelligent Computational Optimization in Engineering: Techniques and Applications;
Köppen, M., Schaefer, G., Abraham, A., Eds.; Springer: Berlin/Heidelberg, Germany, 2011; pp. 357–394,
ISBN 978-3-642-21705-0.
17. Pavementinteractive Roughness. Available online: http://www.pavementinteractive.org/roughness/
(accessed on 11 February 2018).
18. González, A.; O’brien, E.J.; Li, Y.-Y.; Cashell, K. The use of vehicle acceleration measurements to estimate
road roughness. Veh. Syst. Dyn. 2008, 46, 483–499. [CrossRef]
19. Eriksson, J.; Girod, L.; Hull, B.; Newton, R.; Madden, S.; Balakrishnan, H. The pothole patrol: Using a mobile
sensor network for road surface monitoring. In Proceedings of the International Conference on Mobile
Systems, Applications, and Services, Breckenridge, CO, USA, 17–20 June 2008; pp. 29–39.
Sensors 2018, 18, 914 17 of 17

20. Mednis, A.; Strazdins, G.; Liepins, M.; Gordjusins, A.; Selavo, L. RoadMic: Road Surface Monitoring Using
Vehicular Sensor Networks with Microphones. In Proceedings of the International Conference on Networked
Digital Technologies, Prague, Czech Republic, 7–9 July 2010; pp. 417–429.
21. Astarita, V.; Caruso, M.V.; Danieli, G.; Festa, D.C.; Giofrè, V.P.; Iuele, T.; Vaiana, R. A Mobile Application for
Road Surface Quality Control: UNIquALroad. Procedia Soc. Behav. Sci. 2012, 54, 1135–1144. [CrossRef]
22. Astarita, V.; Vaiana, R.; Iuele, T.; Caruso, M.V.; Giofrè, V.P.; De Masi, F. Automated Sensing System for
Monitoring of Road Surface Quality by Mobile Devices. Procedia Soc. Behav. Sci. 2014, 111, 242–251.
[CrossRef]
23. Vaiana, R.; Iuele, T.; Astarita, V.; Caruso, M.V.; Tassitani, A.; Zaffino, C.; Giofrè, V.P. Driving Behavior
and Traffic Safety: An Acceleration-Based Safety Evaluation Procedure for Smartphones. Mod. Appl. Sci.
2014, 8, 88. [CrossRef]
24. Zhao, B.; Nagayama, T.; Makihata, N.; Toyada, M.; Takahashi, M.; Ieiri, M. IRI Estimation by the Frequency
Domain Analysis of Vehicle Dynamic Responses and Its Large-scale Application. In Proceedings of
the International Conference on Mobile and Ubiquitous Systems: Computing Networking and Services,
Hiroshima, Japan, 28 November–1 December 2016; pp. 41–46.
25. Islam, S.; Buttlar, W.G.; Aldunate, R.G.; Vavrik, W.R. Measurement of Pavement Roughness Using
Android-Based Smartphone Application. Transp. Res. Rec. J. Transp. Res. Board 2014, 2457, 30–38. [CrossRef]
26. Mednis, A.; Strazdins, G.; Zviedris, R.; Kanonirs, G.; Selavo, L. Real time pothole detection using Android
smartphones with accelerometers. In Proceedings of the International Conference on Distributed Computing
in Sensor Systems and Workshops, Barcelona, Spain, 27–29 June 2011; pp. 1–6.
27. Alessandroni, G.; Carini, A.; Lattanzi, E.; Freschi, V.; Bogliolo, A. A Study on the Influence of Speed on Road
Roughness Sensing: The SmartRoadSense Case. Sensors 2017, 17, 305. [CrossRef] [PubMed]
28. Tai, Y.C.; Chan, C.W.; Hsu, Y.J. Automatic road anomaly detection using smart mobile device. In Proceedings
of the Conference on Artificial Intelligence and Applications, Hsinchu, Taiwan, 18–20 November 2010.
29. Chang, J.R.; Su, Y.S.; Huang, T.C.; Kang, S.C.; Hsieh, S.H. Measurement of the International Roughness
Index (IRI) Using an Autonomous Robot (P3-AT). In Proceedings of the 26th International Symposium on
Automation and Robotics in Construction (ISARC 2009), Austin, TX, USA, 24–27 June 2009.
30. Ferregut, C.; Abdallah, I.; Melchor, O.; Nazarian, S. Artificial Neural Network-Based Methodologies for Rational
Assessment of Remaining Life of Existing Pavements; Center for Highway Materials Research, The University of
Texas at El Paso: El Paso, TX, USA, 1999.
31. Lin, J.D.; Hsiao, L.H.; Student, G. Correlation Analysis between International Roughness Index (IRI) and
Pavement Distress by Neural Network. In Proceedings of the 82th Annual Meeting of the Transportation
Research Board, Washington, DC, USA, 12–16 January 2003.
32. Perttunen, M.; Mazhelis, O.; Cong, F.; Kauppila, M.; Leppänen, T.; Kantola, J.; Collin, J.;
Pirttikangas, S.; Haverinen, J.; Ristaniemi, T. Distributed Road Surface Condition Monitoring Using
Mobile Phones. In Proceedings of the International Conference on Ubiquitous Intelligence and Computing,
Banff, AB, Canada, 1–4 September 2011; pp. 64–78.
33. Du, Y.; Liu, C.; Wu, D.; Li, S. Application of Vehicle Mounted Accelerometers to Measure Pavement
Roughness. Int. J. Distrib. Sens. Netw. 2016, 12, 8413146. [CrossRef]
34. Witte, T.H.; Wilson, A.M. Accuracy of non-differential GPS for the determination of speed over ground.
J. Biomech. 2004, 37, 1891–1898. [CrossRef] [PubMed]
35. Ranacher, P.; Brunauer, R.; Trutschnig, W.; Van der Spek, S.; Reich, S. Why GPS makes distances bigger than
they are. Int. J. Geogr. Inf. Sci. 2016, 30, 316–333. [CrossRef] [PubMed]
36. City of Copenhagen Bicycle Statistics. Available online: https://web.archive.org/web/20131212093813/http:
//subsite.kk.dk/sitecore/content/Subsites/CityOfCopenhagen/SubsiteFrontpage/LivingInCopenhagen/
CityAndTraffic/CityOfCyclists/CycleStatistics.aspx (accessed on 2 February 2018).

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).