Content uploaded by Emanuele Lattanzi
Author content
All content in this area was uploaded by Emanuele Lattanzi on May 20, 2015
Content may be subject to copyright.
Content uploaded by Emanuele Lattanzi
Author content
All content in this area was uploaded by Emanuele Lattanzi on Feb 02, 2015
Content may be subject to copyright.
Geospatial Data Aggregation and Reduction in
Vehicular Sensing Applications: the Case of Road
Surface Monitoring
Valerio Freschi, Saverio Delpriori, Lorenz Cuno Klopfenstein,
Emanuele Lattanzi, Gioele Luchetti, Alessandro Bogliolo
DiSBeF, University of Urbino, Urbino, Italy
Email: {valerio.freschi, saverio.delpriori, lorenz.klopfenstein, emanuele.lattanzi, gioele.luchetti, alessandro.bogliolo}@uniurb.it
Abstract—Mobile devices present several features which make
them attractive as enabling technology for crowdsensing systems.
In particular, their spectrum of sensing capabilities, together with
consolidated diffusion and ease of use contribute to an increasing
adoption in different mobility-based sensing scenarios. On the
other hand, the availability of massive volumes of geospatial
data provided by large-scale distributed sensing systems prompts
the need for innovative approaches to efficient data gathering
and processing. Data reduction strategies are often necessary in
order to cope with challenges posed by these volumes, for instance
when dealing with real-time visualization of query results. In this
paper we present a data reduction and aggregation approach for
mitigating the imp act of data size in a vehicular sensing appli-
cation aimed at monitoring the roughness of road surfaces. Data
collected by smartphones on board of vehicles is progressively
thinned at different levels of the proposed architecture through
sampling and spatial/temporal aggregation. Preliminary results
show that the proposed methodology p rovides substantial benefits
in terms of reduced impact of data while, at the same time, it
enables full exploitation of statistical error compensation.
I. INTRODUCTION
The increasing diffusion of mobile embedded sensing
devices with wireless communication capabilities opens the
way to unprecedented opportunities in the development of
large scale crowdsensing systems [1]. Current smartphones,
in particular, are commonly equipped with sensors that can
continuously monitor several physical quantities (e.g. acceler-
ation). This provides, in combination with location coordinates
available through GPS or other localization systems, a r ich
source of geo-referenced information. Moreover, the perva-
siveness of these devices makes them the enabling technology
for designing large scale mobile distributed systems aimed at
massive sensing, either volunteer or incentive-based [2].
Needless to say, this perspective also poses significant
challenges to the research community in order to build systems
capable of efficiently and accurately collecting, processing and
making available this wealth of data, ranging from system
architecture to algorithmic design, from communication proto-
cols to database. An additional dimension is represented by the
massive nature of geo-referenced data to be handled by these
hardware/software systems, especially when vehicular sensing
applications are foreseen, which entail further peculiar issues
to be addressed [3]. As such, a common feature of several
research studies in recent scientific literature is represented by
efforts made towards effectively scaling systems for storage
and analysis of big geospatial sensing data.
In this paper we introduce a system for reduction and
aggregation of geospatial data in vehicle-based monitoring
applications. In particular, we describe a novel approach to
manage data produced in a crowdsensing application for road
surface quality control by means of spatial and temporal
aggregation techniques. We demonstrate that the proposed
distributed architecture is suitable to reduce the burden of
large scale geo-referenced data volumes produced by sens-
ing devices mounted on common smartphones. This system
architecture provides the opportunity for fine grained data
gathering and batch processing while, at the same time, it
enables effective visualization and real-time analysis. Progres-
sive spatial and temporal aggregation of data is performed at
different levels of the proposed architecture (from user mobile
devices on vehicles, up to the cloud) resulting into a significant
reduction (w.r.t. raw data logged from smartphones) of the
amount of data to be analyzed and visualized.
The paper is organized as follows: in Section II we sum-
marize the state of the art of related scientific literature. In
Section III we describe the proposed crowdesensing architec-
ture and data aggregation strategy. In Section IV we discuss
experimental results. In Section V we draw final conclusions.
II. RELATED WORK
A huge body of literature has flourished in the last decade
around the vast field of mobile sensing information systems.
In this section we try to summarize which are, in our opinion,
some of the main trends related to the topics which are the
scope of this paper.
Crowdsensing is an increasingly popular paradigm for
gathering si gnificant amounts of data from active communities
of users (i.e. participatory sensing) or agents opportunistically
carrying on sensing tasks (i.e. opportunistic sensing) [1]. Data
is usually sensed by mobile devices whose location can be
tracked with a given precision so that useful geo-referenced
information can be obtained and geographic information sys-
tems (GIS) can be exploited for analytics extraction. The ever
increasing widespread diffusion of commodity smartphones
and the availability of several sensors (e.g. accelerometers,
GPS, ambient light, microphones, cameras, etc.) on board of
them, make these devices the ideal candidate sensing platform
for many large scale mobile monitoring tasks [1], [2].
A. Vehicle-based sensing system architectures
Eriksson et al. proposed in 2008 CarTel, a system for road
surface monitoring focused on pothole detection by means of
embedded accelerometers and GPS sensors mounted in cars
equipped with embedded microprocessors [4]. Mohan et al.
developed Nericell, a smartphone-based mobile sensing system
aimed at detecting traffic conditions, bumps, and honking
events by integrating audio and acceleration data from micro-
phones and triaxial accelerometers mounted on smartphones
[5]. Vtrack is a system that enables road traffic delay est imation
using mobile phones, with emphasis on energy consumption
and noise compensation [6]. A follow-up paper from the same
research group described an approach to trajectory mapping
from cellular GSM fingerprints instead of WiFi and GPS traces
[7]. A prominent example of large-scale system based on
mobile sensing is represented by OpenSense, a system aimed at
monitoring air pollution by means of sensor stations deployed
on public transport vehicles and through participative sensing
from citizens equipped with ad hoc pocket sensors or enhanced
smartphones [8]. A system for road surface collaborative mon-
itoring, called SmartRoadSense, has been recently introduced
[9]. SmartRoadSense is a mobile/cloud architecture designed
for continuous monitoring of road surface quality conditions,
estimated by means of a roughness index computed on board
of smartphones and stored/processed/visualized in cloud.
B. Big geospatial data analysis
Mobile crowdsensing inherently implies dealing with ex-
pected large volumes of data that prompt for efficient and
scalable solutions both at system and at algorithmic level.
The growing research field of the so called spatial BigData
mainly refers to the development of novel methodologies
and approaches to address all issues related to geospatial
massive datasets. Within this framework, some recent works
highlighted the need for new flexible approaches and, at the
same time, pointed out the inadequacy of more tr aditional
approaches rooted in database research [3], [10]. Moreover,
while modern database management systems routinely face
problems related to efficient storage, search, and processing of
data, visualization systems need to be re-designed in order to
keep pace with BigData. According to this perspective, Keller
et al. introduced Vizzly, a middleware designed for interactive
browsing of large data sets in sensor networks applications
which has been integrated in the OpenSense project framework
[11]. Battle et al. stressed the lack of a thorough support of
visualization systems to larger s cales. In order t o overcome
some of the related challenges they proposed ScalaR, a system
for dynamic resolution reduction to be applied when results
of a query are expected to be too big to be handled by
standard data base management systems (DBMS). Reduction
is achieved t hrough a chain of aggregation, sampling, and
filtering operations [12].
III. PROPOSED ARCHITECTURE
This section presents the architecture of SmartRoadSense
and the solutions adopted for data gathering, aggregation,
reduction, and visualization. Scalability issues to be faced at
each stage will be then discussed in next section.
The algorithmic pipeline is distributed at different levels
of SmartRoadSense, a cloud-based system for collaborative
Fig. 1. System overview.
road quality monitoring designed for estimating the surface
roughness of roads by means of a smartphone’s triaxial ac-
celerometers.
The architecture is based on three main components,
schematically represented in Figure 1: i) a mobile application
running on Android devices which reads the data provided by
the embedded GPS and accelerometers and computes every
second a geo-tagged estimate of the roughness of the road
surface; ii) a server that gathers roughness indexes fr om all
the smartphones running the SmartRoadSense application and
makes use of OpenStreetMap [13] to perform spatio-temporal
aggregation and reduction, and iii) a cloud-based front-end for
graphical visualization.
Our approach to data reduction is obtained through dif-
ferent algorithmic strategies developed at different levels of
this system. Figure 2 represents the various phases of the
implemented algorithmic pipeline, labelled (a), (b), (c), (d) and
(e). In the following we sketch the main tasks performed at
each level of the pipeline.
• Phase (a). During phase (a), synthetic numerical
values (called Roughness indexes, RI) are computed
real-time on board of smartphones from accelerations
sensed by the devices. These values, which provide a
reasonable esti mate of the roughness of the underlying
road monitored by the vehicle, represent the result
of a first level of data reduction. In fact, sampling
sensed data according to GPS sampling capabilities
and summarizing the information coming from three
axes into a single numerical estimate value provides
sizeable data compression w.r.t. raw data.
• Phase (b). This step consists of data serialization
and storage of roughness indexes (with geographical
coordinates and timestamp) on the memory of the
smartphones. Batches of stored data are periodically
transmitted to a remote server through GSM channels.
• Phase (c). This phase, implemented on the cloud,
performs consistent spatial aggregation of points re-
ceived by the back-end. Each point i s mapped onto
a map database and aggregated according to specific
geometric constraints. This makes it possible to con-
sistently map the sensed physical quantities of several
adjacent points into a single aggregate, providing
100 x
epochs
t
0
t
-1
t
-2
t
-3
t
-4
t
-5
t
-6
t
-7
t
-n-1
t
-n
...
Fig. 2. Algorithmic pipeline.
further reduction of the number of points which are the
final target of a visualization task and also smoothing
outliers thanks to statistical compensation.
• Phase (d). Phase (d) regards temporal aggregation.
Weighted average values of the monitored quantities
are periodically computed, resulti ng into manifold
benefits. The database is kept updated with last sig-
nificant changes (incrementally down-weighting older
points), data to be (also visually) analyzed is further
thinned, and statistical robustness deriving from mul-
tiple measurements associated to a given location can
be exploited.
• Phase (e). This last step entails the visualization of
data (namely, the aggregated roughness index) pro-
vided as output by previous steps, by means of a
graphical front-end.
In the next subsections, we further detail some aspects
of the implementation of data reduction and aggregation
algorithms, referring them to the three components of the
SmartRoadSense architecture.
A. Smartphone level
The first layer of the system architecture consists of an
Android application which is in charge of gathering data from
the smartphone’s sensors, namely GPS and triaxial accelerom-
eters and implements phases (a) and (c) of the algorithmic
pipeline of Figure 2. Since the sampling frequency of GPS
mounted on current smartphones is much lower than that of
triaxial accelerometers on the same devices (typically 1Hz
and 100 Hz, respectively) the former represents a constraint
on the spatial resolution that can be exploited for a first-cut
reduction of the data to be collected. In fact, the developed
mobile application works on windows of 100 samples (i.e. 100
seconds, taking the above mentioned sampling frequencies)
and computes, for each window, an aggregated roughness
index RI. RI represents the average value of the power of the
prediction errors (named PPE’s) computed when a prediction
filter is applied [9]. Prediction errors are computed for each
time window and along each of the three axial components
of the acceleration. Given the power of the prediction errors
P P E
x
, P P E
y
, and P P E
z
computed by applying a Linear
Predictive Coding algorithm (LPC for short) to the collected
samples, the roughness of the road surface upon which the
vehicle travels is estimated as their arithmetic average.
This estimate provides significant information on the qual-
ity of road surface, given the capability of the LPC algorithm
of filtering out (up to a certain degree) spurious components of
the acceleration signals (engine vibrations, gravitation, inertial
forces, etc.). RI values represent a compact sketch that can
be usefully exploited i n a collaborative setting. In fact, the
contribution of many roughness indexes can be taken into
account to represent the quality of a given road in a specific
geographical position, thus providing a worth of meaningful
information that can be properly averaged. RI values annotated
with a track identification code and with time and position
references are stored in memory according to Java serialization
format and periodically transmitted in batch to a remote server
through GSM connection. Data payload i s encoded in JSON
and HTTP protocol is used for data transfer to the cloud.
B. Cloud back-end level
A server application has been implemented (the YouSense
server of Figure 1) that exposes a set of application program
interfaces (in particular RESTful API’s) in order to allow
permitted users to upload data. The collection back-end has
been designed exploiting PostgreSQL with PostGIS extension
(for geospatial processing) as database. This layer of the
system architecture is in charge of the spatial and temporal
data aggregation corresponding to phases (c) and (d) of Figure
2. First of all it makes use of a map matching algorithm in
order to map each newly received point (composed of spatio-
temporal coordinates, RI and metadata) to its closest road.
Road cartography is provided by OpenStreetMap and map
matching is currently implemented by associating points to
geometrically closest road segments. Artifacts are removed by
a simple post-processing that takes new data points sorted by
timestamp. The list of these points is analyzed using a window
of 3 points, p
1
, p
2
, p
3
. If p
1
and p
3
are matched to the same
road, while p
2
is associated to another road, p
2
is matched
back to the road of p
1
and p
3
, since it is assumed that the road
change was misdetected. Needless to say, several alternatives
could be taken into consideration to enhance the accuracy of
mapping [6].
1) Spatial aggregation: Spatial aggregation is obtained by
uniformly sampling road segments matched by points during
the map-matching phase. Given an input parameter (termed
Spatial Sampling Factor, or SSF) we sample each road uni-
formly every SSF meters obtaining a set of landmark points.
We then track all points falling within a circle of given radius
(called Coverage Circle Radius, CCR) centered around each
landmark point. The average RI value to be associated to
the landmark is then computed as the average value of RIs
associated to points falling within the circle. Data values are
weighted by their distance from the landmark point (i.e. from
the center of the circle) using an inverse exponential function
and annotated with their timestamp.
2) Temporal aggregation: Temporal aggregation (corre-
sponding to phase (d) of the pipeline reported i n Figure 2)
is achieved by aggregating all roughness values computed
for the same position on the same road. Values contributing
to the same point are sorted by descending timestamp. The
contribution of each roughness value decreases exponentially
in time, thus the latest computed value has the highest weight,
while older values are steeply down-weighted. This exponen-
tial decay is simply implemented by updating the temporal
estimate (a daily estimate is a reasonable time horizon in road
quality monitoring) as the average between the current value
and the previous aggregated estimate (regardless of the days
elapsed since last update). This corresponds to an exponential
decay if new esti mates are provided every day. If there are
gaps, they are implicitly filled by assuming the daily value
equals previous estimates.
C. Front-end level
The SmartRoadSense graphical front-end is based on Car-
toDB, a cloud service for visualization t asks of geographical
maps and associated overlays [14]. The service offers web
APIs that allow the back-end to upload updated roughness
values for each geographic point. It also provides functions for
retrieving a list of roughness points for all roads inside a given
geographic area. This is used to populate a map of roughness
points and to display it as an overlay to a geographical map
(e.g. Google Maps [15]), thus implementing the functionalities
associated to phase (e) of Figure 2. Each point is graphically
displayed and filled according to a linear color map that
represents the RI values (green to red, from lowest to highest),
thus providing useful visual information about the roughness
experienced by vehicles travelling along a given road segment.
IV. SCALABILITY ANALYSIS
A full-fledged prototype of SmartRoadSense was devel-
oped at the University of Urbino and systematically used for
one months to monitor the roughness of the roads along a
path of 275Km traveled by public transport twice a day. The
selected path went through 744 roads of the OpenStreetMap
DB adopted in the test bed. Buses were equipped with Android
smartphones (namely, Motorola Moto G) running the mo-
bile SmartRoadSense application. The roughness index values
computed on board of mobile devices once per second were
opportunistically transmitted to the server exploiting either
Wi-Fi or m2m 3G connections when available. Connection
attempts were automatically performed by the application
every 15 minutes.
Spatial aggregation was performed by the server with a
sampling step of 20m along the travelled r oads, corresponding
to a coverage circle radius (CCR) of 40m. Aggregated data
were re-computed every day and stored as geo-tagged time
series. Temporal aggregation was then performed at each
sampling point to obtain a scalar value to be graphically
represented on the map as the combination of all the elements
of the roughness time series associated with that particular
point.
Scalability is analyzed and discussed in the following
section focusing on each single step of the data flow.
A. Data size
In order to investigate the impact of the different phases of
the proposed data flow we studied how the size of the payload
changes at each step, taking into account both the information
content and the encoding adopted. Moreover, we analyzed the
overhead introduced either by the protocols or by the data base
management system for performance optimization.
Figure 3 summarizes the results of this analysis. The figure
is divided into three conceptual sections, from left to right.
On the left a bar plot is used to represent data size (split
into payload, protocol overhead, and DBMS overhead) at each
distinct phase starting from raw sensor data (at the bottom)
up to CartoDB visualization (at the top). The effects of data
serialization, JSON encoding, and HTTP transmission are also
represented. Data size is expressed in bits per sample (bps),
which corresponds to bits per second in the first 5 steps and
to bits per point after spatial aggregation.The phases of the
data flow, labelled (a), (b), (c), (d), and (e) as in Figure 2,
are graphically represented in the middle, pointing out the
element of the architecture involved at each step (smartphones,
cloud, front-end) and the communication among them. Finally,
multiplicative scale factors are reported on the right, using
circles to denote the steps affected by each of them. The
scale factors are: the absolute number of seconds of activity
of the system (s ecs), the average number of simultaneous
users collecting roughness indexes at each second (users), the
absolute number of days of activity of the system (days) which
corresponds to the length of time series associated with each
sampling point, and the total length of the monitored streets
(length) which is upper bounded by the total length of the
streets of the underlying OpenStreetMap DB. White circles
are used to mark phases which are not critically affected by
scale factors in terms of performance. For instance, the number
of users doesn’t impact processing steps carried on board of
smartphones, since we have one device per user.
In the following we detail the results reported in the bar
plot of Figure 3. For what concerns the size of the payload, it is
worth noticing that it depends on the amount of information to
be conveyed but also on the type of encoding, thus determining
different space requirements even without processing steps.
1) Raw sensor data: Payload of raw sensor data is com-
posed of: three values encoding the accelerations (32 bits
each), two values encoding longitude and latitude (64 bits
each) and two values representing other data (termed bearing
and accuracy) to be possibly used for post-processing (each
parameter being 32 bits long). This results into 9856 bits for
each sample of data.
2) Roughness index: After on board-processing, the result-
ing roughness index needs only 320 bits per sample for its
encoding because we need only to keep a single RI value
instead of three acceleration values.
Fig. 3. Data size and scaling factors.
3) Java binary serialization: Since roughness indexes are
written on the memory of mobile devices for batch trans-
mission, we need to take into account the overhead due to
this process (called Java binary serialization). In particular,
the size of the payload is increased because of the encoding
of additional required information (an ID number, start time,
duration, etc.) amounting to 704 bits per sample, while the
overall protocol accounts for a 6504 bits of overhead.
4) JSON over HTTP: In order to be sent over HTTP data
packets are JSON encoded. The resulting payload is 1368 bits
long, while the contribution of protocols to overhead is 1016
bits (JSON), 22 bits (HTTP) and 5 bits (JSON over HTTP):
in total 1043 bits per sample.
5) PostgreSQL storage: When data is received on the
cloud, it is stored in PostgreSQL format. While 1024 bits are
sufficient for encoding the payload, the DBMS introduces a
significant overhead of 4736 bits, that are to be added to 1760
bits needed by the indexing s tructures.
6) PostgreSQL processed: The impact of aggregation at
this phase of the algorithmic pipeline is apparent, since the
payload reduces to 512 bits for each sample, while the over-
head decreases to 802 bits.
7) Final CartoDB data: Data needed by the visualization
front-end consists of 320 bits (needed for encoding geometry
and roughness index) of payload and remaining 2528 bits
required by CartoDB as database overhead.
The effectiveness of data reduction and aggregation strate-
gies is evident from the results reported in Figure 3, previously
described. It is also worthwhile to notice that scalability analy-
sis also clearly points out the impact of database management
systems on the data size. For instance, CartoDB makes use
of indexes in order to provide effective visualization support.
Nonetheless, the adoption of these optimization structures
leads to substantial increase of data dimension (up to 88.7%
of the whole size).
V. CONCLUSION
The increasing diffusion of mobile devices with sensing
and communication capabilities (i.e. smartphones) provides
the opportunity for capillary crowdsensing applications. The
enormous amount of data potentially produced in these set-
tings raises the question of how to handle it for building
efficient analytics frameworks which are the ultimate target of
these systems. In this paper we introduced a data reduction
and aggregation approach aimed at mitigating t he impact
of geospatial BigData in vehicular sensing applications. Our
approach consists of a sequence of geometric sampling and
temporal aggregation steps implemented at different levels
of SmartRoadSense, a system architecture that supports road
quality monitoring. Experimental results show that the pro-
posed methodology is beneficial in terms of the reduced impact
of data size in geospatial applications, while it provides full
support to exploit statistical robustness in a massive distributed
sensing environment.
REFERENCES
[1] R. K. Ganti, F. Ye, and H. Lei, “Mobile crowdsensing: current state and
future challenges,” Communications Magazine, IEEE, vol. 49, no. 11,
pp. 32–39, 2011.
[2] N. D. Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury, and
A. T. Campbell, “A survey of mobile phone sensing,” Communications
Magazine, IEEE, vol. 48, no. 9, pp. 140–150, 2010.
[3] S. Shekhar, V. Gunturi, M. R. Evans, and K. Yang, “Spatial big-data
challenges intersecting mobility and cloud computing,” in Proceedings
of the Eleventh ACM International Workshop on Data Engineering for
Wireless and Mobile Access, ser. MobiDE ’12. New York, NY, USA:
ACM, 2012, pp. 1–6.
[4] J. Eriksson, L. Girod, B. Hull, R. Newton, S. Madden, and H. Balakr-
ishnan, “The pothole patrol: using a mobile sensor network for road
surface monitoring,” in Proceedings of the 6th international conference
on Mobile systems, applications, and services. ACM, 2008, pp. 29–39.
[5] P. Mohan, V. N. Padmanabhan, and R. Ramjee, “Nericell: rich mon-
itoring of road and traffic conditions using mobile smartphones,” in
Proceedings of the 6th ACM conference on Embedded network sensor
systems. ACM, 20 08, pp. 323–336.
[6] A. Thiagarajan, L. Ravindranath, K. LaCurts, S. Madden, H. Balakrish-
nan, S. Toledo, and J. Eriksson, “Vtrack: accurate, energy-aware road
traffic delay estimation using mobile phones,” in Proceedings of the
7th ACM Conference on Embedded Networked Sensor Systems. ACM,
2009, pp. 85–98.
[7] A. Thiagarajan, L. Ravindranath, H. Balakrishnan, S. Madden, and
L. Giraud, “Accurate, low-energy trajectory mapping for mobile de-
vices,” in Proceedings of the NSDI, 2011, pp. –.
[8] K. Aberer, S. Sathe, D. Chakraborty, A. Martinoli, G. Barrenetxea,
B. Faltings, and L. Thiele, “Opensense: open community driven sensing
of environment,” in Proc. of the 1st Intl Workshop on GeoStreaming
(IWGS 10), 2010, pp. 39–42.
[9] G. Alessandroni, L. Klopfenstein, S. Delpriori, M. Dromedari,
G. Luchetti, B. Paolini, A. Seraghiti, E. Lattanzi, V. Freschi, A. Carini,
and A. Bog liolo, “Smartroadsense: Collaborative road surface condition
monitoring,” in Proc. of Eighth International Conference on Mobile
Ubiquitous Computing, Systems, Services and Technologies, Ubicomm
2014. IARIA, 2014, p. accepted to.
[10] B. Simion, D. N. Ilha, A. D. Brown, and R. Johnson, “The price of
generality in spatial indexing,” pp. 8–12, 2013.
[11] M. Keller, J. Beutel, O. Saukh, and L. Thiele, “Visualizing large sensor
network data sets in space and time with vizzly,” in Local Computer
Networks Workshops (LCN Workshops), 2012 IEEE 37th Conference
on. IEEE, 2012, pp. 925–933.
[12] L. Battle, M. Stonebraker, and R. Chang, “Dynamic reduction of
query result sets for interactive visualizaton,” in Big Data, 2013 IEEE
International Conference on. IEEE, 2013, pp. 1–8.
[13] “Openstreetmap,” 2014. [Online]. Available:
http://www.openstreetmap.org
[14] Vizzuality, “Cartodb,” 2013. [Online]. Available: http://cartodb.com
[15] Google, “Google maps,” 2014. [Online]. Available:
http://maps.google.com
