Ieeepro techno solutions 2013 ieee embedded project greatest invehicular accident detection

522 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 2, FEBRUARY 2013
Safe Driving in LA: Report from the Greatest
Intervehicular Accident Detection Test Ever
Gustavo Marfia, Marco Roccetti, Alessandro Amoroso, and Giovanni Pau
Abstract—The UN Economic Commission’s Statistics of Road
Traffic Accidents report of 2011 shows that every year, about
150 000 human beings lose their lives on the roads of the western
world. Although it is a common belief that this figure could shrink
with the use of new sensor and communication technologies, un-
fortunately, none such systems have hit the road to date. Ideally, if
such technologies were put into place, vehicles could be part of a
vehicular ad hoc network (VANET) capable of spreading relevant
information about dangerous events (e.g., car accidents) to all
approaching drivers. However, all this is mainly supported by
simulation studies, as no practical results have been published to
date, revealing the effective performances of such systems at work.
In this paper, we fill this gap, presenting a detailed description of
the greatest experiments (a few thousand throughout the streets of
Los Angeles), to date, ever performed with an accident warning
system specifically devised for highway scenarios. In particular,
among all the possible candidate schemes, we ran a few thousand
experiments with the accident warning system algorithm that was
proven to be optimal in terms of bandwidth usage and covered
distance in realistic scenarios. Our experiments confirm what has
been observed before in theory and simulation, i.e., the use of such
a system can reduce, by as much as 40%, the amount of vehicles
involved in highway pileups.
Index Terms—Design for experiments, prototypes, vehicle
safety, vehicular ad hoc network (VANET) testbed, vehicular and
wireless technologies.
I. INTRODUCTION
THE SOCIETAL costs of car accidents in Europe, in year
2009, amounted to $160 billion according to the Har-
monised European Approaches for Transport Costing project
commissioned by the European Union (EU) [1]. Such value
only partially captures the plethora of problems that have their
origin in such type of events, as it is rather difficult to quantify,
for example, the costs that are imputable to the emotional
distress symptoms or to the increased pollution levels that
usually follow any motor vehicle accident.
Continuing with a policy that recognizes the negative impact
that car accidents have on society, in July 2011, the Transport
Manuscript received January 30, 2012 revised May 7, 2012 and August 31,
2012; accepted October 17, 2012. Date of publication October 24, 2012; date
of current version February 12, 2013. This work was supported by FIRB
DAMASCO, PRIN ALTERNET, NSF/BBN GENI projects, and TOYOTA
InfoTechnology Center, USA. The review of this paper was coordinated by
Dr. P. Lin.
G. Marfia is with the Computer Science Department, University of Bologna,
40126 Bologna, Italy, and also with the University of California, Los Angeles,
CA 90095 USA (e-mail: marfia@cs.unibo.it).
M. Roccetti and A. Amoroso are with the Computer Science Department,
University of Bologna, 40126 Bologna, Italy (e-mail: amoroso@cs.unibo.it;
roccetti@cs.unibo.it).
G. Pau is with the Computer Science Department, University of California,
Los Angeles, CA 90095 USA (e-mail: gpau@cs.ucla.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2012.2226484
Committee of the EU approved an ambitious initiative that
aims at halving the number of road accident victims by 2020
[2]. The path that leads to such aim certainly goes through
the use of advanced technologies and, in particular, through
the adoption of those that can prevent vehicles from being
involved in accidents rather than those that can simply improve
their general outcome. In fact, until now, most in-car solutions
have focused on reducing the damage suffered by passengers
and hence ameliorate the general consequences of an accident
instead of seeking for new ways apt to totally avoid such type
of events. Now, as cars have evidently become more resistant
to impacts, drivers more careful, and roads safer, it is natural to
expect that less space for improvement is left to any traditional
countermeasure.
In this application space, one of the most promising solutions
lies in the realm of communication technologies and is based
on the use of vehicular ad hoc networks (VANETs). When an
accident occurs, involved vehicles can alert all approaching ve-
hicles within a distance of a few hundreds of meters by sending
a short-range alert message. This message can then recursively
percolate throughout the platoon of vehicles that are traveling in
the direction of the accident. Cars, on receiving such message,
can smoothly stop and avoid being involved in the pileup. This
scenario, as it is, appears simple and trivial to implement, but
has been under investigation for the past decade, as it presents
at least two important challenges. The first, with no doubt,
regards the installation of a set of devices that compose the
safety equipment on all vehicles: a Global Positioning System
(GPS) receiver, an accelerometer sensor, and an 802.11 wireless
interface. The second regards the adoption of the best alert
distribution algorithm, whose requirement is to be reliable and
fast, as it should be able to deliver alert information in real time.
As such, the desired algorithm is capable of successfully deal-
ing with 1) the shared nature of the 802.11 channels, avoiding
falling into the well-known broadcast storm problem, and 2) the
challenges of mobile wireless channels, which vary in time and
space because of the fast and slow fading phenomena [3].
Now, although highway safety systems based on VANETs
have been thoroughly studied in the past years, the almost
totality of results has been obtained by means of computer
simulation, with little or no experimentation performed on
real-road scenarios. This biased research approach is certainly
dictated by the fact that the resources required to perform
realistic accident warning system tests are beyond the reach
of most, if not all, research teams. In November 2011, the
main car manufacturers (e.g., Daimler, GM, Ford, Toyota, etc.)
performed a public demonstration in Orlando, FL, without,
however, publishing any of the findings obtained from such a
testing campaign [4].
0018-9545/$31.00 © 2012 IEEE

MARFIA et al.: SAFE DRIVING IN LA: REPORT FROM THE GREATEST INTERVEHICULAR ACCIDENT DETECTION TEST EVER 523
The aim of this paper is to fill such gap, reporting on the
results drawn from the most extensive accident warning system
test performed to date, to the best of our knowledge, on the
streets and highways of Los Angeles, in August 2011. The
tests that have been conducted in Los Angeles involved two
research teams, i.e., one from the University of Bologna, and
one from the University of California, Los Angeles (UCLA),
and entailed performing one entire month of laboratory and
on-the-road experiments, for a total financial commitment that
exceeded $60 000, invested in equipment (including car leases)
and software development costs. To the best of our knowledge,
this experiment campaign is the most extensive, among those
whose results have been made available to date to the scientific
community, in terms of both 1) the number of kilometers driven
by a VANET testbed on a highway scenario and 2) the number
of accident warning system tests performed in reality (over
10 000 tests during that month). Evidence of the importance of
this campaign is witnessed also by the widespread international
media attention that it received (e.g., Wall Street Journal, BBC,
Discovery News, etc.), where also a team composed of journal-
ists, and not only scientists or researchers, followed one of the
experiments held on the road in Los Angeles [5]–[8].
Compared with the work published in [9], where a small
subset of our assessments have been first analyzed, here, we
provide a thorough description of the tests that have been
conducted on the road and of their results, this paper resulting
novel in all of its parts. Differently from the tests performed
more recently in Orlando by the world’s premium automakers,
instead, we produced publicly available data that can generally
benefit the scientific community.
To do all this, we implemented the accident warning system
broadcast algorithm that has been proven to be the optimal
one [15], [16]. Differently from all previous proposals, this
algorithm takes into account the fact that the vehicles’ trans-
mission ranges vary in time and that wireless links can be
anything but symmetric, presenting situations where between
two vehicles, only one can receive the messages transmitted
by the other. With this in mind, in [15], it is individuated that
each vehicle should choose as the best vehicle to retransmit
a given alert message at each hop the farthest spanning relay,
i.e., the relay that can retransmit the message the farthest away,
providing analytic and simulative proof that this is the best
possible strategy in a realistic situation.
Summarizing, the main contributions of this paper are the
following: 1) the dissemination of experimental results drawn
from the largest intervehicular accident warning system experi-
ments ever performed in a live setting; 2) a confirmation, drawn
from experimental data, of the theoretical and simulative results
accomplished by the accident warning system at test; and
3) further evidence that the deployment of such system can
reduce as much as 40% the number of vehicles involved in a
pileup.
The remainder of this paper is organized as follows. In
Section II, we discuss related work, whereas in Section III,
we provide a brief description of the tested accident warning
system. In Section IV, we sketch the experimental scenario,
whereas in Section V, we comment on experiment results. We
finally conclude with Section VI.
II. RELATED WORK
Since the most important contribution of this paper lies in
the construction of a realistic testbed for the assessment of a
VANET-based accident warning system, we begin this section
by illustrating the most relevant experiments performed, to date,
with real VANETs on the road. We then move on to describe the
stream of work in the domain of VANET-based accident warn-
ing systems that fall closest to ours. In fact, a wealth of research
has been carried out in the past decade in the space of VANET-
based applications and safety systems for urban and highway
scenarios [17]–[35], and here, we will be limited to describing
those approaches that are most relevant to the one that is at
test [15].
While a large share of VANET research has addressed, with
no doubt, the simulation of such networks for safety purposes,
the same cannot be said regarding their experimentation on the
road. Realistic experiments with accident warning systems on
highways have been for long impeded by the magnitude of
the testbeds required to perform relevant experiments. Hence,
we will here concentrate on those research efforts that have
obtained most from the limited available vehicular resources.
The authors of [26] carried out a prominent VANET multi-
hop experimentation utilizing 802.11b interfaces installed on
four different vehicles and running the optimized link state
routing protocol to route communication messages between
those vehicles. In their work, the authors present the assessment
of different performance metrics (i.e., packet loss and delay)
obtained while conducting VANET under evaluation through
different scenarios (e.g., urban, highway). A similar approach
can be found in [27], where the authors assessed their vehicular
testbed in static and mobile one- and three-hop scenarios, and
in [37], where authors add to these basic performance metrics
an evaluation of the packet drop rate as a function of the
distance between two moving vehicles. Clearly, the results of
all these works are jeopardized by the fact that only a very
limited amount of hops (i.e., at most three) can be traveled by
any message throughout their testbeds. In our work, we broke
such limit, exploiting a novel methodology that allowed us to
reproduce the conditions that such systems can face in reality
[17], [18]. In particular, we drove a platoon of four vehicles
through city roads and highways and required the system that
was running on the leading vehicle to send a message periodi-
cally. When the message reached the last vehicle of the platoon
directly or through the intervention of one or both of the avail-
able relays, the message bounced back, i.e., was retransmitted
in the opposite direction. The same mechanism was adopted
every time the leading (i.e., the front) vehicle received a copy
of the message it originally sent. In brief, bouncing back the
message at every edge of the network, we were able to infinitely
extend the platoon of vehicles for experimentation purposes
(i.e., a message traveled back and forth indefinitely, as long as
it was not lost). Utilizing the described mechanism, we were
able to reach the most relevant result of this paper, which is
the observation of alert messages that travelled very far away
(e.g., kilometers) from the original place where a hypothetical
accident occurred.
The seminal idea that concerns our work in the area of
VANET-based accident warning systems has been presented in

[38], wherein an algorithm is proposed that limits the number
of forwarders by operating the selection of a unique relay at
each hop. In particular, the rationale is that choosing the farthest
relay among all the potential ones also minimizes the number
of hops required to reach all approaching vehicles. In fact, in
case an accident warning system is implemented following such
philosophy, two main variables influence the delay experienced
by an alert message: 1) the number of hops (i.e., relays) required
by a message to reach all vehicles and 2) the number of for-
warders that intervene every time an alert message is received
by a group of vehicles. A similar idea is implemented in other
protocols that have been proposed successively [39]–[41]. One
limitation of all these approaches derives from the unrealistic
assumption that transmission ranges are constant and unique
in time. A recent variant of the mentioned schemes overcomes
such limitation as it accounts for the variable nature of the wire-
less channel (e.g., devising an automatic transmission range
estimator), without dealing with it in an optimal way [42]. Not
even this last scheme, however, provides a methodology that
efficiently answers to the asymmetries that practically appear
with wireless links (e.g., vehicle A can hear vehicle B but not
vice versa).
In summary, to the best of our knowledge, 1) no extensive
evaluation work has been performed on a real VANET-based
safety system for the simple reason that putting together the
vehicular resources that would be needed for such experimenta-
tion is costly, and 2) no accident warning system, with the sole
exception of [15], can optimally deal with the challenges that
wireless links pose to accident warning system schemes. In the
following, we describe more in detail the motivations that led
us to select [15] to manage the propagation of alert messages
throughout our testbed.
III. ON AN OPTIMAL ACCIDENT ALERT
SYSTEM: A SUMMARY
Rather than here providing a detailed description, which the
interested reader can find published in [15], we supply, in the
following, motivated evidence that explains why, among all
the possible candidates anticipated in Section II, we selected
just this mechanism to implement the accident warning system
subjected to our experiments.
In fact, the algorithm, described in [15], widely demon-
strated, both in simulation and in theory, that it can optimally
deal with time-variant wireless channels and asymmetric trans-
mission as it can disseminate an alert message throughout a
platoon of vehicles requiring the minimum possible number of
transmissions also when such situations arise. For this funda-
mental reason, we implemented [15] for testing as it exhibited
the best characteristics in terms of expected performance in
practical scenarios.
Now, the primary difference between the scheme published
in [15] and all the other algorithms presented before resides
in the fact that this algorithm can guarantee, whenever a route
between a source vehicle and a destination vehicle is available,
the delivery of an alert message along the minimum hop path,
also in realistic scenarios. Such characteristic is not guaranteed
by all the schemes that are described in Section II, as they
assume transmission ranges to be all equal and constant at all
times, and they disregard the fact that communication links can
be anything but symmetric. In practice, as transmission ranges
can strongly vary with surrounding conditions (e.g., obstacles),
the transmission range of the farthest relay may become very
short. In such situation, there might exist another vehicle that is
closer, for example, but that can relay an alert message much
farther away. Hence, in general, at each hop, a vehicle should
choose as its next relay not the farthest one but the farthest
spanning relay, i.e., the one that can relay a message the farthest
away in a given direction.
Differently from all those approaches where a vehicle
chooses as the next relay the one that happens to be the
farthest, the scheme proposed in [15] 1) constantly estimates
the transmission ranges and the positions of all neighboring
vehicles and 2) identifies as the next relay the farthest spanning
relay, i.e., the relay that can retransmit that given alert message
farthest away. In [15], proof is provided that demonstrates that
choosing the farthest spanning relay at each hop provides a
global optimal solution to the problem of finding the minimum
set of relays required to broadcast a message throughout a
platoon of vehicles.
The problem, at this point, becomes that of finding how
each vehicle can identify its farthest spanning relay among the
vehicles that follow it. This is performed by implementing a dis-
tributed software component termed Oracle service, which runs
within the accident warning system of each vehicle. The role
of the Oracle service is twofold: 1) It periodically exchanges
control messages with the Oracle services that run on neighbor-
ing vehicles to discover their positions and transmission ranges
and to resolve any asymmetric communications (i.e., discover
any potential relay that cannot be heard), and 2) it provides to
the accident warning system a list of relays ordered by how far
they can retransmit a message. In such list, termed Out list, the
topmost entry is the farthest spanning relay of that vehicle.
Finally, in [15], a mechanism is described that can be put to
good use by a relay to designate the next relay. This is simply
done by inserting the Out list provided by the Oracle service in
the transmitted alert message. On receiving an alert message,
a given vehicle V 1) checks whether its ID is in the list, and
2) if it is in the list creates a new alert message containing its
own Out list and waits for a time proportional to its position
in the list before transmitting. If V , before sending the relay
message, receives another copy of the same message, it aborts
the transmission, as evidently, another vehicle with a lower
position on the list already relayed that message.
IV. ARCHITECTING THE GREATEST EXPERIMENT EVER
The greatest VANET-based accident warning system exper-
iments that have ever been conducted, to this date, took place
in August 2011 in Los Angeles, as a collaboration between two
research groups: one based at the University of Bologna, Italy,
and one at the UCLA.
In the remainder of this section, we provide information
regarding the magnitude of the experiments that have been per-
formed, details on the testbed methodology that we exploited,

Fig. 1. Exemplar urban scenario.
and, finally, considerations regarding the hardware/software
architecture of the equipment installed on each vehicle.
We begin by mentioning that the tests were performed on
the UCLA campus and throughout the Los Angeles city area
and included the traversal of all types of streets (i.e., almost
a hundred, including illustrious streets such as Wilshire Blvd.,
Sunset Blvd., National Blvd., S. Monica Blvd., Sepulveda
Blvd., Melrose Ave., Mulholland Dr. and I-405), experienc-
ing heterogeneous traffic conditions, while recording logs and
assessing the system. These road experiments lasted, in total,
a week and comprise over 1000 km of traversed roads and
over 10 000 fictitious dangerous situations advertised by the
system at work. One single experiment was run every 10 s while
the platoon of vehicles moved, as this value was regarded as
sufficient to exceed the maximum lifetime of an alert message
in an extreme communication scenario such as a vehicular one.
In the following, we will concentrate on a large subset of such
data, slightly more than 2000 experiments, which comprise all
of the most relevant situations encountered while the system
was at work, while the platoon was moving on the road.
In fact, we ran our experiments in a number of different
road scenarios, which include urban and highway scenarios,
with different types of vehicles, which consisted of both sedans
and vans (a few pictures of our vehicles can be found in [8]).
Although the accident warning system at test has been con-
ceived for highway situations, it demonstrated on the road
that it could easily adapt to urban ones, without requiring any
modification. In particular, the experiments presented in this
paper embraced three different scenarios: 1) densely urban;
2) suburban; and 3) rural. As an example, a densely urban
scenario started at the UCLA campus and ended near the LAX
airport, as depicted in Fig. 1. An exemplar suburban scenario
was, instead, run around the UCLA campus, as it included Sun-
set Blvd., Hilgard Ave., and a portion of the I-405 highway that
divides the UCLA campus from the Getty Museum area (see
Fig. 2. Exemplar suburban scenario.
Fig. 3. Exemplar rural scenario.
Fig. 2). Finally, a sample rural experiment was located along
the Antelope Valley Freeway, as shown in Fig. 3. In general, all
test sites were chosen to assess the performance of our accident
warning system as surroundings varied in terms of 1) traffic;
2) speeds; and 3) wireless channel conditions.
As to the methodology that was used to overcome the limita-
tions imposed by the scarce availability of vehicular resources,
it is worthwhile noticing that this amounts to a major problem
that typically appears when experimenting with VANET-based
accident alert systems. We already documented how this can be
overcome using a so-called back and forth strategy described in
[17] and extended in [18]. Although we do not want to repeat
all the details provided in [17] and [18], this crucial issue needs
to be addressed as follows. For example, if we send a message
from the front vehicle to the back vehicle along a four-car-long

Fig. 4. Back and forth mechanism.
platoon, we could obtain at most three hops, yielding a dis-
tance of a few hundred meters. This clearly is unfit to test
the performance of an alert system that has the potential to
spread a warning kilometers away. Indeed, to overcome such
a problem, we implemented a new experimentation procedure,
as sketched in Fig. 4. In practice, periodically, the front vehicle
sends an alert that keeps bouncing back every time any of the
two edges (i.e., the back and front vehicles) of the platoon
is reached. In other words, the front and back vehicles keep
relaying the alert message they receive from a given direction
along the opposite one. This is the reason why, when utilizing
such methodology, the front and back vehicles are the only
ones that cannot change relative positions within the platoon
(i.e., the middle ones, instead, can overtake each other as often
and as many times as they like). Although when utilizing such
experimentation strategy only sparse network conditions are
met, as, in our case, only four cars were part of the platoon at
any time covering a stretch that ranged between approximately
30 (urban) and 2500 (highway) m, with this technique, an
alert message can virtually travel for an indefinite number of
hops. For example, Fig. 4 shows an alert message that traveled
for 16 hops and was not received after the 17th transmission.
Specifically, in Fig. 4, the value of T represents the number of
hops traveled by the alert message so far, with the exception
of the last hop (number seventeen), where the transmission
finally failed. As the vehicles proceed along the road at speed,
and an alert message keeps traveling between the edges of the
platoon, potentially several times, this guarantees that different
channel and environmental conditions are met, just as in a real
deployment scenario. Clearly, in a real-world scenario, in most
cases, a relay would have a wide selection of next-hop vehicles,
whereas with our experimental system only two choices of
next-hop vehicles were available at each retransmission step,
yielding, hence, suboptimal message relay decisions and con-
sequently suboptimal routes. Under this perspective, the results
reported in this paper represent a conservative snapshot of the
actual system working in reality. For future experimentation
campaigns, such suboptimality, however, can be addressed by
Fig. 5. Accident warning system hardware and software components.
increasing the number of vehicles involved in an experiment
while still utilizing the back and forth strategy. Finally, for what
concerns, instead, the channel and environmental variability
that were experienced during our tests, the results presented in
Section V-F will further witness that the adopted procedure is
sound and that it can be consistently used to test the feasibility
and performance of a highway accident warning system under
realistic conditions.
We now sketch the hardware and software architecture of
our implemented system, as shown in Fig. 5. Specifically, every
car was equipped with one Dell laptop running Ubuntu Linux
distribution version 11.04. Each laptop was connected to a Sirf
Star III GPS receiver and to two high-gain antennas (i.e., one
5-dBi omnidirectional antenna by Hyperlink technologies and
another 8-dBi omnidirectional antenna by l-com mounted at the
center of the roof of each car [43]) via a Ubiquity Networks
SRC PCMCIA/Cardbus 802.11g wireless card adapter exhibit-
ing a 300-mW transmission power and a −93-dBm receiver
sensitivity [44]. Our software system relied on two fundamental
components: 1) Jpcap [45], a Java library that supports low-
level networking operations, and 2) gpsd [46], a GPS monitor
daemon. Jpcap, in particular, supports the use of raw sockets,
which permitted us to bypass the overhead introduced by the
IP layer. Hence, all messages, both Oracle messages and alert
messages, were sent as broadcast layer-2 frames and were
received and processed by all the other vehicles in the pla-
toon. This means that vehicles sent messages at a fixed base
rate as Auto Rate Fallback is disabled for broadcast frame
transmissions, which was equal to 6 Mb/s under our particular
configuration. The gpsd utility, instead, was utilized to read
and translate the serial GPS stream in the JSON lightweight
data interchange format, making it available through a local
socket. For a more clear understanding, we summarized all of
the described information in Table I.
Now, to conclude this section, it is worthwhile to spend a few
words on how the results that have been observed utilizing the
described setting can be put in relation with those that would
have been achieved when adopting an accident warning system
based on VANET 802.11p technologies. Any performance gap,
in favor of any of the two telecommunication systems, heavily
depends on the transmission ranges reached when opting for

TABLE I
EXPERIMENT INFORMATION
any of the two solutions. To provide such value for 802.11p,
we resorted to scientific literature, as well as to the standard.
Interestingly, this is an ambiguous question whose answer is
not clear, yet, to this date. In fact, when resorting to literature,
we find that different studies experienced different transmission
ranges, from as low as 70 m to almost 500 m [10]–[12]. Such
uncertainty is confirmed by standardization entities, which have
not yet finally determined how dedicated short range communi-
cation channels will be utilized [13]. Now, an indication of how
far safety messages could be heard is given by governmental
institutions, which report that they could reach as far as 1 km
[14]. This said, how our results can be interpreted clearly
depends on how 802.11p will be employed in its final version
for vehicular safety applications. How far our messages traveled
is clear in Fig. 11 (the ratio of lost messages exceeded the 20%
percentage above 300 m). For what concerns a comparison with
802.11p, instead, if safety messages will be able to travel as far
as 1 km, as the standardization and governmental documents
that we cite report, we can conclude that with this study we took
a conservative path, putting ourselves in a worst-case scenario.
V. EXPERIMENTAL OUTCOMES
This section is organized as follows. We will first provide
the assessment of the most important performance figures that
characterize an accident warning system and then discuss in
detail and confute, supplying further experimental evidence,
the possible objections that could be raised regarding those
results. Hence, step by step, we will provide data supporting
the feasibility and practicality of the accident warning system
at test. As a final result, we will show that the accident warning
system presented in [15] meets the expectations and can reduce
by approximately 40% the number of vehicles involved in a
pileup, compared with situations where it is not adopted.
A. How Far Could an Alert Be Spread?
In Fig. 6, we summarize the distances that were reached
by alert messages while the vehicular testbed was operating.
In a relevant number of cases, as shown in the figure, alert
messages reached distances that well exceeded those necessary
to correctly operate (e.g., a few kilometers).
Not only the sole distance but also the number of times
alert messages were successfully relayed represent a relevant
Fig. 6. Overall distribution of the distances reached by alert messages (one of
them travelled as far as 36 km).
Fig. 7. Number of hops versus travelled distance when traveling at a speed
above 54 km/h.
performance metric of a highway accident warning system. For
this reason, we plotted both figures of merit in Fig. 7, whereas
the platoon was moving at at least 54 km/h (at speed, as in a real
case). The reached distance is computed as the total distance
covered while a message travels back and forth, between the
front and back vehicles. Hence, for example, if an alert message
traveled 200 m, experiencing first a single transmission from
the front to the back vehicle, and then another one from the

TABLE II
AVERAGE NUMBER OF HOPS PER REACHED DISTANCE WHILE VEHICLES MOVED AT MORE THAN 54 km/h
TABLE III
REACHED DISTANCE
back vehicle to the front one, before being lost, this experiment
would be represented in Fig. 7 with a point plotted in (2, 400).
One of the situations where we observe the greatest reached
distance (24.3 km) was obtained when forwarded 81 times,
traveling an average of 300 m with each relay procedure.
However, even farther distances than those shown in Fig. 7 were
reached, as one message traveled as far as 36 km (for the sake
of clarity, we set an upper limit on the x-axis to 25 km).
The points plotted in Fig. 7 exhibit an average traveled
distance per hop of about 215 m, matching in general the
approximate expected transmission range for 802.11g links.
Also of interest is the regression line plotted in Fig. 7, which
gives an approximate relation between the number of times
an alert message was relayed and the distance, expressed in
kilometers, that it reached. In Table II, we provide the average
number of hops that applies to a given distance of interest,
whereas our experiments were obtained from the regression line
of Fig. 7.
B. After How Many Relays Did Alerts Vanish?
A second important question entails deriving for how long,
on average, an alert message traveled before vanishing at
some point. Clearly, an alert message can disappear, not being
received by any of the designated relays, for a number of
different reasons, which are generally related with the traversal
of degraded wireless channels.
In Fig. 7, we provided the number of times a message was
relayed when at least one hop was traveled (i.e., the message
was successfully received by at least one vehicle). When at
least one vehicle was reached, we found that the distance
that was traveled by an alert message, on average, was high.
In Table III, we report that for average platoon speeds that
exceeded 54 km/h, the average reached distance was 3897 m
(approximately 19 hops according to Table II), and this value
did not significantly decrease when moving at higher speeds,
being equal to 3625 m for speeds that exceeded 90 km/h.
When, instead, alert messages that immediately vanished
(i.e., were not received by any relay) are also accounted for, we
observe that the number of hops traveled by an alert message
drops from an average slightly above 16, when the platoon
moved at speeds confined between 54 and 90 km/h, to an
average of about six hops, when the platoon exceeded a speed
of 90 km/h (see Fig. 8). The reason for such drop, as will also be
confirmed by the results discussed in the following sections, is
due to the high amount of alert message that incurred in highly
degraded wireless channels, especially when the sources and
destination distances fell beyond a given value.
Fig. 8. Average number of times an alert message is relayed as a function of
the average platoon speed.
Fig. 9. Delay versus traveled distance when traveling at a speed above
54 km/h.
C. How Long Did it Take to Spread an Alert at a
Given Distance?
To correctly analyze the delay incurred by an alert message
when reaching a given distance, we first measured the overhead
introduced by our software implementation. To do this, we
measured during all of our on-the-road experiments at the
front vehicle the time that elapsed between the reception of
an alert message and the time at which that same message
was successively relayed, finding that, during our on-the-road
experiments, while traveling at a speed that exceeded 54 km/h,
such a value amounted to 72 ms on average. This delay value,
measured utilizing the laptop OS clock, was clearly dominated
by our particular software implementation.
Now, keeping such information in mind, we are interested
in knowing how long it took to reach a given distance, when
traveling at a speed that exceeded the value of 54 km/h. Such
information is shown in Fig. 9, where we provide a scatterplot
where each single point denotes the reached distance and the
time taken to reach such a distance.

TABLE IV
AVERAGE DELAY PER REACHED DISTANCE WHILE VEHICLES MOVED AT MORE THAN 54 km/h
TABLE V
AVERAGE NUMBER OF HOPS PER REACHED DISTANCE OVERALL EXPERIMENTS
TABLE VI
AVERAGE DELAY PER REACHED DISTANCE OVERALL EXPERIMENTS
We again obtained and plotted in Fig. 9 a linear regression
formula, this time computing the relation that best fits the
samples that put in correspondence the traveled distance and the
delay experienced by the alert messages. We then utilized such
formula to compute the average delays per reached distance
reported in Table IV.
The average delay experienced, per hop, can now be ob-
tained, reminding that, according to Table III, on average, alert
messages reached a distance of 3897 m (i.e., 3.9 km) when the
platoon moved at speeds that exceeded 54 km/h. The average
delay, hence, experienced by an alert message amounted to
1806 ms, also according to Table IV. The corresponding aver-
age number of times an alert message was relayed equals 19.3,
in agreement with the values reported in Table II. This means
that the average delay experienced per each relay procedure
is equal to 92 ms. Recalling that the measured processing
time (which included only the time spent within our software
implementation at the front vehicle) amounted to 72 ms, we find
a difference of 20 ms between these two values that cannot be
justified by a cumulative contribution of Java Native Interface
calls of Jpcap, OS scheduling procedures, and message trans-
mission times. Leaving the solution of this puzzle to the next
two sections, we here note that 72 ms would not be wasted in
a final hardware implementation. Assuming, hence, an efficient
hardware implementation processing time negligible, in such a
situation, the average time required to receive an alert message
with a single transmission would drop to 20 ms.
D. Is There a Correlation Between Vehicle Speed and
Alert Relay Time?
Since the average time required for an alert message to be
relayed exceeded our expectations (i.e., in simulation this value
lied below 10 ms), we performed an additional analysis over all
the data set that was collected without limiting to the samples
that were recorded while the average speed of the platoon lied
above 54 km/h.
The first difference that we encountered was that the average
processing time measured at the front vehicle lowered to 68 ms;
hence, we found a value that is 4 ms lower than the value com-
puted before. We then reorganized the data and computed the
regression formulas that best fit it according to the methodology
used for the samples plotted in Figs. 7 and 9. In fact, the linear
regression formulas also changed slightly and, consequently,
the average number of hops and the average delay per reached
distance, as reported in Tables V and VI.
Visual inspection of the average values reported in Tables V
and VI with those reported in Tables II and IV reveals that
the former numbers increased. Nonetheless, when computing
now the average time spent during each hop, we find that
approximately 74 ms was required during each relay procedure.
Finally, recalling that the measured processing time amounted
to 68 ms, we now found that the time wasted in processing
procedures at the front vehicle now amounted to the 92% of the
total time and that 6 ms, on average, was realistically required
for each relay procedure, hence, meeting our expectations.
E. Why Does Average Relay Time Increase While
Traveling at Speed?
An interesting point now lies in understanding why, when
traveling at an average speed that exceeded 54 km/h, the platoon
experienced a higher alert message average relay delay. We
found that the reason for such delay increase is due to the
high amount of alert messages that vanished when the sources
and destination distances fell beyond a given value. Here, we
provide clues that corroborate such a fact.
The first is that when traveling at high speeds (i.e., on
highways), the distance between vehicles was generally high.
This fact clearly emerges from Fig. 10, which represents the
speed of the platoon (gray line expressed in meters per second),
and the relative distances between the front vehicle and the
three other vehicles, as these evolved in time in the situation
depicted in Fig. 3. During the time that the vehicles moved on
the I-405 Freeway, the platoon stretch grew from 100 to 1200 m,
with a consequent increase of the distance travelled by single
transmissions.
Fig. 11, instead, provides the second fact, exhibiting the
impact of source–destination distance on alert message losses:
beyond 300 m, the loss probability reached relevant values, with
a peak of almost 80% in the 500- to 600-m range. Such situation
clearly jeopardizes the performance of any accident warning
system operating in very sparse networks. In a case where four
vehicles are lined, each at a distance of 300–400 m from the
vehicle that is immediately in front, an alert message would
be relayed six times with a probability of (0.6)6
= 0.046, and,
thus, almost never.
Now, the increased delay value per each relay procedure is
easily explained considering that losses were frequent and that

Fig. 10. Distances from the front vehicle (red, yellow, and blue lines) and
average platoon speed (gray line) during the rural scenario experiment.
Fig. 11. Alert message loss rate as a function of the distance between the
source and destination vehicles.
the system at test implements a redundancy mechanism where,
if the transmission of the best relay is not heard by the second
best relay, the second best relay transmits the message, which
is assumed to be vanished, after waiting for 50 ms. The original
mechanism includes up to 10 different relays to improve the
robustness of the accident warning system. In our case, at most
two other relays could intervene, with a total increase of at most
100 ms due to retransmissions.
This said, it is now easy to solve the puzzle left unanswered
in Section V-C. The value of 72 ms, in fact, represents the
time elapsed between the arrival and the departure of an alert
message, as seen inside our accident warning system running
on the front vehicle. Reminding here how the back-and-forth
mechanism works, the front vehicle is the best relay at all
times; thus, every time a message is received by the front
vehicle, no retransmission timer is started. This means that
Java Native Interface calls of Jpcap, OS scheduling procedures,
and message transmission time contribute to those 20 ms as
expected, but alert message retransmissions do as well. Taking
now into account the delays reported in Sections V-C and D,
we observe that the 20 ms is heavily dominated by alert
message retransmissions, which we here estimate amounting to
at least 20 − 6 = 14 ms, where 6 ms was the time computed for
each relay procedure in the previous section.
Finally, the ground truth cannot be found without performing
direct measurements of the delays that composed the total
one that was experienced by an alert message during relaying
operations. This operation has been performed recording the
timestamps during the experimental campaign at all vehicles. In
particular, we measured the contributions of the Java Native In-
terface calls of Jpcap, OS scheduling procedures, and message
transmission times recording the timestamp right before a JNI
call was operated and recording additional timestamps at all of
the intermediate steps required before an alert message transited
through the wireless interface. In addition, as a crosscheck, we
captured the time instants at which that given alert message
was first heard by other vehicles. Table VII reports the values
that were measured from the experimental testbed for those
messages that reached distances between 3 and 5 km (to be
comparable with the values that we have so far indicated) while
traveling at speed. These numbers roughly confirm those that
have been obtained so far and reveal one additional interesting
fact (considering that retransmission timeout times were set
to 50 ms during the experimental campaign): When traveling
at speed, one retransmission was required every three relay
operations, on average.
Clearly, lower retransmission times potentially yield lower
average delays, but on the other side, lower retransmission
time values would have triggered unwanted retransmissions
(and collisions), as the best relay could have been anticipated
while still in the process of sending the received alert message.
The value of 50 ms, hence, was set as the right compromise
between these two needs. However, lower retransmission values
can be experimented with more efficient implementations of our
accident warning system.
F. Have Our Experiments Been Realistic, At Last?
The effective realism of the entire experimentation campaign
lies upon the assumption that the hops that are traveled by
an alert message between the vehicles of our testbed are rep-
resentative of those that would be encountered in a practical
situation. While it directly follows from the construction of
the experiment that the wireless channel conditions that are
experienced between distinct experiments radically differ (i.e.,
each experiment is characterized by a brand new accident
warning message, which is sent every 10 s), as the platoon
moved at speed, the same is not as obvious within a single
experiment. We will address such issue here.
In a practical situation, we generally expect that the wire-
less channel encountered by an alert message can radically
differ from transmission to transmission. This is intuitively
true, as between different hops the sender and the receiver
may be 1) traveling at radically different speeds and distances,
2) equipped with devices that exhibit diverse power and sen-
sitivity characteristics, 3) experiencing increasing or no inter-
ference at all, 4) surrounded by reflecting objects (i.e., other
cars, buildings, etc.), and 5) traveling in different or in the

TABLE VII
ALERT MESSAGE AVERAGE DELAY COMPONENTS PER VEHICLE WHEN TRAVELING AT SPEED
same direction, also with respect to the propagation of an alert
message. All these can be summarized by the fact that, in
practice, we cannot expect the wireless channel encountered at
one hop to be correlated with the wireless channel traversed at
another, because vehicles are moving at speed [47]. Our scope,
hence, is that of understanding to what extent the alert messages
that traveled back and forth, from one edge to the other of the
testbed platoon, found similar, or different, channel conditions
when traversing a hop that was previously encountered. A reli-
able indicator of such similarity is given by the coherence time
of the wireless channel between the vehicles of the platoon,
i.e., the time interval over which the impulse response of those
channels is considered correlated to their previous values [48].
Before, however, proceeding and resorting to the theory of
wireless propagation, we would like to appeal to the principle
of reality that lies behind any experimental work: our tests have
been carried out on real roads with vehicles that moved at speed
and with highly varying boundary conditions. This means that
no two transmissions of alert messages encountered the same
line-of-sight (LOS) and Non-LOS (NLOS) conditions. Hence,
even when a source and a destination traveled exactly at the
same speed (event that was rather unlikely), the relative posi-
tions of the surrounding objects rapidly changed, also rapidly
changing propagation conditions.
Now, the average value of the channel coherence time can
be found as the inverse of the Doppler shift, i.e., the change in
frequency that an electromagnetic wave experiences as a result
of a difference in speed between a given source and a given
receiver. This value is simply found as vs,r × f0/c, where vs,r
is the relative speed between the source and the receiver, f0
is the carrier frequency of the transmitted signal, and c is the
speed of light. At a differential speed of 1 m/s (i.e., 3.6 km/h)
between a source and a receiver, the channel coherence time
is approximately 125 ms for a carrier frequency of 2.4 GHz
(i.e., the carrier frequency of employed 802.11g interfaces). At
differential speeds of 2, 4, and 8 m/s, such value decreases to
62, 31, and 15 ms, respectively [49]. All of such values are
confirmed in both theory and by previous experimental works,
which also observe how channel coherence times are two to
three times shorter in highway situations than in urban ones due
to rapidly changing boundary conditions [50]–[53].
To further support these arguments, an additional set of
experiments was performed with the scope of empirically es-
timating the longest channel coherence time that was expe-
rienced, thus in a worst case scenario for our experimental
methodology. In particular, two vehicles in LOS were parked,
in a static position, along the curb of Wilshire Blvd. One
of the vehicles, the transmitting vehicle, steadily sent User
Datagram Protocol packets matching the link capacity to the
second vehicle, the receiver. The receiver could slowly perform
small forward or backward movements (well below the speed
of 1 m/s), just as when proceeding in line, never exceeding,
however, a maximum distance of 5 m from its original position.
In the meantime, intense traffic flowed through Wilshire, thus
Fig. 12. RSSI values over a 7.5-min interval.
Fig. 13. Autocorrelation function of the RSSI plot shown in Fig. 12.
creating varying surrounding conditions. Ten of such experi-
ments were performed, each lasting 15 minutes. During each
experiment, the link received signal strength indication (RSSI)
value was recorded every 15 ms at the receiver. In fact, the
RSSI time series can be put to good use to empirically estimate
the channel coherence time, as follows [54], [55]: the time
instant where the autocorrelation function of the RSSI time
series falls to 90% of its initial value. Fig. 12 shows the
RSSI readings in the 7.5-min interval that exhibited the highest
average autocorrelation. Fig. 13, instead, shows the autocor-
relation function that refers to the same time interval. The
aforementioned definition applied to the RSSI readings plotted
in Fig. 12 returns a channel coherence time of about 180 ms (the
horizontal line in correspondence of an autocorrelation value of

Fig. 14. Alert message loss rate as a function of the differential speed between
the source and destination vehicles.
0.9 crosses the edge of 12th time lag in Fig. 13), agreeing with
other works that performed similar analysis before in similar
conditions [50]–[53].
Now, putting such information into context, we observe that
the same source–destination pairs are typically repeated every
time an alert message reaches the front vehicle after bouncing
off the back vehicle. In fact, when, for example, an alert
message sent by the front vehicle flows through a third vehicle,
namely, V , to reach the back vehicle, a source–destination
repetition will appear along the path travelled by that alert
message when the front vehicle relays that same alert message
again to V . Hence, the distribution of back and forth times (i.e.,
the round trip experienced by an alert message at the front
vehicle) can indicate whether an alert message traversed the
same source–destination pair experiencing the same channel
conditions. Our results show that the average time elapsed
between two subsequent transmissions of an alert message from
the front vehicle amounts to 186 ms (the fifth percentile is equal
to 104 ms and the overall minimum value to 52 ms). These
values indicate that the chance that the same wireless channel
conditions are met by a message during two subsequent loops
is very low. To further support such argument, we analyzed the
number of one-hop transmission events that occurred during
all of our experiments as a function of the speed differential
between a source and a destination pair, whereas the platoon
moved at an average speed that exceeded 15 m/s (see Fig. 14).
Observing the figure, the relative majority of transmissions
occurred between a sender and a receiver that moved at a
relative speed below 1 m/s. However, more interesting is the
fact that the fraction of messages that got lost under given speed
drift conditions kept quite constant up to a source–destination
speed difference of 4 m/s (i.e., 26% between 0 and 1 m/s,
30% between 1 and 2 m/s, 34% between 2 and 3 m/s, and
33% between 3 and 4 m/s). Hence, for all those cases where a
sufficient amount of experimental data is available, we found
that the chance for an alert message of getting lost when
the relative speed between the source and the destination was
below 1 m/s (i.e., in a situation where the channel coherence
is expected to be longer) was very similar to the chance of
getting lost when the relative speed was instead higher. All this
indicates, once more, that the experiment that has been carried
out clearly resembles a practical case, as the channel conditions
changed with the same modalities, just as in reality.
Fig. 15. Traffic scenarios: 1 = free flow traffic conditions, 2 = congested
traffic conditions.
For the sake of completeness, we observe that, with our
methodology, no transmission of alert messages between two
vehicles that were traveling in opposite directions occurred.
The worst-case scenario, from a wireless propagation point
of view, is here represented by all those situations where two
vehicles travel in the opposite directions at relative speeds
that can be as high as 220 km/h (e.g., 110 km/h in the two
respective directions). Although we believe that such type of
event would not be as frequent as the event of a message
being sent between two vehicles moving in the same direction,
as only a few seconds, in fact, would be available for the
discovery of a candidate relay and the successful transmission
of an alert message to a vehicle that is traveling at a relative
speed of 220 km/h, this case might as well occur. Interestingly,
relevant previous work, which utilizes hardware settings that
are very similar to ours, reveals the amount of data that can be
exchanged between two vehicles that are moving in opposite
directions on a highway [56]. The work that has been presented
in the cited paper only focuses on the exchange of bulk data
between two moving vehicles but can still be of interest to
understand what contact time and what transfers of information
can be achieved in such situation. In particular, in [56], average
contact times that were slightly less than 15 s and average
amounts of transferred data that were approximately equal to
8 MB were observed when utilizing packet sizes that lied be-
tween 1000 and 2304 B. A rough analysis of such numbers re-
veals that the transmission range conditions experienced in [56]
are compatible with those observed during our experiments;
in fact, vehicles traveling in opposite directions at a speed of
110 km/h would begin their transfer of information at a distance
of (110/3.6) ∗ 7.5 ≈ 230 m from each other. Concluding, al-
though we were not able to test our system in this last particular
situation, the numbers that are available from relevant literature
reveal that our accident warning system could be capable of
also facing such adverse conditions, as alert messages weigh at
most a few hundred of bytes, and the use of 802.11p, which
is capable of achieving higher transmission ranges, and hence,
contact times could not but improve its performance.
G. In Summary: How Many Vehicles Can Be Saved?
We are now able to conjecture how many vehicles could be
saved by utilizing a system like the one that is described here. In

fact, we have found that, on average, the delay incurred by our
system to deliver an alert message one hop away was confined
between the values of 6 and 20 ms, depending on the particular
wireless propagation scenario.
We simulated a vehicle crash on a three-lane highway and
estimated how many vehicles, on average, could be involved
with and without an intervehicular and also considering an ideal
accident warning system, where all vehicles learn about the
occurrence of an accident with no delay at all.
We considered two realistic traffic flow scenarios, a realistic
pavement condition that corresponded to the dry conditions that
we experienced in Los Angeles, realistic driver response times,
and vehicle lengths as well. Vehicle speeds and related time-
headway distributions are also realistic as drawn from a real
measurement campaign [57]. Specifically, in the congestion-
free scenario, vehicles move at 110 km/h and the time-headway
distribution induces a linear density of 20 vehicles/km on a per
lane basis. In the congested scenario, the speed is 40 km/h,
and the linear density is 40 vehicles/km. The kinetic friction
constant between the tires and the asphalt is set to 0.8 as the
pavement is dry, as suggested in [58]. The drivers’ response
times are in the [0.75, 1.4] s range, and the vehicles’ lengths are
randomly drawn from the [3.5, 5] m interval according to [59]
and [60]. The vehicle following model, used when no accident
warning system is operating, entails that a driver brakes after
a random response time since seeing the preceding vehicle
braking.
We ran 100 simulations, whose results are as follows. Un-
der free-flow conditions and without the use of any accident
warning system, we find that the average number of vehicles
involved in an accident equals 8.85. When, instead, our accident
warning system is adopted, we find an average value equal to
5.76 when retransmitting an alert message with an average de-
lay of 20 ms at a distance of 215 m. In the congested flow case,
both values are lower and equal to 5.40 (no accident warning
system) and 3.13 (20-ms delay). We plot such information in
Fig. 15, normalizing to 100 the highest values pertaining to
each respective experiment (i.e., 8.85 in the free-flow case and
5.40 in the congested case) and scaling the other accordingly
within each experiment scenario. The leftmost group of bars,
which correspond to the experiment labeled as 1, results from
the free-flow scenario, whereas the rightmost group of bars,
labeled as 2, results from a congested scenario. Within each of
the two scenarios, the leftmost bar represents the base case of
100 vehicles involved in an accident when no accident warning
system is deployed, whereas the fourth bar instead refers to
those situations where alert messages experience a delay of
20 ms at each relay procedure. In summary, our simulation
results, based on real delay values drawn from our experiments,
definitely confirm that the use of an accident warning system
like ours can potentially reduce the number of crashed vehicles
by nearly 40%.
VI. CONCLUSION
In this paper, we have filled an important gap of VANET sys-
tems research, presenting a detailed description of the greatest
experiments, to date, ever performed with an accident warning
system specifically devised for highway scenarios. In particular,
we implemented and tested an accident warning system based
on the multihop broadcast algorithm that has been proven to be
optimal in terms of bandwidth usage and covered distance in
realistic scenarios. We demonstrated, obtaining a confirmation
drawn from experimental data of simulative and theoretical
results, that such systems can really improve the conditions
experienced on the road by drivers, as we collected further
evidence that close to 40% of vehicles could be saved from
being involved in a pileup when an accident occurs.
ACKNOWLEDGMENT
The authors wish to thank F. Bronzino for his help in carrying
out the experiments and F. Callegati, M. D. De Amorim,
M. Gerla, I. Seskar, and the anonymous reviewers that helped
improve the quality of this paper.
REFERENCES
[1] Accessed on Jan. 20, 2012. [Online]. Available: http://heatco.ier.
uni-stuttgart.de/
[2] Accessed on Jan. 20, 2012. [Online]. Available: http://www.europarl.
europa.eu
[3] M. Torrent-Moreno, D. Jiang, and H. Hartenstein, “Broadcast reception
rates and effects of priority access in 802.11-based vehicular ad-hoc
networks,” in Proc. 1st ACM Int. Workshop VANET, 2004, pp. 10–18.
[4] Accessed on Jan. 20, 2012. [Online]. Available: http://www.
itsworldcongress.org/
[5] Accessed on Jan. 20, 2012. [Online]. Available: http://blogs.wsj.com/
tech-europe/2011/07/13/car-to-car-communication-could-reduce-jams/
[6] Accessed on Jan. 20, 2012. [Online]. Available: http://www.bbc.co.uk/
news/technology-14125245
[7] Accessed on Jan. 20, 2012. [Online]. Available: http://news.discovery.
com/autos/accident-alert-app-taken-for-test-drive-110914.html
[8] Accessed on Jan. 20, 2012. [Online]. Available: http://gadget.wired.it/
news/applicazioni/2011/10/25/app-anti-incidente-ucla.html
[9] A. Amoroso, G. Marfia, M. Roccetti, and G. Pau, “To live and drive
in L.A.: Measurements from a real intervehicular accident alert test,” in
Proc. IEEE WCNCW, 2012, pp. 328–332.
[10] F. Martelli, M. E. Renda, G. Resta, and P. Santi, “A measurement-based
study of beaconing performance in IEEE 802.11p vehicular networks,” in
Proc. 31st IEEE INFOCOM, 2012, pp. 1503–1511.
[11] F. Bai, D. D. Stancil, and H. Krishnan, “Toward understanding char-
acteristics of Dedicated Short Range Communications (DSRC) from a
perspective of vehicular network engineers,” in Proc. 16th Annu. Int. Con.
MobiCom, 2010, pp. 329–340.
[12] K. Hong, D. Xing, V. Rai, and J. Kenney, “Characterization of DSRC
performance as a function of transmit power,” in Proc. 16th ACM Int.
Workshop VANET, New York, NY, 2009, pp. 63–68.
[13] Soc. Automotive Eng. Int., Warrendale Marshall, PA, DSRC Implementa-
tion Guide, 2012.
[14] M. McGurrin, “Vehicle information exchange needs for mobility applica-
tions,” U.S. Dept. Transp., Res. Innovative Technol. Admin., Washington,
DC, Final Rep. FHWA-JPO-12-021, Feb. 2012.
[15] A. Amoroso, G. Marfia, and M. Roccetti, “Going realistic and optimal: A
distributed multi-hop broadcast algorithm for vehicular safety,” Comput.
Netw., vol. 55, no. 10, pp. 2504–2519, Jul. 2011.
[16] A. Amoroso, G. Marfia, and M. Roccetti, “Optimal assessments in
VANET: The Oracle,” in Proc. IFIP WD, 2010, pp. 1–5.
[17] A. Amoroso, G. Marfia, M. Roccetti, and G. Pau, “Creative testbeds for
VANET research: A new methodology,” in Proc. IEEE CCNC, 2012,
pp. 477–481.
[18] G. Marfia, M. Roccetti, A. Amoroso, M. Gerla, G. Pau, and J. H. Lim,
“Cognitive cars: constructing a cognitive playground for VANET re-
search testbeds,” in Proc. 4th Int. Conf. Cogn. Radio Adv. Spectr. Manag.
(CogART), 2011, pp. 1–29.
[19] I. Leontiadis, G. Marfia, D. Mack, G. Pau, C. Mascolo, and M. Gerla,
“On the effectiveness of an opportunistic traffic management system for
vehicular networks,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 4,
pp. 1537–1548, Dec. 2011.

[20] S. R. Cavalcanti, M. E. M. Campista, F. B. Abdesslem, L. H. M. K. Costa,
M. D. de Amorim, and O. C. M. B. Duarte, “Veer: A trajectory-based peer
selection algorithm for networks of vehicles,” in Proc. 9th IFIP Annu.
Med-Hoc-Net Workshop, 2010, pp. 1–8.
[21] M. Gerla, B. Zhou, Y.-Z. Lee, F. Soldo, U. Lee, and G. Marfia, “Vehicular
grid communications: The role of the Internet infrastructure,�� in Proc. 2nd
Annu. Int. Workshop Wireless Internet, 2006, pp. 112–120.
[22] S. Savasta, M. Pini, and G. Marfia, “Performance assessment of a com-
mercial GPS receiver for networking applications,” in Proc. IEEE 5th
CCNC, 2008, pp. 613–617.
[23] L. Bononi and M. Di Felice, “A cross-layered MAC and clustering scheme
for efficient broadcast in VANETs,” in Proc. IEEE 2nd Int. Workshop
MOVENET, 2006, pp. 190–197.
[24] C. E. Palazzi, F. Pezzoni, and P. M. Ruiz, “Delay-bounded data gathering
in urban vehicular sensor networks,” Pervasive Mobile Comput., vol. 8,
no. 2, pp. 180–193, Apr. 2012.
[25] J. J. Blum and A. Eskandarian, “A reliable link-layer protocol for robust
and scalable intervehicle communications,” IEEE Trans. Intell. Transp.
Syst., vol. 8, no. 1, pp. 4–13, Mar. 2007.
[26] M. Moske, H. Fussler, H. Hartenstein, and W. Franz, “Performance mea-
surements of a vehicular ad hoc network,” in Proc. IEEE 59th VTC, 2004,
pp. 2116–2120.
[27] J. Santa, M. Tsukada, T. Ernst, and A. F. Gomez-Skarmeta, “Experimental
analysis of multi-hop routing in vehicular ad-hoc networks,” in Proc. 5th
Int. Conf. TridentCom Workshops, 2009, pp. 1–8.
[28] N. Wisitpongphan, O. K. Tonguz, J. S. Parikh, P. Mudalige, F. Bai,
and V. Sadekar, “Broadcast storm mitigation techniques in vehicular
ad hoc networks,” IEEE Wireless Commun., vol. 14, no. 6, pp. 84–94,
Dec. 2007.
[29] F. Ye, M. Adams, and S. Roy, “V2V wireless communication protocol
for rear-end collision avoidance on highways,” in Proc. IEEE Int. Conf.
Commun. Workshops, 2008, pp. 375–379.
[30] R. A. Santos, R. M. Edwards, and A. Edwards, “Cluster-based location
routing algorithm for vehicle to vehicle communication,” in Proc. IEEE
Radio Wireless Conf., 2004, pp. 39–42.
[31] G. Marfia, G. Pau, E. De Sena, E. Giordano, and M. Gerla, “Evaluating
vehicle network strategies for downtown Portland: Opportunistic infras-
tructure and the importance of realistic mobility models,” in Proc. 1st Int.
MobiSys Workshop MobiOpp Netw., 2007, pp. 47–51.
[32] S. Ferretti and M. Roccetti, “Fast delivery of game events with an opti-
mistic synchronization mechanism in massive multiplayer online games,”
in Proc. ACM SIGCHI Int. Conf. Adv. Comput. Entertainment Technol.,
2005, pp. 405–412.
[33] C. E. Palazzi, S. Ferretti, M. Roccetti, G. Pau, and M. Gerla, “How do
you quickly choreograph inter-vehicular communications? A fast vehicle-
to-vehicle multi-hop broadcast algorithm, explained,” in Proc. IEEE 4th
Annu. CCNC, 2007, pp. 960–964.
[34] F. Malandrino, C. Casetti, C.-F. Chiasserini, and M. Fiore, “Content
downloading in vehicular networks: What really matters,” in Proc. IEEE
INFOCOM, 2011, pp. 426–430.
[35] M. Cesana, L. Fratta, M. Gerla, E. Giordano, and G. Pau, “C-VeT
the UCLA campus vehicular testbed: Integration of VANET and Mesh
networks,” in Proc. EW Conf., 2010, pp. 689–695.
[36] G. Marfia and M. Roccetti, “Vehicular congestion detection and short-
term forecasting: A new model with results,” IEEE Trans. Veh. Technol.,
vol. 60, no. 7, pp. 2936–2948, Sep. 2011.
[37] H. Venkataraman, A. d’Ussel, T. Corre, C. H. Muntean, and
G.-M. Muntean, “Performance analysis of real-time multimedia trans-
mission in 802.11p based multihop hybrid vehicular networks,” in Proc.
6th Int. Wireless Commun. Mobile Comput. Conf., 2010, pp. 1151–1155.
[38] G. Korkmaz, E. Ekici, F. Ozguner, and U. Ozguner, “Urban multi-hop
broadcast protocol for inter-vehicle communication systems,” in Proc. 1st
ACM Int. Workshop Veh. Ad Hoc Netw., 2004, pp. 76–85.
[39] A. Zanella, G. Pierobon, and S. Merlin, “On the limiting performance
of broadcast algorithms over unidimensional ad-hoc radio networks,” in
Proc. Wireless Personal Multimedia Commun., 2004, pp. 165–169.
[40] E. Fasolo, R. Furiato, and A. Zanella, “Smart broadcast algorithm for
intervehicular communication,” in Proc. Wireless Personal Multimedia
Commun., 2005, pp. 1–5.
[41] S. Biswas, R. Tatchikou, and F. Dion, “Vehicle-to-vehicle wireless com-
munication protocols for enhancing highway traffic safety,” IEEE Com-
mun. Mag., vol. 44, no. 1, pp. 74–82, Jan. 2006.
[42] C. E. Palazzi, S. Ferretti, and M. Roccetti, “An inter-vehicular commu-
nication architecture for safety and entertainment,” IEEE Trans. Intell.
Transp. Syst., vol. 11, no. 1, pp. 90–99, Mar. 2009.
[43] S. Kaul, K. Ramachandran, P. Shankar, S. Oh, M. Gruteser, I. Seskar,
and T. Nadeem, “Effect of antenna placement and diversity on vehicular
network communications,” in Proc. IEEE Conf. Sensor, Mesh Ad Hoc
Commun. Netw., 2007, pp. 112–121.
[44] Super Range Cardbus, Ubiquiti Netw., Inc., San Jose, CA, Apr. 25, 2012.
[Online]. Available: http://www.metrix.net/datasheets/src_datasheet.pdf
[45] Accessed on Jan. 20, 2012. [Online]. Available: http://netresearch.ics.uci.
edu/kfujii/Jpcap/doc/
[46] Accessed on Jan. 20, 2012. [Online]. Available: http://gpsd.berlios.de/
[47] M. Jerbi, P. Marlier, and S. M. Senouci, “Experimental assessment of V2V
and I2V communications,” in Proc. IEEE Int. Conf. Mobile Adhoc Sens.
Syst., Pisa, Italy, 2007, pp. 1–6.
[48] L. Cheng, B. E. Henty, R. Cooper, D. Stancil, and F. Bai, “A measurement
study of time-scaled 802.11a waveforms over the mobile-to-mobile vehic-
ular channel at 5.9 GHz,” IEEE Commun. Mag., vol. 46, no. 5, pp. 84–91,
May 2008.
[49] T. S. Rappaport, Wireless Communications: Principles and Practice.
Englewood Cliffs, NJ: Prentice-Hall, 1999.
[50] G. Acosta-Marum and M. A. Ingram, “A BER-based partitioned model
for a 2.4 GHz vehicle-to-vehicle expressway channel,” Int. J. Wireless
Personal Commun., vol. 37, no. 3/4, pp. 421–443, May 2006.
[51] A. Goldsmith, Wireless Communications. Cambridge, U.K.: Cambridge
Univ. Press, 2005.
[52] H. Kremo, I. Seskar, and P. Spasojevic, “On spatial and temporal selec-
tivity of a class of vehicular short range channels,” in Proc. 20th Virginia
Tech Symp. Wireless Personal Commun., 2010, pp. 1–7.
[53] I. N. L. Tan, W. Tang, K. Laberteaux, and A. Bahai, “Measurement and
Analysis of Wireless Channel Impairments in DSRC Vehicular Communi-
cations,” Elect. Eng. Comput. Sci. Univ. Calif., Berkeley, CA, Tech. Rep.
UCB/EECS-2008-33, 2008.
[54] D. D. Stancil, F. Bai, and L. Cheng, “Communication systems for car-2-x
networks,” in Vehicular Networking. Hoboken, NJ: Wiley, 2010.
[55] J. Gross, O. Punal, and M. Emmelmann, “Multi-user OFDMA frame ag-
gregation for future wireless local area networking,” in Proc. IFIP Netw.,
Aachen, Germany, 2009, pp. 220–223.
[56] T. Zahn, G. O’Shea, and A. Rowstron, “Feasibility of content dissemi-
nation between devices in moving vehicles,” in Proc. 5th Int. CONEXT
Technol., 2009, pp. 97–108.
[57] L. Neubert, L. Santen, A. Schadschneider, and M. Schreckenberg,
“Single-vehicle data of highway traffic: A statistical analysis,” Phys. Rev.
E, Stat. Phys. Plasmas Fluids Relat. Interdiscip. Top., vol. 60, no. 6,
pp. 6480–6490, Dec. 1999.
[58] C.-G. Wallman and H. Astrom, “Friction Measurement Methods and the
Correlation Between Road Friction and Traffic Safety,” Swedish Nat.
Road Transp. Res. Inst., Linköping, Sweden, Tech. Rep., 2001.
[59] M. Bilban, A. Vojvoda, and J. Jerman, “Age affects drivers response
times,” J. Coll. Antropol., vol. 33, no. 2, pp. 467–471, Jun. 2009.
[60] R. Akcelik, “On the validity of some traffic engineering folklore,”
presented at the 22nd Conf. Australian Inst. Transp. Res., Canberra,
Australia, Dec. 2000.
Gustavo Marfia received the Laurea degree in
telecommunications engineering from the University
of Pisa, Pisa, Italy, in 2003 and the Ph.D. degree in
computer science from the University of California,
Los Angeles, in 2009.
He is currently a Postdoctoral Researcher with
the Department of Computer Science, University of
Bologna, Bologna, Italy. He is also with the Univer-
sity of California, Los Angeles. His research interests
include biomedical, transportation, and multimedia
networking systems, fields in which he has published
over 60 technical papers.
Dr. Marfia serves as an Associate Editor for the Elsevier Journal of Visual
Communication and Image Representation and for the Association for Com-
puting Machinery Computers in Entertainment.

Marco Roccetti received the Italian Laurea degree
in electronic engineering from the University of
Bologna, Bologna, Italy.
Since 2000, he has been a Full Professor with
the Computer Science Department, University of
Bologna. His research interests include computer
entertainment, intelligent transportation systems, and
web-based applications, ﬁelds in which he has au-
thored more than 250 technical papers.
Prof. Roccetti serves as an Associate Editor for
many international journals and is active in several
Italian and international projects.
Alessandro Amoroso received the Laurea degree in
physics from the University of Bologna, Bologna,
Italy, in 1987.
Since 1994, he has been with the Department of
Computer Science, University of Bologna, where
he is currently an Associate Professor in computer
science. He has participated in several scientiﬁc
projects, including the National Research Council,
National Energy Board, and University of California
at San Diego. His main research areas are mobile de-
vices, multimedia systems, and distributed systems.
His recent research activity has included work on vehicular ad hoc networks.
Giovanni Pau received the Italian Laurea doctor-
ate degree in computer science and the Ph.D. de-
gree in computer engineering from the University
of Bologna, Bologna, Italy, in 1998 and 2002,
respectively.
He is currently an Adjunct Associate Professor
with the Computer Science Department, University
of California, Los Angeles (UCLA). He is currently
leading the UCLA vehicular laboratory aimed at
building a publicly available vehicular testbed for the
research community. His main research interests lie
in the area of vehicular ad hoc networks and urban sensing.
Dr. Pau served as Vice Chair for North America for the IEEE Communi-
cations Society Multimedia Technical Committee and serves on the technical
program committee of several IEEE and Association for Computing Machinery
conferences. Finally, he is an Associate Editor for the Elsevier International
Journal of Ad Hoc Networks and the Springer International Journal on Peer to
Peer Systems.

Ieeepro techno solutions 2013 ieee embedded project greatest invehicular accident detection

Related slideshows

More Related Content

Ieeepro techno solutions 2013 ieee embedded project greatest invehicular accident detection