IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 7, NO. 4, DECEMBER 2006 429
The Impact of Cooperative Adaptive Cruise Control
on Traffic-Flow Characteristics
Bart van Arem, Member, IEEE, Cornelie J. G. van Driel, and Ruben Visser
Abstract—Cooperative adaptive cruise control (CACC) is an ex-
tension of ACC. In addition to measuring the distance to a prede-
cessor, a vehicle can also exchange information with a predecessor
by wireless communication. This enables a vehicle to follow its
predecessor at a closer distance under tighter control. This paper
focuses on the impact of CACC on traffic-flow characteristics. It
uses the traffic-flow simulation model MIXIC that was specially
designed to study the impact of intelligent vehicles on traffic flow.
The authors study the impacts of CACC for a highway-merging
scenario from four to three lanes. The results show an improve-
ment of traffic-flow stability and a slight increase in traffic-flow
efficiency compared with the merging scenario without equipped
vehicles.
Index Terms—Adaptive cruise control (ACC), intelligent
vehicles, traffic-flow simulation, vehicle–vehicle communication.
I. INTRODUCTION
D
URING THE past few decades, western society has been
constantly confronted with problems caused by increas-
ing road traffic. This increase in traffic demand leads to a
heavily congested network and has a negative effect on traffic
safety, air pollution, and energy consumption. The expectations
of the use of telematics technology in road traffic in this respect
are high, since this technology could lead to system innovations
(e.g., advanced vehicle guidance), which in the long term can
contribute to the problems faced [1], [2].
Advanced driver-assistance (ADA) systems support a driver
in his driving tasks. These systems are being developed because
they have the perspective to increase the driver’s safety and
comfort. Additionally, ADA systems can have a positive impact
on traffic-flow performance and reduce emissions and fuel
consumption. Examples of ADA systems are various forms
of cruise control, lane-keeping systems, and collision-warning
systems.
Manuscript received October 14, 2005; revised June 20, 2006, July 31, 2006,
and August 4, 2006. This work was supported by the Applications of Integrated
Driver Assistance (AIDA) program of The Netherlands Organization of Ap-
plied Scientific Research TNO and the University of Twente. The Associate
Editor for this paper was L. Vlacic.
B. van Arem and C. J. G. van Driel are with the Research Program of
Applications of Integrated Driver Assistance (AIDA), Centre for Transport
Studies, Faculty of Engineering Technology, University of Twente, 7500
Enschede, The Netherlands (e-mail: b.vanarem@utwente.nl; c.j.g.vandriel@
utwente.nl.).
R. Visser was with the Research Program of Applications Integrated Driver
Assistance (AIDA), Centre for Transport Studies, Faculty of Engineering
Technology, University of Twente, 7500 Enschede, The Netherlands. He is now
with 4Motion Consultancy, 3521 Utrecht, The Netherlands (e-mail: ruben@
4motion.nl).
Digital Object Identifier 10.1109/TITS.2006.884615
An ADA system that has been introduced by the automotive
industry is adaptive cruise control (ACC). ACC is a radar-based
system, which is designed to enhance driving comfort and con-
venience by relieving the driver of the need to continually adjust
his speed to match that of a preceding vehicle. The system slows
down when it approaches a vehicle with a lower speed, and
the system increases the speed to the level of speed previously
set when the vehicle upfront accelerates or disappears (e.g., by
changing lanes).
Vehicle-to-vehicle communication can further advance the
development of ADA systems. Cooperative ACC (CACC) is a
further development of ACC that adds vehicle-to-vehicle com-
munication, providing the ACC system with more and better
information about the vehicle it is following. With information
of this type, the ACC controller will be able to better anticipate
problems, enabling it to be safer, smoother, and more “natural”
in response. Although CACC is primarily designed for giving
the driver more comfort and convenience, CACC has a potential
effect on traffic safety and traffic efficiency. It is of importance
to understand the traffic-flow effects of CACC early in the
development so that, if they are discovered to inadvertently
create problems, the design can be adjusted accordingly before
adverse traffic effects are widely manifested. Apart from that,
it is recommended to study the traffic-flow effects of CACC so
that these (comfort) systems can be developed to best support
future advances.
Uncertainties exist about the traffic-flow impact of the rela-
tively new developed system CACC. The objective of this paper
is to assess the impact of CACC on traffic-flow characteristics
in terms of traffic stability and throughput.
This paper is organized as follows. In Section II, we review
the relevant literature. In Section III, we describe the traffic-
flow simulation model MIXIC. The CACC system is then
explained in Section IV. We describe the setup of the simulation
study and its results in Sections V and VI, respectively. The
results are discussed in Section VII. Section VIII contains our
conclusions.
II. R
EVIEW OF LITERATURE
According to two literature studies, ACC can contribute
to the stability of traffic flow limitedly and affect traffic
performance both positively and negatively [3], [4]. A low-
penetration level of ACC does not have any effect on traffic
flow, regardless of the time-gap set [5]. Even under the most
favorable conditions, with ideal ACC system design and per-
formance, it appears that ACC can only have a small impact on
highway capacity [6].
1524-9050/$20.00 © 2006 IEEE
430 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 7, NO. 4, DECEMBER 2006
In contrast to the voluminous literature on autonomous ACC,
the literature related to CACC is limited. In a number of studies,
the functionality, architecture, or design of CACC systems have
been described. However, extensively exploring the traffic-flow
effects, quantitatively in terms of throughput and capacity, has
been done by only a few researchers.
The concept of full platooning is described in [7]. High-
capacity values of up to 8500 vehicles an hour per lane can
be achieved, if separate infrastructure is available and all ve-
hicles using this lane can communicate with each other. This
concept of automated highway systems (AHSs) is especially
envisioned by American researchers. AHS have been defined
as vehicle–highway systems that support hands-off and feet-off
driving on dedicated freeway lanes. Different AHS have been
explored, and some have been investigated in depth.
Cooperative following (CF) uses automated longitudinal
control combined with intervehicle communication [8]. It al-
lows for anticipation to severe braking maneuvers in emerging
shock waves with the aim of smoothening traffic flow and
enhancing traffic safety. The functionality of CF has been
modeled in the microscopic-traffic simulation model MIXIC,
and the simulation has been run with a platoon of mixed CF
equipped and nonequipped vehicles. Although at a platoon level
better stability was achieved, the potential advantages on traffic-
flow efficiency could not be confirmed.
The CarTALK 2000 project focuses on developing coopera-
tive driver-assistance systems, which are based upon mobile-
intervehicle communication [9]. The traffic impacts of two
applications [basic warning function and early braking (i.e.,
a continuation of CF)] were assessed using MIXIC. Both ap-
plications concern vehicles that broadcast a message to other
vehicles when accelerating with −2.0 m/s
2
or less. The results
indicate an improvement of traffic stability in terms of a reduc-
tion in the number of shockwaves for all penetration rates and
headways tested.
The effects on capacity of increasing market penetra-
tion of both ACC and CACC vehicles, relative to manually
driven vehicles, was examined in a quantitative way by using
microscopic-traffic simulation [6]. A single highway lane with
a ramp–highway junction consisting of a single-lane offramp
followed immediately by a single-lane onramp has been con-
sidered. The analyses were initially conducted for the distinct
cases of 100% manually driven vehicles, 100% ACC (time gap
of 1.4 s), and 100% CACC (time gap of 0.5 s) to verify the
reasonableness of the results under these simplest cases. The
nominal-capacity estimates for the manual driving, ACC, and
CACC cases were 2050, 2200, and 4550 vehicles per hour,
respectively. Next, mixed vehicle populations were analyzed
in all feasible multiples of 20% of each vehicle type. It was
concluded that CACC can potentially double the capacity of a
highway lane at a high-CACC market penetration. The capacity
effect is sensitive to market penetration, based on the fact that
the reduced time gaps are only achievable between pairs of
vehicles that are CACC equipped.
The CHAUFFEUR 2 project has addressed three approaches
aiming to reduce a truck driver’s workload by developing truck-
platooning capability [10]. First, we have an electronic tow
bar in which a vehicle automatically follows a manually driven
leading vehicle. Due to vehicle-to-vehicle communication, the
following distance is very close (6–12 m, which is equal to
0.3–0.6 s at 80 km/h). Second, we have the Chauffeur assistant,
which enables the truck to follow any other vehicle on a
highway with a safe following distance using an ACC and a
lane-keeping system. Third, we have electronic coupling of
three trucks in a platooning mode. The leading vehicle is driven
conventionally, and both following vehicles follow. Also, in
this project, the following distance is very close (6–12 m). For
this last platooning mode, the trucks are equipped with vehicle-
to-vehicle-communication systems. Considering the results of a
simulation study with the microscopic-traffic simulation mod-
els VISSIM and FARSI, it is concluded that the main effects of
the CHAUFFEUR 2 systems are a better usage of road capacity,
up to 20% reduction in fuel consumption, and increased traffic
safety. However, it has been remarked that platooning is
mostly feasible at night or on sections with low-traffic volume
because, during high-traffic volume, the stability of traffic flow
decreased.
From this literature review, the following conclusions can
be drawn. First, vehicle-to-vehicle communication can provide
an ACC system with more and better information about the
vehicle it is following. Not only following distance and speed
difference with respect to its direct predecessor are considered,
but also speed changes can be coordinated with each other. The
information could include precise speed information, acceler-
ation, fault warnings, warnings of forward hazards, maximum
braking capability, and current braking capability. With infor-
mation of this type, the ACC controller can better anticipate
problems, enabling the vehicle to be safer, smoother, and faster
in response and, as a result, enable closer vehicle following.
Time gaps could be as small as 0.5 s.
Second, CACC has the potential to increase capacity by
minimizing time gaps between consecutive vehicles and traffic-
flow stability by improving string stability. This improved
performance can only be achieved when pairs of vehicles are
equipped with the CACC system. Therefore, the improvement
of traffic efficiency largely depends on the level of CACC
deployment.
Third, extensive research into the traffic-flow impact of
CACC, in terms of traffic-flow stability and throughput, is
lacking. The limited CACC effect studies that have been per-
formed emphasize that CACC is able to increase the capacity
of a highway significantly. CACC can potentially double the
highway capacity at a high market penetration.
III. T
RAFFIC SIMULATION MODEL MIXIC
In order to study the potential impact of ADA systems,
such as CACC, the modeling approach should be suitable for
analyzing different assumptions for ADA-system functionality,
roadside systems, driver behavior, and vehicle dynamics. Fur-
thermore, it should be capable of assessing impacts on traffic
performance, traffic safety, exhaust-gas emission, and noise
emission. In order to meet these requirements, a stochastic
simulation model MIXIC was developed [11].
The microscopic-traffic simulation model MIXIC simulates
traffic on a link level in a network. Given an input of traffic flow,
VA N A R E M et al.: IMPACT OF COOPERATIVE ADAPTIVE CRUISE CONTROL 431
MIXIC simulates traffic behavior on this link and produces
traffic statistics. MIXIC uses real traffic measurements (time
instant, lane, speed, and vehicle length) to generate traffic at
the start of the simulation run. At each step (set to 0.1 s),
new vehicle positions are calculated by a driver model and a
vehicle model. The driver model produces driver actions, such
as lane changing and new pedal and gear positions. These driver
actions are input for the vehicle model, which calculates the
resulting acceleration and position of the vehicle. The main
components of MIXIC are described in the remainder of this
section.
The traffic generator decides whether or not to place new
vehicles at the start of the first road link. Vehicles are generated
from so-called traffic “injection” files and are assigned specific
vehicle/driver data and an initial state. A traffic injection file
consists of recorded real-world data of individual vehicles
(arrival time, position, lane, speed, and length). Vehicle types
and driver types (reaction time, desired speed, etc.) are assigned
randomly (using occurrence frequencies). The use of injection
files to generate input for a microscopic model has three ad-
vantages. First, it is by definition realistic. Second, it puts the
traffic-evolution model to the test (this model should in any
case be able to process the amount of traffic offered by the
injection file, traffic being actually observed). Third, it allows
for calibration and validation by comparing the model with
measurement further downstream [12].
The driver model consists of three main components: the
lane change model, the longitudinal model, and a component,
which describes the interaction between the driver and the
ADA system. The lane-change model consists of a mandatory
lane-change model (represents forced lane changing due to
geometric factors) and a free lane-change model (represents
overtaking). The longitudinal model distinguishes free-driving
behavior (the driver attempts to reach or maintain his intended
speed) and car-following behavior (the driver adjusts his speed
and/or following distance with respect to traffic ahead). The car-
following model implemented in MIXIC is derived from the
optimal control model of Burnham et al. [13]. It is based upon
the assumption that drivers try to keep the relative speed to the
lead car zero and simultaneously attempt to keep the clearance
at a desired value. In addition to the original model, also, the
relative speed to the vehicle ahead of the lead vehicle is taken
into account because it contributes to the stability of the traffic
flow [12]. In contrast with other microscopic-traffic models, it
is possible in MIXIC to specify the driver interaction with an
ADA system.
The vehicle model describes the dynamic vehicle behavior
as a result of the interaction with the driver and the road,
taking into account the ambient conditions. The vehicle model
uses information on the characteristics of the vehicle, the road
geometry, the condition of the road, and the wind. The output
of the model is an updated vehicle acceleration, which is used
to calculate a new speed and position of the vehicle.
To study the effects of ADA systems, MIXIC also contains
a detailed model describing the behavior of the ADA system in
the paper. An ACC model is already available in MIXIC. If the
driver has switched on the ACC system, the ACC model takes
over the longitudinal driver model. To obtain the information
Fig. 1. Position of CACC model in MIXIC.
needed for this control task, a radarlike sensor is used. The
sensor is characterized by its delay and its maximum detection
range. Failures in the operation of the sensing and communica-
tion equipment were not considered.
The simulation model MIXIC was originally developed in
the early 1990s. It was calibrated for different two-, three-,
and four-lane motorway situations. The motorway situation
studied in this paper was calibrated against data from the A4
motorway near Schiphol [12], [14]. Both the longitudinal and
lateral models appeared to represent real-life traffic adequately.
It appeared that the traffic flows on a four-lane setting were
considered realistic. On the part where the traffic merges to
a three-lane motorway, traffic volumes as high as 7700 pcu/h
were observed for 5-min intervals. Although this appeared to
be high, empirical studies have indicated that such peak flows
are indeed high but not unrealistic [15].
IV. C
OOPERATIVE ADAPTIVE CRUISE CONTROL (CACC)
The characteristics of MIXIC are suitable for exploring the
impact of CACC on the traffic flow. First, however, a new
CACC model had to be designed in MIXIC. The basis of the
longitudinal driver model in MIXIC is the calculation of a
desired acceleration of a driver. If a CACC system takes over
a part of the longitudinal driving task of a driver, a reference
acceleration of the CACC controller is calculated instead. This
reference acceleration is used to determine the real acceleration
of the vehicle in the MIXIC vehicle model. The position of
the CACC model in MIXIC as described above is presented
in Fig. 1.
For modeling purposes, a CACC-equipped vehicle can be
divided into two distinct components: the CACC controller de-
livering reference values and a vehicle model transforming the
reference values into actually realized values, using the vehicle
model. The acceleration reference demand from the CACC
controller must be determined and then fed into the vehicle
model. The equations for the speed and distance controller have
been derived from [16]. Since MIXIC differs from the model
432 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 7, NO. 4, DECEMBER 2006
used in [16], some changes have been made in these algorithms
in order to develop a workable model.
The acceleration demand can be computed on the basis of the
difference between current and intended speed (a
ref_ν
) or on
the basis of the distance and speed difference between the ego
vehicle (i.e., CACC-equipped vehicle) and target vehicle (i.e.,
predecessor of ego vehicle) (a
ref_d
). The acceleration demand
is given by the most restrictive one:
a
ref
= min(a
ref_ν
,a
ref_d
). (1)
The resulting reference acceleration is limited between the
maximum comfortable acceleration 2 m/s
2
and the maximum
comfortable deceleration -3m/s
2
.
Let ν
int
and ν denote the intended and current speed, respec-
tively, of the CACC set by the driver in meters per second. The
acceleration demand, based on speed difference, is given by
a
ref_ν
= k · (ν
int
− ν) (2)
with k as a constant-speed error factor.
The computation of the reference acceleration, based on the
distance and speed difference between the ego vehicle and
target vehicle, is slightly more complicated. Let ν
p
denote the
speed of the target vehicle, and let r and r
ref
denote the current
and reference clearance to the target vehicle in m, respectively.
Let a
p
denote the acceleration of the target vehicle. The refer-
ence acceleration, based on the distance and speed difference
between the ego vehicle and target vehicle, is given by
a
ref_d
= k
a
· a
p
+ k
ν
· (ν
p
− ν)+k
d
· (r − r
ref
) (3)
with k
a
, k
ν
, and k
d
as constant factors.
The reference clearance r
ref
is defined as a maximum among
safe following distance (r
safe
), following distance according
to the system time setting (r
system
), and a minimum allowed
distance (r
min
), set at 2 m. The safe following distance is
computed on the basis of the speed ν of the ego vehicle and
the deceleration capabilities d and d
p
of the ego vehicle and the
target vehicle, respectively:
r
safe
=
ν
2
2
·
1
d
p
−
1
d
(4)
where, for simplicity, we have assumed a communication delay
equal to zero. The following distance according to the system-
target time-gap setting is given by
r
system
= t
system
· ν (5)
where t
system
is chosen equal to 0.5 s, if the target vehicle has
CACC, and 1.4 s otherwise.
The constant factor k was chosen equal to 0.3 in accordance
with earlier MIXIC studies [12]. The constant factor k
a
was
chosen equal to 1.0 in accordance with [6]. The default values
of MIXIC for k
ν
and k
d
are 3.0 and 0.2, respectively. In a recent
MIXIC study, these values were set more “strongly” to 0.58 and
0.1, respectively [17]. Fig. 2 illustrates how combinations of
the respective CACC parameters of the reference-acceleration
function (by systematically varying the values for k
ν
and k
d
)
Fig. 2. Relative speed of a CACC vehicle approaching a slower CACC
equipped vehicle.
Fig. 3. Speed of manually controlled vehicles.
influence the relative speed of a CACC-equipped vehicle ap-
proaching a slower CACC-equipped vehicle.
In this paper, the parameter setting of k
d
=0.1, k
ν
=0.58,
and k
a
=1.0 was selected, since this setting resulted in the
most smooth and fast reaction of the CACC controller without
leading to unsafe situations compared to the other settings
(Fig. 2).
Regarding the human interaction with the CACC, it is as-
sumed (similarly to the operation of an ACC in MIXIC) that the
driver can switch the system on and off. The driver will switch
on the CACC as much as possible but will turn the CACC off if
a deceleration required is stronger than the CACC capability or
in case of a mandatory lane change.
Further tests to check the CACC operation in MIXIC were
performed, assuming a platoon of five vehicles. The first vehicle
drives at 80 km/h, and the other vehicles approach in a platoon
of four vehicles. In the reference situation, all vehicles are under
manual control; in the other situation, the vehicles are CACC
controlled. Fig. 3 displays the speed of manually controlled
vehicles. Fig. 4 displays the speed of the CACC-controlled
vehicles.
We conclude from Figs. 3 and 4 that the approaching CACC-
equipped vehicles react faster on the decelerating predecessor
than in the scenario in which all vehicles are manually driven.
The lines in the diagrams of the 100% CACC-equipped platoon
stick close together, which indicates that the time between an
accelerating or decelerating action of the predecessor and the
VA N A R E M et al.: IMPACT OF COOPERATIVE ADAPTIVE CRUISE CONTROL 433
Fig. 4. Speed of CACC controlled vehicles.
Fig. 5. Simulated highway configuration.
same action of the successor is less than in the scenario of
100% manually driven vehicles. Furthermore, the curves of
the CACC-equipped platoon are smoother, which illustrates a
smoother behavior of the CACC-equipped vehicles above man-
ually driven vehicles. Analysis of the acceleration reveals that
the CACC vehicles do not decelerate stronger than −2 m/s
2
,
while the manually driven vehicles decelerate as strong as
−2.5 m/s
2
.
V. S
IMULATION SETUP
The traffic simulation model MIXIC was used to examine
the impact of CACC on the traffic flow. The basic configuration
used in this paper is a four-lane highway with a road narrowing
by a lane drop. A lane drop corresponds to a mandatory
lane change in the MIXIC traffic simulation model. When
a mandatory lane change is carried out, the drivers turn off
their CACC system. Once the mandatory lane change has been
carried out, the system is turned back
ON under the normal
conditions maintained by MIXIC. A lane drop makes it possible
to measure the maximum traffic volume at different penetration
rates of CACC when the traffic volume on the link before the
lane drop nears a congestion state. In addition, a number of
experiments were conducted, with a special lane for CACC
vehicles, to study whether this would lead to additional traffic-
flow benefits. The highway configuration, presented in Fig. 5, is
split into six links of 1000 m each. A left lane drop is modeled
from 4 km after the start of the simulation.
A prewarning of the “merge” is given to the driver, 1350 m
before the transition, from four to three lanes. In the scenarios
with a CACC lane, the CACC lane is introduced after 2000 m
on the left most lane, expanded with one lane after 3000 m, after
which the left most CACC lane ends after 4000 m (see dark lane
sections). CACC drivers were not assumed to go to the CACC
lane consciously. However, when they are driving on the CACC
lane, they will not leave it. The simulation time was 150 min per
run. A measurement point is placed in the middle of each link
for statistical analysis. Data from the first and last sections were
not analyzed to avoid transient aspects.
For the traffic generation, data from the A4 highway, near
Schiphol in The Netherlands, was used. The MIXIC behavior
for this data set was previously calibrated [12]. The data set
contains sufficiently high-traffic volumes (up to 7600 pcu/h)
that can lead to congestion in the narrowing scenario under
study. The number of trucks and vans in this data set is small
(vehicles 94%, vans 4%, and trucks 2%).
Regarding the operation of the CACC, the time-gap setting
of the CACC system is set on 0.5 s, if it is following a CACC-
equipped vehicle, and on 1.4 s, if it is following a non-CACC-
equipped vehicle, respectively. The penetration rate of CACC
systems was varied in multiples of 20%. This resulted in one
reference case with no CACC vehicles, five CACC scenar-
ios without a CACC lane, and three CACC scenarios with a
CACC lane.
To ensure statistical validity, five stochastically independent
simulations were performed for each selected scenario. Analy-
sis of variance (ANOVA) was used to test whether the means of
an indicator in different scenarios were significantly different
from each other. Post hoc Tukey’s tests were used to study
whether the value of an indicator on a specific link differed
from the values on other links. To test to what extent a CACC
scenario was significantly different from the reference case,
post hoc Dunett’s tests were performed. These tests enable to
compare the means of a group to a control group. In this case,
the control group is the 0%-CACC scenario (reference case).
VI. S
IMULATION RESULTS
We illustrate the results by the following output variables
that were analyzed on link four (just before the lane drop) and
link five (just after the lane drop). Since shockwaves represent
variations in the flow that propagate through the traffic, the
number of shock waves on a highway stretch is used as a
measure for traffic stability. It is defined as an observation of at
least three vehicles on the same lane within a mutual distance of
50 m and within a time period of 3 s with a deceleration stronger
than −5 m/s
2
. Furthermore, we used the average speed on a
link as an indicator for traffic throughput. Finally, we measured,
the three highest 5-min average traffic volumes are used as an
indication of the roadway capacity.
In general, we observed that during the simulations, a number
of vehicles did not succeed to merge from link four to link five
(i.e., these vehicles were removed from the simulation). This
especially occurred in the case of a CACC lane and in the case
of a high-penetration of CACC vehicles, leading to reduced
merging gaps because of the close following.
Fig. 6 shows that the number of shockwaves, just before the
lane drop, decreases drastically when more CACC-equipped
vehicles are present. The same was found just after the lane
