Академический Документы
Профессиональный Документы
Культура Документы
DOI 10.3233/ICA-150486
IOS Press
Abstract. The vehicular collision can lead to serious casualties and traffic congestions, especially multiple-vehicle collision.
Most recent studies mainly focused on collision warning and avoidance strategies for two consecutive vehicles, but only a few on
multiple-vehicle situations. This study proposes a coordinated brake control (CBC) strategy for multiple vehicles to minimize the
risk of rear-end collision using model predictive control (MPC) framework. The objective is to minimize total impact energy by
determining the desired braking force, where the impact energy is defined as the relative kinetic energy for a consecutive pair of
vehicles. Under the MPC framework, this problem is further converted to a quadratic programming at each time step for numerical
computations. To compare the performance, three other control strategies, i.e. direct brake control (DBC), driver reaction based
brake control (DRBC) and linear quadratic regulator (LQR) control are also considered in this paper. The simulation results,
in both a typical scenario and a huge number of scenarios under stochastic situations, show that CBC strategy has the best
performance among these four strategies. The proposed CBC strategy has the potential to avoid the collision among a group of
vehicles, and to mitigate the impact in cases where the collision is unavoidable.
Keywords: Driver assistance systems, longitudinal collision mitigation, model predictive control, multiple vehicles, V2V
ISSN 1069-2509/15/$35.00
c 2015 IOS Press and the author(s). All rights reserved
172 J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC
tions. The problem with such systems is that vehicular rectly used the master cylinder pressure as the control
accidents on highways, particularly under congested variable [23]. Technologies of this nature can be used
conditions, result from the collective behavior of mul- as the foundation for implementation of the approach
tiple closely spaced vehicles (i.e., a coupled group) in proposed in this study.
string formation. Therefore, the last vehicle in the for-
mation experiences an accumulated time delay in re- 1.2. Preview of key results
sponding to the behavior of the vehicles in the front.
In fact, a string of coupled vehicles on a highway The objective of this study is to design a coordi-
is more likely to result in multiple-vehicle accidents, nated brake control (CBC) strategy to avoid longitudi-
which is believed to be one of the detrimental effects of nal collision, or, if impossible to avoid, to mitigate the
conventional longitudinal collision avoidance systems. impact of multiple-vehicle longitudinal collision. The
With increasing computing and wireless communica- total relative kinetic energy is used as a novel perfor-
tion abilities, such as Vehicle-to-Vehicle (V2V) [4], the mance minimization index rather than minimizing rel-
cooperative driver assistance systems (DAS) have the ative speed or relative distance errors, which has been
potential to further mitigate this effect [34,39]. The co- most often employed in previous studies. Intuitively, a
operative DAS has a great potential for avoiding such vehicular system of low kinetic energy poses a low risk
serious accidents or, at least, for minimizing the im- of damage and is more stable. Hence, we assume that
the safety of the multiple coupled vehicles increases
pact if the collision is unavoidable by simultaneously
with decreasing relative kinetic energy. The use of the
controlling the braking of multiple vehicles [25]. In
total relative kinetic energy facilitates more safety un-
V2V systems, the use of Dedicated Short Range Com-
der conditions where coupled vehicular systems are
munication (DSRC) or Wireless Access in Vehicular
composed of a variety of vehicle masses, which is gen-
Environments (WAVE) has the overall advantages of
erally the case under realistic conditions.
extensive network radio communication capabilities
The proposed study is different from the concept of
that demonstrate low-latency and high throughput, and
the automated platoon, in which all vehicles have au-
which is both robust and scalable [12]. Due to these
tomated control capabilities, and are coordinated and
kinds of characteristics, DSRC/WAVE is suitable for
organized in a hierarchical automated highway sys-
basic safety applications, and can be also combined tem [24]. This study employs V2V and vehicle con-
with other on-board sensor systems, such as vehicle trol capabilities applicable to current practical highway
on-board sensors (e.g. lidar/radar), precision position systems by coordinating on local vehicle behaviors.
systems (e.g. GPS/DGPS), and vehicle state estimation Therefore, it is necessary to consider the relationship
systems (e.g. wheel speed, vehicle velocity, accelera- between the coupled vehicle group and its surrounding
tion and yaw rate). In addition, the relative position and environment, i.e., those vehicles in the front and rear
speed of vehicles can be obtained and shared by multi- of the group that cannot be coupled into the system be-
source information fusion or communication [35]. The cause of the lack of V2V or automated control capabil-
accuracy of these parameters can meet the require- ity. Under actual traffic conditions, the coupled group
ments for application in connected vehicles [10,39]. is usually heterogeneous rather than homogeneous,
The main advantage of the cooperative DAS is its and the group is generally comprised of vehicles of dif-
use of wireless communication for information pass- ferent masses, lengths, and braking systems operated
ing to reduce time delays and to compensate for the de- by drivers exhibiting a variety of car-following [3,30]
ficiencies associated with conventional sensors [4,34]. and lane-changing behaviors [33]. The heterogeneous
The internal stability and string stability for multiple nature of a coupled group can seriously affect platoon
vehicle following was analyzed in [24,43], which was control [24] and also the capacity of signalized inter-
critical for the practical traffic flow safety. Naus et al. sections [40]. Some other studies also put a lot atten-
derived a necessary and sufficient frequency-domain tion to heterogeneous issue, e.g. [20,37]. Here, we con-
condition for vehicle string formation stability and ex- sider the case of a heterogeneous group of coupled
perimented with two vehicles equipped with coopera- vehicles. To verify the performance of the proposed
tive adaptive cruise control (CACC) [28]. For heavy- CBC strategy, three other control strategies: 1) Direct
duty trucks, reducing the actuator delay is also critical Brake Control (DBC), 2) Driver Reaction based Brake
for control system performance. For the application of Control (DRBC), and 3) Linear Quadratic Regulator
such systems, Lu and Hedrick et al. proposed the de- (LQR) control, are also built for simulation studies,
sign of a hydraulic brake for passenger vehicles that di- and compared with the proposed approach.
J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC 173
Fig. 1. Overall configuration of multiple vehicle collision mitigation control, where Fi is the actual braking force of the i-th vehicle, Fi,des is
the desired braking force, dr is the actual relative distance, and ds is the safe distance gap.
The main contributions of this study are as follows. standard highway driving, the distance between adja-
(a) Multiple coupled vehicles are directly taken into ac- cent vehicles in terms of time according to their speed
count rather than considering only two vehicles, as typ- (time headway (THW)) ranges from 0.7 to 3.8 s [6].
ically done in previous work. (b) The total relative ki- Therefore, most vehicles in car-following scenarios are
netic energy between each consecutive pair of vehicles actually coupled in highway traffic and can be orga-
is used as the objective function for collision avoid- nized into groups. The boundaries among groups de-
ance and impact mitigation, and the problem is formu- pend on several factors, including (1) the type of ve-
lated as an optimal control, which is further simpli- hicles involved, (2) the maximum deceleration capa-
fied as a linear model predictive control (MPC) for fu- bility, (3) current speed and THW, (4) traction condi-
ture real-time implementation. (c) Simulation is con- tions between the tires and roadway, and (5) wireless
ducted using Matlab, where the vehicle masses of the communication availability. Wireless communication
coupled vehicle group are generated and assigned ran- is critical for coordinated control of braking of each ve-
domly and the vehicle mass in turn determines vehi- hicle and for broadcasting the vehicle types and decel-
cle length and deceleration capability according to an eration capabilities of the vehicles of which the group
empirical formula.
is composed in the message set.
The remainder of this paper is organized as follows.
The overall control system for each group of coupled
Section 2 presents the system analysis and modeling,
vehicles contains both upper and lower level controls,
and Section 3 presents vehicle longitudinal dynamics.
as shown in Fig. 1. The function of the upper level con-
Section 4 is devoted to MPC design depending on the
trol is to generate the desired braking force for each ve-
vehicle longitudinal dynamic model. The simulation
hicle to avoid collision or minimize potential impact if
scenario and analysis are discussed in Section 5, and
Section 6 presents concluding remarks. collision is unavoidable. The controller is also subject
to several constraints, including the deceleration capa-
bilities of each vehicle. The lower level control acti-
2. System analysis and modeling vates the brake system to achieve the desired braking
force. This study primarily focuses on the upper level
For this study, we consider a system composed of control. From a practical viewpoint, each vehicle may
closely following vehicles, i.e. multiple longitudinally have a different powertrain depending on the vehicle
coupled vehicles, for collision avoidance and impact type. Most heavy-duty vehicles are equipped with vari-
mitigation. Adjacent vehicles in a given lane are cou- ous braking systems, such as pneumatic brakes, engine
pled if their states (speed and distance) jointly satisfy retarders, and transmission retarders, whereas many
some predefined conditions, i.e., if any longitudinal passenger cars are equipped with disk or drum brakes,
maneuver of the subject vehicle would require its fol- which may be hydraulic or electric. For simplicity of
lower to take immediate action to avoid collision. For modeling, the differences between vehicle types in a
174 J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC
coupled group are quantified according to the deceler- Here, the magnitude of each period must be quanti-
ation capability of each vehicle. tatively determined in a reasonable manner. We focus
As suggested in [17], the collision impact mitiga- exclusively on the control action by braking beginning
tion system is only activated if a certain collision threat at t0 to the time point when the maneuver (collision
threshold is reached. This condition can be determined avoidance or impact mitigation) is completed. The sys-
by sensor detection, risk assessment, and warning the tem considers only coupled vehicles.
driver. If a driver fails to respond appropriately to the
warning of an accident risk, the CBC strategy is acti- 2.2. Upper level control strategy
vated as the last resort.
To properly define the system, the following as- The differences between vehicles include mass, de-
sumptions are made. celeration capability, sensor detection, control actua-
1) Assumption 1: The desired minimum decelera- tion, and overall time delay. It is impractical to apply a
tion of the first vehicle and the desired maximum constant deceleration to each vehicle. Instead, the brak-
deceleration of the last vehicle in the coupled ing torque of each vehicle must be adjusted according
group are known. to the current and predicted states on the basis of sen-
2) Assumption 2: Each vehicle has a front-looking sor measurements. The upper level control strategy can
and rear-looking radar/lidar to detect inter- be stated as follows.
vehicle distance and relative speed reliably. 1) The first vehicle should brake as little as possible
3) Assumption 3: Vehicles in the coupled group can to fully use the space in front and to minimize
communicate with each other (duplex communi- impact with following vehicles. However, the de-
cation). gree of braking is not completely free but is po-
4) Assumption 4: All vehicles in the coupled group tentially limited by vehicles ahead that are not in
are equipped with devices for controlling the the coupled group. Hence, a tradeoff is always
braking force. made for the first vehicle between reducing the
impact with the following vehicle and the risk of
2.1. Overall system structure collision with those ahead.
2) The last vehicle should apply the brake to the
Figure 1 shows the overall system structure for greatest extent possible to fully use its rear space
multiple-vehicle brake control. The system includes and to minimize its impact to the vehicle in front
remote sensor detection, system state estimation, a of it. However, again the degree of braking is not
DSRC communication network, control and decision completely free but is potentially limited by ve-
making, and control of actuators. The time to collision hicles not in the coupled group that are following
(TTC) is a critical variable for automatic control. The behind. Hence, a tradeoff is always made for the
TTC is refined into the following time point sequences last vehicle between reducing the impact with the
for improved performance in active vehicle safety. vehicle ahead and the risk of collision with those
1) System coupling time: A group of vehicles in the behind.
same lane can be in coupled status for a certain 3) The desired braking force between all other ve-
time period denoted as the coupling time. hicles will be determined by the current and pre-
2) System warning time: The driver receives a warn- dicted states of each vehicle, which are estimated
ing signal before brake activation for passive col- according to real-time sensor measurements.
lision avoidance. 4) State prediction in the finite time horizon of each
3) Braking start time t0 : The CBC actuation starts vehicle is conducted by an MPC approach.
time necessary to avoid collision if the driver The algorithm for longitudinal collision avoidance
does not appropriately respond to the warning, and impact minimization should consider the follow-
or, if unavoidable, the earliest time possible to ing factors: (1) Multiple-vehicle scenario, (2) mechan-
activate the brake for collision active impact mit- ical definition of impact, (3) mass of each vehicle, (4)
igation. relative distance and relative speed at braking, (5) rel-
4) Collision start time: Time instant at which at ative distance and speed at collision, (6) braking capa-
least one pair of vehicles collides. bility of each vehicle, (7) braking constraint applied to
5) Collision end time: Time instant at which the col- the first vehicle, and (8) braking constraint applied to
lision scenario ends. the last vehicle.
J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC 175
2.3. Relative kinetic energy consideration of this factor could yield a more
accurate prediction of the braking performance,
A quantitative description of the multiple-vehicle it would complicate the MPC design. Hence, we
collision scenario also depends on vehicle types. In neglect the pitching motion to reduce complexity
these cases, the kinetic energy during collision is not of the controller design.
fixed. Moreover, it would be difficult to accurately es- 5) The vehicle moves in a flat plane.
timate the collision impact for each pair of vehicles. Based on the assumptions above, the longitudinal
Therefore, we use the relative kinetic energy before dynamics for each vehicle in the coupled group are as
collision as the measure of the potential impact of col- follows [18]:
lision between a pair of vehicles, which in turn is used
to estimate the overall system potential impact. This is xi (t) = vi (t)
physically justified because the energy of the collision 1 , (2)
is the work done by the relative kinetic energy. The fol-
vi (t) = Fi (t) CA vi2 mi gf
mi
lowing total relative kinetic energy E(t) as a function
i = 1, 2, . . . , N
of time t serves as the integrand of objective function:
where xi (t) is the longitudinal distance with respect
1
N
2 to a common inertial coordinate system, CA is the
E(t) = mi (vi1 (t) vi (t)) , (1)
2 i=2 lumped aerodynamic drag coefficient, is the accelera-
tion due to gravity, f is the coefficient of rolling re-
where vi and mi are the velocity and mass of the i- sistance, and Fi (t) denotes the actual braking force
th vehicle, respectively, and N denotes the number of in the longitudinal direction of the i-th vehicle. Be-
vehicles in the coupled group. Here, the mass of the cause sensors and actuators in mechanical systems
first vehicle is not considered. As discussed previously, have their own inherent dynamics, the actual braking
the minimization of the total relative kinetic energy is force Fi (t) cannot accurately follow the desired brak-
helpful in reducing the impact of collision when colli- ing force Fi,des (t) immediately. Hence, a lumped first-
sion is unavoidable. order inertial dynamics between Fi (t) and Fi,des (t) is
assumed [19]:
min
where k represents the time step.
ai ai ( k + j| k) amax
The objective function over the predictive horizon is
amin a ( k + j| k) c
i
1 1 1
,
cN a1 ( k + j| k) amax
1
Np N
N
J(k) = mi xi1 ( k + j| k) xi ( k + j| k) ds
2 j=1 i=2 (5)
i = 1, 2, . . . , N, j = 0, 1, . . . , Np 1,
[vi1 ( k + j| k) vi ( k + j| k)]2 ,
where (k + j|k) denotes the predicted state at time The abovementioned control aims to mitigate the
k + j obtained by starting from the current time k, and impact of collision, but may not be able to avoid colli-
Np denotes the length of the predictive horizon. sion completely under some conditions. In some cases,
The system is constrained according to vehicle de- collision is unavoidable and corresponds to the infea-
celeration abilities, as expressed in Eq. (6a). In addi- sibility of problem Eq. (7). One solution is to set a
tion, the first vehicle in the front end and the last ve- slack variable in the constraint Eq. (6d), which enables
hicle in the rear end of the coupled group provide the Eq. (6d) to be imposed as a soft constraint. Another
constraints expressed in Eqs (6b) and (6c). Finally, the method is to use the last control efforts when encoun-
coupled group must have the ability to avoid collision, tering infeasibility. In our design, constraint violation
so the constraint in Eq. (6d) is established. usually indicates the inevitability of a collision. Hence,
we employ the second method.
amin
i ai (k + j|k) amax
i , (6a)
The solution to the problem given by Eq. (7) is the
amin
1 a1 (k + j|k) c1 , (6b) sequence of input signal Fdes (k) = [Fdes (k|k), . . . ,
Fdes (k + Np 1|k)] RN Np 1 , where Fdes (k + j|k)
cN a1 (k + j|k) amax
N , (6c)
= [F1,des (k + j|k), . . . , FN,des (k + j|k)] RN 1 , at
xi1 (k + j|k) xi (k + j|k) ds , (6d) the current time instant k.
The solution of Eq. (7) minimizes the cost function
i = 1, 2, . . . , N, j = 0, 1, . . . , Np 1,
in Eq. (5) subject to constraints in Eqs (6a) to (6d).
where, amax
i , amin
i are the maximum possible accelera- Only the first part of the computed optimal control in-
tion and maximum possible deceleration; c1 is the min- put sequence is applied to the coupled system during
imum deceleration value of the first vehicle of the cou- the next sampling interval. At the next time step k + 1,
pled group for avoiding collision with a vehicle ahead; Eq. (7) is resolved over a shifted horizon to utilize
cN is the maximum deceleration value of the last vehi- the updated information regarding the coupled vehicle
cle in the coupled group for avoiding collision with a state.
vehicle behind; and ds is the safe distance gap.
Hence, we can write the control problem in the fol- 4.2. Linear MPC formulation
lowing compact form.
The nonlinear optimization of Eq. (7) is nontrivial
1
Np N
min J(k) = mi [vi1 ( k + j| k) and must be solved numerically. Here, Eq. (7) is lin-
Fdes (k) 2 j=1 i=2 earized with respect to the operating point, and then
transformed into a constrained quadratic programming
vi ( k + j| k)]2 (QP) problem at each control instant. The prediction
subject to horizon is 0.1 s; therefore, the model is linearized at
the current speed. Such a strategy naturally ensures
xi ( k + j + 1| k) = xi ( k + j| k) that vehicle speed is nearly constant over the predictive
+vi ( k + j| k) t horizon. Then, the braking force, aerodynamic drag,
i k + j + 1| k) = vi 2( k + j| k)
(
v and rolling resistance can be regarded as a single vari-
1 1 1
u2,des (k + j|k) , (12)
cN a1 (k + j|k) aN max
udes (k + j|k) = u3,des (k + j|k) RN 1 .
xi1 (k + j|k) xi (k + j|k) ds
..
.
uN,des (k + j|k) i = 1, 2, . . . , N, j = 0, 1, . . . , Np 1.
178 J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC
Here, Q has the appropriate value and dimension. A. Coordinated brake control (CBC)
Moreover, Eq. (12) has a quadratic optimal target term
and linear constraints. Hence, Eq. (12) can be recast as CBC coordinates the braking force for each vehicle
the constrained QP problem given by Eq. (13). using MPC, which is formulated into a constrained QP
1 T problem. The design details of CBC have been given
min J(k) = U HU + P T U + C in Sections 3 and 4.
U 2 (13)
subject to Acon U bcon . B. Direct brake control (DBC)
Here, U , H, P , C, Acon , bcon have appropriate val-
DBC employs equivalent assumptions and condi-
ues and dimensions. Model Eq. (12) is linearized
tions as CBC. Vehicles in the coupled group must have
around an operating point. Therefore, the linear time-
invariant model Eq. (12) at time k is used to predict the V2V communication and automatic braking capabil-
coupled group state X ( k + j| k) , j = 0, 1, . . . , Np ities. The difference between DBC and CBC is that
1 in the predictive horizon. Accordingly, the optimiza- DBC directly employs the full deceleration ability of
tion variables udes ( k| k) are used to solve the braking each vehicle rather than coordinating the braking force
force Fi,des ( k| k) for each vehicle during the braking for each vehicle under emergency conditions. There-
process. fore, each vehicle in the coupled group immediately
and fully employs braking when confronting critical
4.3. MPC algorithm in implementation events.
For numerical calculation, it is suggested to sequen- C. Driver-reaction based brake control (DRBC)
tially perform the following steps.
DRBC does not assume that vehicles in the coupled
Step 1: Initialize X(k) according to sensor mea-
surements and communication to predict the group have V2V communication and automatic brak-
forward Np time steps. ing capabilities. Only the driver judges the driving sit-
Step 2: Find the optimal desired braking force Fdes uation and brakes the vehicle fully after a certain driver
(k|k) numerically using the QP method. reaction time for emergency situations.
Step 3: Apply the control (desired braking force)
D. Linear quadratic regulator control (LQR) control
Fdes (k|k) to the system and then generate
the system state of the next time step.
This method employs LQR control during braking,
Step 4: If the speed of each vehicle is not reduced
which is a classical optimal control strategy for Coop-
to zero, estimate the system state to obtain
erative Adaptive Cruise Control (CACC) during nor-
X (k + 1) according to sensor detection.
mal driving conditions. In this study, the constant THW
Step 5: Set k := k + 1; X(k) = X (k + 1), and go
policy is used for the car-following model, and the con-
to Step 2.
trol objective is to follow the drivers desired spacing
(see [26] for more details).
5. Simulation and analysis
5.1. Simulation parameters
The coupled group must allow for different types
of vehicles because the control algorithm is designed Vehicles with large masses are usually also long and
for practical coordination of multi-vehicle longitudinal have limited deceleration capabilities. Here, the mass
collision mitigation. Vehicle types are characterized by of each vehicle is generated randomly and is used to
parameters such as vehicle mass, body length, dynam- determine both the vehicle body length and decelera-
ics of the braking system, and deceleration capabil- tion capability. The minimum mass used is Mmin =
ity. To verify the performance of the proposed CBC 1000 kg (passenger car), and the maximum mass is
strategy, three other control strategies, namely, Direct Mmax = 15000 kg (half-loaded heavy-duty truck). For
Brake Control (DBC), 2) Driver-Reaction based Brake a given number of vehicles N , the mass of each vehicle
Control (DRBC) and 3) Linear Quadratic Regulator is randomly generated as follows.
(LQR) control, are simulated for comparison with the
mi = Mmin + (Mmax Mmin) rand(1),
proposed strategy. Three strategies are defined as fol- (14)
lows. i = 1, 2, . . . , N.
J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC 179
Table 1
Then, the vehicle body length is determined as fol- Vehicle parameters used in the simulation
lows:
Id m (103 kg) amin (m/s2 ) L (m) THW Td (s)
mi Mmin 1 8.66 4.87 13.95 1.57 0.73 0.42
Li = 3.0(1)+23.0, = . (15) 2 2.38 6.12 4.97 1.63 0.63 0.24
Mmax
3 12.45 4.11 19.35 1.35 0.74 0.53
4 9.62 4.68 15.32 1.57 0.62 0.45
As such, the minimum body length is 3.0 m (pas- 5 11.99 4.20 18.71 1.49 0.77 0.51
senger car) and maximum body length is 23 m (heavy- 6 7.50 5.10 12.28 1.49 0.57 0.39
duty truck: tractor trailer combination). Li et al. de- 7 5.31 5.54 9.15 1.52 0.61 0.32
termined that the time constant for the braking system 8 14.23 3.75 21.90 1.58 0.70 0.58
9 7.43 5.11 12.18 1.65 0.56 0.38
of a typical heavy-duty truck is about 0.45 s [19]. For
simplicity, we assume that large vehicles have large
200
time constants for their braking systems. Hence, the
1
time constant for a vehicle braking system is given by 2
the following. 0 3
4
Distance(m)
mi Mmin 5
i = 0.2 (1 ) + 0.6, = , (16) -200 6
Mmax
7
8
The vehicle deceleration capability is determined by -400 9
the following expression.
mi -600
amin = 3.0 2.2 , i = 1, 2, . . . , N. 0 5 10 15 20
i
15000 Time(s)
400 3 4
-100
Distance(m)
4 5
300
5 -200 6
200 6 7
-300 8
100 7
9
8 -400
0 Crash point
9
-500
0 5 10 15
Time(s) -600
0 5 10 15 20
Dcceleration profiles Time(s)
6
(a)
1
5 200
Deceleration(m.s -2)
2
4 1
3 100
2
3 4
0 3
5 4
2 6 -100
Distance(m)
5
1 7 6
-200
8 7
0
9 -300 8
0 5 10 15 9
-400
Time(s)
Crash point
-500
Fig. 5. Nine vehicle braking force profiles and deceleration profiles
by CBC. -600
0 5 10 15 20
Time(s)
spectively. The braking force and deceleration profiles (b)
are shown in Fig. 5. The simulation results demonstrate 200
collision avoidance for all vehicles. From Figs 3 and 1
100
2
4, it can be seen that, after about 10 s, all the vehi- 3
0
cles stopped when employing the CBC strategy and all 4
-100
Distance(m)
Frequency
3 1000
Total Relative Kinetic Energy(MJ)
CBC 500
2.5 DBC 0
0 2000 4000 6000 8000 10000 12000 14000 16000
DRBC Mass(kg)
2 LQR
Frequency
1000
500
1.5
0
5 10 15 20
1 Length(m)
Frequency
1000
0.5 500
0
-6.5 -6 -5.5 -5 -4.5 -4 -3.5
0
0 5 10 15 20 25 Maximun deceleration(m.s -2)
Frequency
Time(s) 1000
500
Fig. 7. Comparison of the total relative kinetic energies produced by 0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
the four control strategies over the entire braking process. Time constant for braking system(s)
between vehicles 7 and 8 using the DBC strategy. Fig. 8. Vehicle attribute distributions for 1000 simulations.
From Table 1, we can see that vehicles 2 and 7 have
good deceleration capabilities (amin2 = 6.12 m/s2 100
and amin7 = 5.54 m/s2 ). Vehicles 2 and 7 stopped Successful Rate
Failure Rate
within 100 m and 110 m, respectively. However, vehi- 80
cles 3 and 8 have relatively poor deceleration capabil-
ities (amin3 = 4.10 m/s2 and amin8 = 3.75 m/s2 ).
Percentage(%)
60
Vehicle 3 requires over 150 m to stop fully and vehicle
7 requires over 170 m. The difference between these
40
two stop distances is larger than the car-following dis-
tance. Hence, collision occurred between these vehi-
cles. As shown in Fig. 6(b), two accidents between ve- 20
Table 2 2.5 20
Failure rates of the control strategies under the various simulation
Relative Speed(m.s-1)
15
Different Failure Failure rate in collision cases 1.5
strategies cases CBC DBC DRBC LQR 10
CBC 102 100% 73.5% 99.0% 99.0% 1
DBC 364 20.6% 100% 100% 65.1% 5
DRBC 881 11.5% 41.3% 100% 55.7% 0.5
LQR 527 19.17% 44.97% 93.17% 100%
0 0
Distance(m)
Energy(MJ)
2
of success (Figs 10(c) and (d)). As shown in Figs 10(c) 30
1.5
and (d), among the successful braking cases, the max- 20 1
imum relative kinetic energy for the proposed CBC
10 0.5
strategy is the smallest, and the minimum relative dis-
0
tance is the largest during the braking process relative 0
to the other three strategies. As such, when the colli- CBC DBC DRBC LQR CBC DBC DRBC LQR
sion is unavoidable, the collision impact for the CBC (c) (d)
strategy is the lowest (Figs 10(a) and (b)). Therefore, Fig. 10. Box plots of the statistical data of 1000 simulations using
we can claim that the proposed CBC strategy demon- the four control strategies. (a) Relative kinetic energy and (b) relative
strates the best performance for collision avoidance speed distributions in the case of failure. (c) Minimum relative dis-
among the four control strategies considered. On the tance and (d) maximum relative kinetic energy in the case of success.
Note that the number of failures is CBC: 102, DBC: 364, DRBC:
other hand, LQR control for CACC demonstrates good 881, and LQR: 527.
performance under normal conditions [26]. However,
LQR control provides degraded performance for colli- 16 16
12 12
As shown in Table 2, for the 102 failed simula-
tions using the CBC strategy, the failure rates of DBC,
8 8
DRBC, and LQR under equivalent conditions were
73.53%, 99.02%, and 99.02%, respectively. For the 4 4
364 failure cases using DBC, the failure rates of CBC,
DRBC, and LQR were 20.6%, 100%, and 65.11%, re- 0
0 5 10 15
0
0 5 10 15
spectively. The statistical data in Table 2 also verifies Leading Vehicle Mass(103kg) Leading vehicle mass(103kg)
that the proposed CBC strategy demonstrates the best (a) (b)
performance. 16 16
Following vehicle mass(103kg)
12 12
5.4. Discussions on unavoidable collision
8 8
Intuitively, a collision will occur between two vehi-
cles in the coupled group if a vehicle has good deceler- 4 4
Table 3
Success rate for collision avoidance under different mass distribution
The MPC is reduced to a constrained QP problem
ranges* for faster computation. The algorithm can be imple-
mented in a centralized or decentralized scheme. To be
CBC DBC DRBC LQR
Case 1 100% 98% 81.8% 75%
implemented as a centralized scheme, one vehicle in
Case 2 100% 99.6% 59.8% 83.4% the coupled group must act as the coordinator to calcu-
*Note. Case 1: Vehicle masses generated between 1000 kg and late the desired deceleration for each vehicle and com-
5000 kg; Case 2: Vehicle masses genrated between 10000 kg and municate it in real time. For a decentralized scheme,
15000 kg. Each case is based on 500 simulations. each vehicle must calculate and activate its desired de-
celeration independently. Three other control strate-
ficient distance between them, a collision may be un- gies, namely DBC, DRBC, and LQR, are built for fair
avoidable regardless of the control strategy employed. comparison with CBC. The findings include.
A possible solution for collision avoidance involv- (1) In the absence of a minimum deceleration limit
ing a large vehicle is to increase the car-following for the first vehicle and a maximum deceleration
distance or reduce the driving speed. Using the sim- limit for the last vehicle in the coupled group,
ulation results in the previous section, the distribu- it is likely that most collisions could be avoided
tions of collision vehicle types can be obtained (see using the proposed braking control. However,
Fig. 11). It can be observed that the majority of colli- in practice, the coupled group cannot be com-
sion cases occurred between a large vehicle and a small pletely isolated from other vehicles in front of
vehicle, and that the proposed CBC strategy can re- and behind the coupled group.
duce the collision probability between large and small (2) Intuitively, a system with less kinetic energy
vehicles. However, doing so may result in the colli- will have a reduced risk to safety and be more
sion between small vehicles because the collision im- stable. The control objective in the proposed
pact (relative kinetic energy) between small vehicles strategy is to minimize the total relative kinetic
is typically smaller than for the case when large ve- energy. The extensive simulation results verify
hicles are involved. Hence, from this viewpoint, the the effectiveness of the proposed strategy. The
proposed CBC strategy has good performance and en- reduced relative kinetic energy leads to reduced
hances overall safety. risk of collision.
The mass distribution (or deceleration capability (3) Heterogeneous braking capabilities are a ma-
distribution) in the coupled group has a significant in- jor cause of collisions. The multi-simulation re-
fluence on the success rate of collision avoidance. If sults show that unavoidable collisions often oc-
we alter the simulation parameter to make the vehicle cur among vehicles that follow a smaller leader
mass distribution follow a narrow range, thereby pro- in the coupled group, because large vehicles typ-
ically require a longer distance to stop relative
ducing a more homogeneous coupled group, the suc-
to small vehicles. A possible solution for colli-
cess rate of all four control strategies are substantially
sion avoidance is to increase the car-following
improved, as demonstrated in Table 3. In this case, the
distance or reduce the driving speed.
coupled group can be treated as one extended vehicle,
and collision is usually avoided. There are still a few open questions in the frame-
work of cooperative braking control. Future work will
include how to implement the proposed strategy with
passenger cars and/or trucks, and how to implement
6. Conclusions
wireless connection technology and enhance automatic
brake control capabilities for various environmental
This study considers the cooperative braking control
conditions.
(CBC) strategy for a group of vehicles with the pur-
pose of minimizing the risk of rear-end collision. The
upper level controller is designed using model predic- Acknowledgments
tive control (MPC) theory, in which the total relative
kinetic energy between consecutive pairs of vehicles This work was mainly supported by NSFC with
is used as the cost function. This cost function is able 51175290 and 51205228. The authors would like to
to generate the desired braking force for each vehicle thank the support of the Chinese National Programs
in the coupled group required to avoid collision or to for High Technology Research and Development with
minimize the total impact if collision is unavoidable. 2012AA111901.
184 J.-Q. Wang et al. / Longitudinal collision mitigation via coordinated braking of multiple vehicles using MPC
ligent Transportation Systems, IEEE Transactions on 4(3) nalized intersection capacity, Computer Aided Civil and In-
(2003), 143153. frastructure Engineering 25(6) (2010), 452467.
[37] D., E. Vega, Medina, R. Messeguer, D. Royo, F. Freitag, S.F. [41] W. Young, A. Sobhani, M.G. Lenne and M. Sarvi, Simulation
Ochoa and J.A. Pino, Sharing hardware resources in heteroge- of safety: A review of the state of the art in road safety simula-
neous computer-supported collaboration scenarios, Integrated tion modelling, Accident Analysis and Prevention 66 (2014),
Computer-Aided Engineering 20(1) (2013), 5977. 89103.
[38] J. Wang, D. Ni and K. Li, RFID-based vehicle positioning [42] Y. Zheng, J. Wang, X. Li, C. Yu, K. Kodaka and K. Li, Driv-
and its applications in connected vehicles, Sensors 14 (2014), ing risk assessment using cluster analysis based on natural-
42254238. istic driving data, in: IEEE 17th International Conference on
[39] J. Wang, D. Zhang, J. Liu, M. Lu and K. Li, Multi-objective Intelligent Transportation Systems, IEEE (2014), 25842589.
driving assistance system for intersection support, in: 13th [43] Y. Zheng, S. Eben Li, J. Wang, L. Wang and K. Li, Influence
International IEEE Conference on Intelligent Transportation of information flow topology on closed-loop stability of vehi-
Systems Madeira, (2010), 348353. cle platoon with rigid formation, 17th Intelligent Transporta-
[40] S.S. Washburn and C.O. CruzCasas, Impact of trucks on sig- tion System Conference, IEEE, (2014), 20942100.
Copyright of Integrated Computer-Aided Engineering is the property of IOS Press and its
content may not be copied or emailed to multiple sites or posted to a listserv without the
copyright holder's express written permission. However, users may print, download, or email
articles for individual use.