1

Cooperative Method of Traffic Signal Optimization

and Speed Control of Connected Vehicles at
Isolated Intersections∗
Biao Xu, Xuegang Jeff Ban, Yougang Bian, Wan Li, Jianqiang Wang, Shengbo Eben Li and Keqiang Li

Abstract— Signalized intersections play an important role in and idling. This further aggravates traffic jams and decreases
transportation efficiency and vehicle fuel economy in urban areas. the vehicle fuel economy.
This paper proposes a cooperative method of traffic signal control Researchers have made great efforts to improve trans-
and vehicle speed optimization for connected automated vehicles,
which optimizes the traffic signal timing and vehicles’ speed portation efficiency and vehicle fuel economy at signalized
trajectories at the same time. The method consists of two levels, intersections. In the transportation field, traffic signal con-
i.e., roadside traffic signal optimization and onboard vehicle trol/optimization has been an extensively studied area for
speed control. The former calculates the optimal traffic signal various types of signal control such as fixed-time and actuated
timing and vehicles’ arrival time to minimize the total travel time signal control [3]. More advanced methods, such as adaptive
of all vehicles; the latter optimizes the engine power and brake
force to minimize the fuel consumption of individual vehicles. The signal control, have also been proposed and deployed (e.g.,
enumeration method and the pseudospectral method are applied OPAC [4], RHODES [5], SCOOT [6], and Ren et al. [7]).
in roadside and onboard optimization, respectively. Simulation In the vehicle control field, vehicle engine control, startup
studies are conducted to compare the proposed method with control and speed control have been studied to improve the
benchmark methods. The results show significant improvement transportation efficiency and vehicle fuel economy. Chen et
of transportation efficiency and fuel economy by the cooperation
method. al. investigated the control strategy of engine start/stop system
that can shut off the engine automatically during idling to
Index Terms— Traffic Signal Optimization, Vehicle Speed Con- decrease fuel consumption [8]. Wang et al. proposed a startup
trol, Connected Automated Vehicle, V2I Cooperation, Intelligent
Transportation Systems, Travel Time, Fuel Consumption assistance system at signalized intersections to help vehicles
start up with less delay for improving the transportation
efficiency [9]. Li et al. demonstrated an eco-departure system
I. I NTRODUCTION that aims to optimize the speed profiles of the vehicle departure
operations at signalized intersections and to decrease the fuel
Urban traffic is critical to businesses, people’s daily life, consumption during this period [10].
and local economies, which also causes heavy congestion Recent technology developments on connected automated
and related energy use and emission issues. The US Energy vehicles (CAVs) have shown great potential to significantly
Information Administration reported that, of the total petro- improve the mobility, fuel economy, and safety of urban traffic
leum consumed in the US in 2016, vehicles accounted for [11], [12]. CAVs, especially the vehicle to vehicle (V2V) and
a staggering portion of about 50% [1]. Furthermore, Texas vehicle to infrastructure (V2I) communications and vehicle
Transportation Institute reported that, in 2014, traffic jams automation, also provide new opportunities for traffic signal
wasted over 3.1 billion gallons of fuel and 6.9 billion hours control and vehicle control at signalized intersections.
in urban areas in the US [2]. Vehicle fuel economy and First, with V2I and V2V communications, traffic signal
transportation efficiency are affected by many factors such as controllers can acquire the more exact positions and motion
road capacity and infrastructure design. Among these factors, information of approaching vehicles in real time, which can
signalized intersections in urban areas play an important role, be used for more effective traffic signal control. Lee et al. [13]
which could block traffic flow and cause vehicle deceleration obtained vehicles’ states from connected vehicles to estimate
the travel times that were used for arriving intersection control,
∗Research supported by National Natural Science Foundation of China
(NO. 51475254 and NO. 51625503), National Science and Technology Major which proved to improve the total delay and average speed of
Project (No. 2016ZX03002019), National Science Foundation Grant CMMI- vehicles compared to the actuated signal control. Priemer and
1719551, National Key R&D Program in China (No. 2016YFB0100906), Bernhard [14] designed a decentralized adaptive traffic signal
and the US Department of Transportation Tier 1 University of Transportation
Center (UTC) on Connected Cities for Smart Mobility towards Accessible & control method that estimated the queue length and traffic flow
Resilience Transportation (C2SMART). (corresponding author: Keqiang Li). using the V2I data. The simulation results showed that the
Biao Xu, Yougang Bian, Jianqiang Wang, Shengbo Eben Li and Keqiang proposed method performed better than TRANSYT-7F signal
Li are with the Department of Automotive Engineering, Tsinghua Univer-
sity, Beijing 100084, China; and Collaborative Innovation Center of Elec- control when the penetration exceeded 20%. Goodall [15] used
tric Vehicles in Beijing, Beijing 100084, China. (e-mail: {b-xu13, byg14} simulation methods to predict the queue length and delay after
@mails.tsinghua.edu.cn, {wjqlws, likq, lishbo}@tsinghua.edu.cn). obtaining vehicle position and speed information via V2I,
Xuegang Ban and Wan Li are with the Department of Civil and Environmen-
tal Engineering, University of Washington, Seattle, WA 98105, USA. (e-mail: which was then used to optimize traffic signal timing. This
{banx, wan5}@uw.edu). method was able to improve the traffic mobility compared
Digital Object Identifier: 10.1109/TITS.2018.2849029

1558-0016 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
to coordinated actuated signal control in different levels of the vehicle speed optimization is rule-based, which may not
saturation rates according to the simulation results. Zhao et lead to the optimal speed trajectories.
al. [16] developed a V2I-based signal timing optimization Besides traffic signal control and vehicle control, there are
method by considering individual vehicles’ fuel consumption other CAV-based traffic intersection management methods.
characteristics of which the simulation results suggested the Dresner and Stone [26] designed a reservation-based method
potential to improve vehicle fuel economy around the inter- to coordinate the motion of CAVs at unsignalized intersections.
section. Feng et al. [17] presented a bi-level adaptive signal Lee and Park [27] proposed a so-called Cooperative Vehicle
control algorithm using the connected vehicle data to minimize Intersection Control method to manipulate individual CAVs’
the vehicle delay and queue length, and the simulation results maneuvers to avoid vehicle collision at unsignalized intersec-
showed that it reduced total delay significantly under high tions. Tachet et al. [28] proposed a capacity-optimal Slot-based
penetration rates compared to the actuated control. Actually, Intersection management system for CAVs which allocates
the ability to acquire the exact vehicle motion information the vehicles’ access time with the consideration of safety
promotes estimation accuracy of the queue length, vehicle constraints. Yang et al. [29] proposed a method to optimize
travel time, and vehicle fuel consumption for traffic signal the departure sequence and trajectories of the vehicles at an
control, which results in better signal control performance. intersection which had potential to minimize the traffic delay
Secondly, an approaching vehicle can also obtain traffic for a mix-traffic environment with conventional, connected,
signal phases/timing and traffic conditions by V2I/V2V com- and automated vehicles. Most of these methods, by eliminating
munications in real time, through which the speed trajectory the traffic signal completely, can only work when CAVs reach
of the vehicle can be optimized and controlled to reduce fuel the 100% penetration, and no pedestrian or bicyclist is present
consumption or other desirable objectives. Such individual at the intersection. This, however, will not happen in the near
vehicle control can be more easily deployed on CAVs since the future. Additionally, they did not separate the conflicting traffic
recommended vehicle speed trajectory may be implemented flow, which needs high robustness requirement for vehicle
as part of the automation algorithm. Asadi and Vahidi [18] control to ensure traffic safety at intersections.
proposed a predictive cruise control method at signalized In this paper, we propose a cooperative method of traffic
intersections, which aimed to make vehicles cruise at the signal optimization and vehicle speed control which can simul-
pre-set speed and arrive at the green light timely with the taneously optimize traffic signal timing and vehicle speed tra-
minimal use of braking. Jin et al. [19] suggested a power-based jectories. The purpose is to improve transportation efficiency
optimal longitudinal control for ICE (Internal combustion and decrease CAV fuel consumption. The cooperation method
engine) vehicles to optimize the speed profiles considering consists of traffic optimization and vehicle optimal control.
the brake specific fuel consumption map, the traffic signal, Traffic optimization calculates the optimal traffic signal timing
and the road grade. Wu et al. [20] made effort to optimize the and vehicles’ arrival times cycle by cycle, according to the
speed profiles for electric vehicles on signalized arterials. Xu vehicles’ initial speed and position information. Here we
et al. [21] used the branch and bound algorithm to optimize apply the best practices in signal design, i.e., the dual-ring
vehicle speed profiles in adjacent signalized intersections. He diagram for signal design [3]. Such method has been proven
et al. [22] introduced a multi-stage optimal control method to be effective in ensuring safety (e.g., separating conflicting
that optimizes the vehicle speed considering the traffic sig- movements and providing proper clearance times), while at
nal and the queue length at intersections. HomChauduri et the same time maximizing the efficiency for vehicles passing
al. [23] developed a speed optimization method for a group signalized intersections. The vehicle optimal control is applied
of connected vehicles using decentralized model predictive to each individual vehicle to obtain the optimal trajectories
control, which was proved to be effective for vehicle platoon based on the optimal arrival time calculated in the traffic
going through signalized intersections. In Europe, eCoMove optimization to minimize the vehicle fuel consumption. Here
project [24] built a prototype system for energy efficiency we assume that all vehicles are CAV-equipped so that the
and mobility of vehicles which was able to provide driving optimized speed trajectory can be calculated and executed by
assistance with the human-machine interface to avoid vehicle the vehicles automatically without human involvement.
stops at traffic signals. The main contributions of this paper are as follows:
In summary, CAV-based traffic signal control has great (1) We develop a cooperative method for the simultaneous
potential to improve the transportation efficiency and fuel optimization of traffic signal timing (macro level) and vehicle
economy for all vehicles in the system (defined as the macro control (micro level), by considering two objectives: trans-
traffic level in this paper), while CAV-based vehicle speed portation efficiency and vehicle fuel consumption.
optimization can improve efficiency and reduce fuel use at the (2) We consider transportation efficiency at the macro signal
individual vehicle level (defined as the micro vehicle level). It timing control level and fuel economy at the micro vehicle
is expected that combining the control at the macro and micro control level. This also implies that the primary goal of the
levels can help further improve transportation efficiency and proposed method is to ensure the efficiency of all vehicles,
decrease fuel consumption. For this, Li et al. [25] developed a while at the same time to minimize vehicle fuel consumption.
joint optimization method for traffic signal timing and vehicle Note that these two goals are interrelated: better efficiency
speeds, which was shown to decrease the travel time in will usually lead to less fuel consumption. Such consideration
different traffic demands. However, they did not consider the also helps to decompose the method into two interactive
fuel consumption for the approaching vehicles. Additionally, components, which makes the cooperative method easier to
 3

construct and solve.

This paper is structured as follows. In Section II, we
define the problem studied in the paper. Section III presents
the methodology of traffic signal optimization and vehicles
speed control. Section IV shows the enumeration method for
solving the traffic optimization problem and the pseudospectral
method for solving the vehicle optimal control problem. The
simulation and results are presented in Section V. Finally,
Section VI gives some concluding remarks.

II. P ROBLEM S TATEMENT

We consider a typical 4-leg, 2-lane (in each direction)
signalized intersection (see Fig. 1). The two lanes in each
incoming approach of the intersection are a left-turning lane Fig. 1. Studied Scenario
and a through/right-turning lane, respectively. Here, we assume
that vehicles for different destinations are already on the
corresponding lanes so that we do not need to consider the few signal cycles, which essentially indicates the occurrence
lane changing behavior. In other words, vehicle lane changing of oversaturation due to heavy traffic demands. The design
has been done before they enter the study area (when they of the signal clearance time will make sure that the vehicles
approach the intersection) or after they leave the study area. As passing the stop line at the end of the effective green time can
a result, vehicles from the same approaching direction will also be cleared during the clearance time, which is standard signal
follow FIFO (first-in-first-out) and no vehicle overtaking will design practice [3].
occur. The study area is defined by the communication range In Fig. 2, we demonstrate the schematic diagram of the
of V2I as shown in Fig. 1, which is usually a few hundred dual-ring signal control design. The green lines represent
meters in radius. the effective green intervals of the traffic signal phases and
There are thus eight approaching traffic movements in the red lines represent the clearance time R. tNBL , tSBT ,
the studied intersection, after we combine the through and tWBL , tEBT , tSBL , tNBT , tEBL and tWBT denote the phase
right turn movements for the same direction. They are north- time of NBL, SBT, WBL, EBT, SBL, NBT, EBL, and WBT
bound left (NBL), northbound through (NBT), westbound left movement, respectively. tNS and tEW are the phase time of the
(WBL), westbound through (WBT), southbound left (SBL), northbound/southbound and eastbound/westbound movements,
southbound through (SBT), eastbound left (EBL) and east- respectively. C is the (fixed) cycle length.
bound through (EBT) movements. The purpose of the traffic In practice, left turn phases or phases of the minor streets
signal control is to coordinate those movements. For a vehicle, may be skipped. Hence, we use binary variables bNBL , bSBT ,
the movement it belongs to is denoted as L when L ∈ F = bWBL , bEBT , bSBL , bNBT , bEBL , and bWBT to denote if the
{NBL, NBT, WBL, WBT, SBL, SBT, EBL, EBT}. We number corresponding phases are skipped. Specifically, if b L = 0 (L ∈
the vehicles in movement L with 1∼ N L according to the F), then the phase for the movement L is skipped and the
vehicles’ distance to the stop line where N L represents the phase time t L = 0. When b L = 1 (L ∈ F), the phase for the
total vehicle number of traffic movement L. As a result, a movement L is reserved and the phase time t L is greater than
vehicle can be represented with a pair (L, i ) where i is the the minimum green time. Briefly, ∀L ∈ F, we have
serial number of the vehicle and i ≤ N L .
As aforementioned, we apply the dual-ring approach for (1 − b L ) t L = 0, (1)
signal design [30] with predetermined fixed cycle length C (to b L ∈ {0, 1} . (2)
readily extend the proposed model to consider the coordination
As discussed before, we assume that all vehicles are
of multiple signals in future research); see Fig. 2. The phase
equipped with positioning and V2V/V2I devices so that they
sequence of Ring 1 is NBL – SBT – WBL – EBT and the
can send location and movement information to the sig-
phase sequence of Ring 2 is SBL – NBT – EBL – WBT.
nal controller when they enter the communication range of
Each phase can be divided into effective green interval and
the intersection. Additionally, all approaching vehicles are
clearance time (i.e., yellow and all-red times). Clearance time
assumed to be automated vehicles so that vehicles can control
is also predetermined and fixed using standard procedures [3].
their speed, strictly follow the optimized speed profile, and
Notice that since a particular phase may be skipped, the actual
pass the intersection automatically. That is, we assume 100%
phase sequence after optimization may be different as that
penetration of CAV equipped vehicles in this research.
shown in Fig. 2.
We assume that vehicles can pass the stop lines during the
III. M ETHODOLOGY
effective green intervals of the traffic signal, once they appear
in the study area. Note that this does NOT mean that vehicles A. Overview of Cooperation
must pass the intersection during the same cycle they arrive The proposed cooperative method consists of two major
at the stop lines. They may pass the intersection in the next components: traffic optimization and vehicle control. As
 vehicle (L, i ) meets the first condition above, we have
(i)
τ L ∈ g L , ∀i ≤ N L , L ∈ F, (3)

where τ L(i) is the arrival time of the vehicle (L, i ), which is

the vehicle’s running time from the initial position to the stop
line, and g L is the union set of the traffic signal green interval
for the traffic movement L. Therefore Equation (3) means the
vehicle’s arrival time falls into the green intervals’ union set.
Given the phase time of NBL, SBT, WBL, EBT, SBL, NBT,
EBL and WBT as tNBL , tSBT , tWBL , tEBT , tSBL , tNBT , tEBL and
tWBT , the effective green interval at the beginning of the cycle
can be written as:
 K −1
Fig. 2. Signal phase sequence gNBL = [kC, kC + (tNBL − R) bNBL ] (4)
k=0
 K −1
gSBT = [kC + bNBL tNBL ,kC + bNBL tNBL
k=0
+ bSBT (tSBT − R)] (5)
 K −1
gWBL = [kC + bSBT tSBT + bNBL tNBL ,kC + bSBT tSBT
k=0
+ bNBL tNBL + bWBL (tWBL − R)] (6)
 K −1
gEBT = [kC + C − bEBT tEBT ,kC + C − bEBT R] (7)
k=0
 K −1
gSBL = [kC,kC + (tSBL − R) bSBL ] (8)
k=0
 K −1
gNBT = [kC + bSBL tSBL ,kC + bSBL tSBL
k=0
+ bNBT (tNBT − R)] (9)
 K −1
Fig. 3. Overview of the cooperation of traffic signal and vehicles
gEBL = [kC + bNBT tNBT + bSBL tSBL ,kC + bNBT tNBT
k=0
+ bSBL tSBL + bEBL (tEBL − R)] (10)
 K −1
shown in Fig. 3, the traffic optimization resides at the traffic gWBT = [kC + C − bWBT tWBT ,kC + C − bWBT R]
controller. It acquires the speed and position information k=0
(11)
from approaching vehicles, calculates the optimal traffic signal
timing, and plans the vehicles’ arrival times with the aim to where R refers to the clearance time, k means the cycle
decrease the total travel time of all vehicles at the macro level. number, and K denotes the total number of cycles considered
The optimized vehicles’ arrival times are then sent to each by the method and is pre-set. Here, we take (4) as an example
corresponding vehicle. The onboard vehicle control decides to explain the equations. As shown in (4), gNBL is a union
the optimal speed trajectory of the vehicle to optimize the set of several green intervals in K cycles, of which kC is
engine power and brake force by considering the constraint the start time of the green interval in the k th cycle and
of the arrival time to minimize the fuel consumption in the kC+(tNBL − R)bNBL is the end time of the green interval in the
whole trip of the vehicle at the micro vehicle level. k th cycle. It means that the vehicles may pass the intersection
It is worth noting that the traffic optimization and the vehicle at any green interval of the K cycles.
control are conducted with the rolling horizon procedure. For In addition, the allocated time of different traffic phases
the traffic optimization, it optimizes the traffic signal timing must meet the requirement of the dual-ring constraints:
and vehicles’ arrival time at the end of each signal cycle, which
means its control period is one signal cycle. The calculated bSBT tSBT + bNBL tNBL = bNBT tNBT + bSBL tSBL = tNS , (12)
optimal signal timing is adopted by the traffic signal for the
bEBT tEBT + bWBL tWBL = bWBT tWBT + bEBL tEBL = tEW , (13)
next cycle. For the vehicle control, the control period is much
shorter, e.g., 0.2s. tNS + tEW = C, (14)
tL − R ∈ [gmin , gmax ] ∪ {0} , L ∈ F. (15)
B. B. Traffic Optimization In (15), gmin means the minimum green time, and gmax
1) Traffic model: As all vehicles are assumed to be auto- means the maximum green time. Notice that a phase L may
mated vehicles, we use some basic rules to build the traffic be skipped, in which case t L is zero.
model in this paper. Specifically, when vehicles are approach- Since the signal timing is modeled via the dual-ring diagram
ing the intersection, they must (i) arrive at the stop line within in Fig. 2, constraints (1) – (15) ensure that movement conflicts
the green interval of the corresponding phase; and (ii) keep a will not occur while compatible movements may occur con-
safe time headway to avoid collisions. Mathematically, if the currently if needed.
 5

Fig. 4(b)), the minimum arrival time is


(i) (i)2 (i)
−v L + v L +amax d L v2 −v(i)2
τ L(i),min = , if d L(i) < max L .
2amax 2amax
(18)
It is obvious that the vehicle’s arrival time cannot be less
than its minimum arrival time (see constraints (19)). These
minimal arrival time constraints consider the vehicle kinematic
Fig. 4. Two cases for vehicle arriving with minimal time
limitation, and ensure the optimized arrival time satisfies the
vehicle kinematic constraints.
Additionally, to ensure safety, vehicle (L, i ) must keep a τ L(i) ≥ τ L(i),min , ∀i ≤ N L , L ∈ F. (19)
safe time headway to the front vehicle (L, i − 1):
3) Cost Function: In traffic signal control, minimizing delay
(i) (i−1) + is often used as the objective. In this paper, we use the total
τL − τL ≥ T H Wmin , ∀2 ≤ i ≤ N L , i ∈ N , L ∈ F,
(16) travel time (from the initial position to the stop line) of all
the approaching vehicles as the cost function for traffic signal
where T H Wmin denotes the minimum safe time headway optimization:
which we can set as 2s [31].   NL
Remark: In Traffic Optimization, headway constraint (16) (i) (i)
f τL = τL .
can only ensure the overall collision avoidance of two consec- L∈F i=1
utive vehicles, in particular at the stop line. Vehicle collision
may still occur before they get to the stop line. This will be 4) Optimization Model: The traffic optimization aims to
further considered in Vehicle Optimal Control with a safety optimize the phase time of the traffic signal and the arrival time
distance constraint imposed to ensure collision avoidance of of all the approaching vehicles. Based on the traffic model, the
the trajectories of two consecutive vehicles. vehicle model, and the cost function, the traffic optimization
2) Vehicle Kinematic Model: A vehicle is a complex non- model is as follows:

linear dynamic system with numerous nonlinear components ∗
tSBT ∗
, tNBL ∗
, tNBT ∗
, tSBL ∗
, tEBT ∗
, tWBL ∗
, tWBT ∗
, tEBL , τL(i)∗
such as the engine and the transmission. In this paper, we
NL
build a simple vehicle kinematic model to describe the vehicle  (i)
movement in the traffic optimization model since the model = arg min τL ,
needs to consider all approaching vehicles. In the vehicle con- L∈F i=1

trol model, a more detailed vehicle model will be developed subject to: constr ai nts : (1) ∼ (19), (20)
to more accurately control the speed of a single vehicle. ∗ , t∗ , t∗ , t∗ , t∗ , t∗ ∗ ∗ (i)∗
where tSBT NBL NBT SBL EBT WBL , tWBT , tEBL , and τ L
It is assumed that the vehicle tends to accelerate to the are the optimal signal timing and the vehicles’ optimal travel
desired speed and keep uniform motion when running through time.
the intersection. Based on this assumption, the vehicle model Note that when the total cycle number K in the optimization
can be written as: model is large enough, we can always find the feasible
arrival time for vehicles in the optimization model under

a (t) , if v = vt
ḋ = −v, v̇ = a, a = any circumstances. Hence, the optimization model is always
0, if v = vt ,
feasible.
where t, d, v, and a are the time and distance to the inter- Remark: The traffic optimization is conducted at the end
section, and the velocity and the acceleration of the vehicle of each traffic signal cycle. The vehicle speed and position
respectively. Denote vt the target speed; in this paper, we are regarded as input to the optimization model. Furthermore,
assume the maximum target speed is the road speed limit. we conduct the optimization in a “receding horizon” way.
Given the speed limit vmax and the maximum comfortable We first consider multiple cycles (K ) in one optimization
acceleration amax , we can calculate the minimum arrival time and a vehicle may pass the intersection within the K cycles.
τ L(i),min of the vehicle (L, i ) with the distance d L(i) and the The optimized signal timing is only applied for the next
velocity v(i) L . When a vehicle’s distance is far enough so that
cycle after each optimization. This method essentially is a
the vehicle can speed up to the speed limit before arriving at receding horizon optimization technique, which enhances both
the intersection (see Fig. 4(a)), the minimum arrival time is the global optimality and the model precision.

(i) 2
 
(i) (i)2
2amax d L + vmax −v L v2 −v C. Vehicle Optimal Control
(i) (i)
τ L ,min = , if d L ≥ max L .
2amax vmax 2amax 1) Vehicle Dynamic Model: In vehicle control, we use a
(17) longitudinal dynamic model as the vehicle model [18],
On the other hand, when the distance is not far enough (see ṡ = −v, (21)
 1 P 1 to accelerate to a higher speed (maybe the speed of free
v̇ = − Fb − C D Aρa v2 − mg f cos α − mg sin α ,
m v 2 traffic flow) after going through the intersection. During the
(22) acceleration, its engine power increases sharply, which results
where P is the engine power, Fb means the braking force, m is in lots of fuel consumption. Since the terminal speed of the
the vehicle’s mass, CD is the drag coefficient, A is the vehicle vehicle has great influence on the future fuel consumption, we
frontal area, ρa is the air density, f is the rolling resistance considered the cost function with two parts, i.e., the vehicle
coefficient of the tires, g is the gravitational acceleration, fuel consumption of the total trip from the initial position to
α is the road slope, v is the vehicle speed, and s is the the intersection and the future fuel consumption after leaving
vehicle displacement. Notice that (21) & (22) constitute a more the intersection (represented by the terminal speed). Hence,
detailed vehicle model compared with the vehicle model in the cost function [33] is:
 tf
traffic optimization. 2
k1 FC (t) dt − k2 v t f , (28)
We can adjust the vehicle speed by controlling either the 0
engine power P or the braking force Fb with the uncontrolled
where k1 and k2 are the weight coefficients of the fuel
one being zero. We then select the Virginia Tech Comprehen-
consumption and vehicle kinetic energy.
sive Power-Based Fuel Consumption Model (VT-CPFM) [32]
5) Vehicle Optimization Model: An optimal control problem
as the vehicle fuel consumption model, which is a power-
is built to optimize the vehicle engine power P and the
based model to calculate the fuel consumption. It is defined
brake force Fb . Combining the vehicle dynamics, the safety
as follows:
constraint, the terminal constraints and the cost function, we
α + α1 P + α2 P 2 , if P > 0

FC = 0 (23) construct the vehicle optimal control model as:
α0 , if P ≤ 0,  ∗ ∗
P , Fb
where Fc is the fuel consumption rate, and α0 , α1 , and α2 are  tf
2
the coefficients of the fuel consumption model. = arg min k1 FC (t) dt − k2 v t f ,
2) Safety Constraint: When a vehicle is approaching the 0
intersection, it must keep a safe distance to the front vehicle subject to: vehicle dynamics: (21) and (22),
to avoid front collision: fuel consumption model: (23),
safety constraint: (24),
s (t) − s p (t) ≥ dsafe , (24)
terminal constraints: (25), (26), and (27).
where sp is the distance of the preceding vehicle to the stop (29)
line and dsafe is the safe following distance. Here (24) ensures
the safe distance at any time, i.e., between the trajectories of As a summary of this section: by considering the total
the two vehicles. vehicle arrival time in the traffic optimization model and the
3) Terminal Constraints: Some terminal constraints includ- fuel consumption in the individual vehicle control model,
ing initial state constraints and the terminal state constraints the proposed cooperative method (i) simultaneously considers
should be met. traffic efficiency and fuel use of the intersection; and (ii)
Firstly, the initial state is determinate: considers efficiency with higher priority, i.e., the primary goal
is to ensure efficiency of all vehicles, while at the same time
s (0) = d0 , (25) to minimize vehicle fuel consumption. Such consideration also
v (0) = v0 , (26) decomposes the efficiency and energy objectives, which makes
the cooperative method easier to construct and solve.
where v0 is the initial velocity of the vehicle and d0 is the
initial distance to the stop line.
The vehicles must follow the time schedules planned by IV. S OLUTION M ETHODS
Traffic Optimization, so the terminal time of Vehicle Optimal A. Solution Method for Traffic Optimization
Control refers to arrival time in Traffic Optimization. There- The proposed traffic optimization model is not a standard
fore, nonlinear programming (NLP) problem because of the discrete
s t f = 0, (27) interval constraints in (4)∼(11). Thus, it is difficult to solve the
optimization model by standard NLP solvers. Here, we adopt
where tf is the fixed terminal time which is optimized in Traffic an enumeration method to achieve the minimal total travel
(i) (i)
Optimization as τ L , i.e. tf = τ L . time of the approaching vehicles. The detailed procedures of
4) Cost Function: Primarily, the fuel consumption of the the enumeration method are shown below:
whole trip is considered in the cost function in individual (i) We first discretize the traffic signal timing [tNBL , tSBT ,
vehicle control, while ensuring the arrival time constraint from tWBL , tEBT , tSBL , tNBT , tEBL , tWBT ] considering the constraints
the traffic optimization model. Besides, the terminal velocity (1), (2), (12), (13), (14), and (15), which leads to N signal
of the vehicle (which reflects the vehicle kinetic energy at the timing schemes. Here, N is the number of traffic signal
terminal time) has great influence on the fuel consumption timing schemes after we conduct the discretization. We give
after the vehicle passes the intersection. This is because when an example to show how to discretize the traffic signal timing.
a vehicle passes an intersection with a low speed, it has Assume C = 60s, gmin = 8s, gmax = 24s, R = 2s, and the
 7

Fig. 5. Vehicle state and control variables of ASC, VC, and CTV

discrete step of 2s. We first discretize tNBL as [0, 10, 12, 14, precision.
16, 18, 20, 22, 24, 26]. Secondly, for given tNBL (say 14), Taking a typical case with C = 80s, gmin = 8s, gmax = 30s,
we can discretize tSBT as [0, 10, 12, 14, 16, 18, 20, 22, 24, and the discrete step of 2s as an example, we conduct the
26]. Thirdly, for given tSBT (say 20), we discretize tSBL and calculation for 500 times, of which the results show that the
tNBT as [(10, 24), (12, 22), (14, 20), (16, 18), (18, 16), (20, average and maximum CPU times are 0.28s and 0.45s.
14), (22, 12), (24, 10)] so that they add up to 34s. Then, we
discretize tWBL and tEBT as [(10, 16), (12, 14), (14, 12), (16, B. Solution Method for Vehicle Control
10)]. Finally, we discretize tEBL and tWBT as [(10, 16), (12,
14), (14, 12), (16, 10)] so that they add up to 26s. After the The proposed vehicle control method is an optimal control
discrete procedures described above, we get 1092 (N) signal (OPC) problem with nonlinear objective function and kinetic
timing schemes. equations, which is challenging to solve. Since the 1950s, sev-
(ii) Next, we will calculate the total travel time for each eral theoretical developments, such as the variational method
signal timing scheme with the constraints of (3)∼(11) and and maximum principle, have been introduced to solve such
(16)∼(19). OPCs. These methods are generally demanding computation-
(iii) Finally, we select the optimal signal timing scheme with ally and highly dependent on the initial values. They also tend
the minimal total travel time. to have a small radius of convergence, which may result in
Notice that the discrete interval of the traffic signal timing suboptimal or infeasible solutions, especially for complex and
should not be too small. Otherwise, the number of traffic signal high-dimensional OPCs. To solve high-dimensional and com-
timing schemes N will be too large, which is computationally plex nonlinear OPCs, the shooting method, the dynamic inver-
demanding. In this sense, the enumeration method is a brute sion method, and the pseudospectral method have been pro-
force method that may lead to a huge computation load, espe- posed [34]. Among these methods, the pseudospectral method
cially when the traffic signal cycle length is long or the discrete is of higher computational efficiency and accuracy [35]. In
step of signal timing is small. However, if the discrete step of this study, we use the Legendre pseudospectral method to
the traffic signal timing is chosen properly, the enumeration solve the proposed OPC in (29). The method uses Legendre
method is able to calculate the approximate global optimum interpolating polynomial to approximate the state and control
of the traffic signal timing with fairly efficient computation. variables.
In this paper, we set the discrete step to 2s because: (1) the The steps of Legendre pseudospectral method are shown as
smaller discrete step leads to a huger calculation load; (2) step 1 ∼ step 6 [35]. The main idea of Legendre pseudospec-
2s is a common discrete step in traffic optimization which can tral method is to transform a continuous optimal control
also help us find a suboptimal solution. In a word, the discrete problem into a discrete combinational optimization problem
step to 2s is a balance of the computation load and the solution using a global interpolating polynomial. Step 1 and Step 2
try to find the discrete time, i.e., the collocation points for
the differential state equations and the integration objective where D j i denotes the element of the differential matrix D
function using Legendre polynomial. Step 3 ∼ Step 5 carry and
out the discretization of states and control variables in state ⎧
PN τ j
, if l = j
⎪
equations and the integration objective function. Additionally, ⎪
⎪
⎨ PN (τi ) τ j − τi
⎪
in Step 6, we eventually obtain the discrete NLP. D j i = −N (N + 1) /4, (40)
if l = j = 0
Step 1: time domain transformation. We transform the time ⎪
⎪
⎪ N (N + 1) /4, if l = j = N
domain t ∈ [0, tf ] to τ in the standard time interval [−1, 1] ⎪
⎩
0, otherwise.
as (30).
Then the state space equation can be transformed as the
2t
τ= − 1, τ ∈ [−1, 1] . (30) following constraints.
tf ⎡ ⎤
−v j
N
Step 2: collocations calculation. We use the Newton-
  ⎛ ⎞
 s tf ⎢ Pj 1
D ji i − ⎢ − Fbj − C D Aρa v2j − ⎠⎥
⎥
1
Raphson method to iterate and finally find the collocation vi 2 ⎣ ⎝ v j 2 ⎦ = 0,
points. Here, collocation points denote the discrete points for i=0 m mg f cos α − mg sin α
the state and control variables. The initial iteration value of the (41)
collocation points τ j,0 is selected as the roots of the N-order
Chebyshev-Gauss-Lobatto polynomial. where j = 0, 1, · · · , N.
Step 5: objective function transformation. We transform the
πj objective function using the Gauss-Lobatto quadrature rule to
τ j,0 = cos , j = 0, 1, · · · , N. (31)
N enhance the accuracy of transforming the integral:
For each collocation point τ j , we use Legendre polynomial t f N
J = k1 ω j FC P j , τ j − k2 v2N , (42)
to iteratively calculate the final collocations as (32), 2 j =1
where w j denotes the integration weight, defined as:
τ j,i PN τ j,i − PN−1 τ j,i
τ j,i+1 = τ j,i − (32) 2
N PN τ j,i ωj = . (43)
2
N (N + 1) PN τ j
where
Step 6: NLP transformation. Using the above steps, we can
P1 τ j,i = 1 (33) transform the proposed OPC to NLP problem as follows,
P2 τ j,i = τ j,i , (34)

P j∗ , Fbj
∗
(2n + 1) τ j,i Pn τ j,i − n Pn−1 τ j,i
Pn+1 τ j,i = . (35) t f N
n+1 = arg min k1 ω j FC P j , τ j − k2 v2N ,
2 j =1
When the iteration errors are small enough, we can finish N  
 s
the iteration and find the final N collocation points τ j ( j = subject to: D ji i
vi
0, 1, 2, · · · , N). i=0
⎡ ⎤
Step 3: states and control variables discretization. After we −v j
determine the collocation points, we can use the Lagrange tf ⎢ "
P
#
− ⎣ 1 v jj − Fbj − 12 C D Aρa v2j − ⎦ = 0,
⎥
interpolation as (36)∼(38) to approximate the state and the 2 m
mg f cos α − mg sin α
control variables using the value of the state and the control
variables at the collocation points. s j − s f j ≥ d0 ,
N s0 = 0,
x (τ ) ≈ X (τ ) = L i (τ ) X i , (36) v0 = 0,
i=0
N
u (τ ) ≈ U (τ ) = L i (τ ) Ui , (37) s N = d. (44)
i=0
N τ − τj We conducted the Legendre pseudospectral method for 500
L i (τ ) = . (38)
j =0, j =i τi − τ j times in cases with the collocation number of 20, the random
initial speed from 5m/s to 15m/s, the random distance from
In this paper, x = [s, v]T , u = [P, Fb ]T , X i =[si , vi ]T , and 500m to 800m, and random arrival time, of which the results
Ui =[Pi , Fbi ]T , where si , vi , Pi , and Fbi are the displacement, show that the average CPU time is 0.16s and the maximum
the speed, the engine power and the brake force respectively CPU time is 0.3s.
at the collocation points. The NLP problem in (44) is a nonlinear problem with a
Step 4: state space equation transformation. We transform non-convex objective and non-convex constraints. Generally,
the differential form of the state space equations into the this kind of problems is challenging to solve. Fortunately, our
algebraic form with the values of the state and the control vehicle control problem here has a relatively small dimension
variables at the collocation points, as (39). (in hundreds of variables and constraints), which can be solved
N N by standard NLP solvers such as SNOPT [36]. These solvers
ẋ τ j ≈ L̇ i τ j X i = D ji Xi always result in a local solution, which is acceptable for the
i=0 i=0
( j = 0, 1, · · · N ) , (39) vehicle control model here.
 9

TABLE I
ASC S IMULATION PARAMETERS

TABLE II
V EHICLE PARAMETERS

Fig. 6. Fuel consumption

V. S IMULATION AND R ESULTS

A. Simulation Setup
To verify the proposed cooperation method, we conducted
simulation with VISSIM and MATLAB. First, we imple-
mented the actuated signal control [3] (ASC) method in
VISSIM and considered it as one of the baseline algorithms.
Some parameters of the ASC simulation can be seen in
Fig. 7. Travel time
TABLE I. Meanwhile, the vehicles’ inputs and trajectories
were exported via the COM port of VISSIM. Secondly, we
used Equation (23) to calculate the fuel consumption of the
vehicles simulated in ASC with the recorded trajectories data.
TABLE II shows some vehicle parameters used in the fuel
consumption calculation which we obtained from a vehicle
model of CARSIM [37]. Thirdly, we conducted simulations
of the Cooperative method of Traffic signal optimization and
Vehicle control (CTV) using the same vehicle input data from
VISSIM. Then, we used the Vehicle Control algorithm (VC)
with fixed traffic signal timing as another baseline algorithm.
In this algorithm, vehicles will be firstly assigned the optimal
arrival times based on the fixed timing of traffic signal. Their
trajectories will be optimized by the vehicle control model in Fig. 8. Number of oversaturated vehicles
Section III. The same vehicle model as TABLE II was used
in CTV simulation and VC simulation.
Fig. 5(b) shows the speed profiles. We can see that the
vehicle with ASC runs with higher speed during 5s ∼ 40s,
B. Case Study and has to brake and to keep idling after 40s. The vehicles
We first show the results of one case study with the with CTV and VC slow down ahead and then go through
traffic demand of 300 vehicle/(hour·lane) in the northbound/ the intersection without any stop and idling. Furthermore, the
southbound flow and the eastbound/ westbound flow. The vehicle under VC reaches a lower speed because of the longer
communication ranges of CTV and VC are both 600m. The travel time compared with that under the CTV.
cycle length of CTV and VC is 60s. Fig. 5(c) shows engine power curves. The vehicle speed
Fig. 5 shows the trajectories, speed profiles, engine power with ASC fluctuates because of the car following model
curves, and brake force curves of one randomly selected used in VISSIM, which causes 0.9g more fuel consumption
vehicle from NBL with the CTV, VC and ASC control. according to our computation. Furthermore, the average engine
Fig. 5(a) shows the trajectories. The bold horizontal lines power of ASC during 5s ∼ 40s is around 4.5kW which is
indicate the effective green intervals of different algorithms. It higher than those of CTV and VC.
can be seen that the effective green interval of CTV is earlier Fig. 5(d) shows the brake force curves. We can see that the
than those of ASC and VC. This leads to the shorter travel ASC vehicle brakes at about 40s, while the CTV vehicle and
time compared with the ASC and CV algorithms. the VC vehicle pass the intersection without any braking.
Fig. 9. Simulation results of Group A

It is worth noticing that the traffic optimization of CTV was drops from 85g to about 80g, while fuel consumption keeps
conducted in 20s (shown as the vertical bold dashed lines in nearly unchangeable (about 76g) after the communication
Fig. 5), after which the vehicle travel times were re-optimized. range reaches over 800m. It means that we do not need to
It caused the saltation of the engine power curve and the speed extend the communication range to acquire better performance
profile of the CTV vehicle in Fig. 5. when the communication range is over 800m. Actually, in
Since vehicle fuel consumption depends on the specific the situation studies, when the communication range reaches
driving pattern (i.e., the trajectory or speed profile) of the 800m, the vehicles and traffic signal have enough time and
vehicle, we expect that the three algorithms will lead to very space to adjust their speed and timing, respectively. In this
different fuel consumption on the selected vehicle. The simu- case, increasing the communication range further does little
lation results confirm this. In particular, the vehicle consumes to improve the performance of the proposed algorithm.
less fuel under CTV because it slows down ahead to avoid From Fig. 6 and Fig. 7, we can see that the communication
braking and idling at the intersection compared with the ASC. range correlates with the cycle length. The best cycle length
Besides, the vehicle under CTV passes the intersection using for communication range of 400m, 600m, 800m, 1000m, and
shorter time because its green interval starts earlier compared 1200m are about 40s, 50s, 60s, 60s, and 60s, respectively.
with the VC. Specifically, the vehicle fuel consumption under Fig. 8 shows the number of oversaturated vehicles in the
CTV, VC, and ASC for the whole trip are 32.8g, 38.9, and scenarios with different cycle length and communication range
57.8g, respectively. Furthermore, the vehicle travel times of conditions. We can see from the figure that longer cycle length
CTV, VC, and ASC for the whole trip are 49.7s, 59.2s, and leads to fewer oversaturated vehicles. It is because the longer
79.5s, respectively. cycle length means longer green intervals that result in more
vehicles passing the intersection within a cycle length.
C. Impact of Cycle Length and Communication Range When the communication range is small, the travel time
within the communication range is short. In this case, shorter
The cycle length and the communication range are both cycle length can lead to more frequent adjustment of the
critical parameters for the proposed cooperative method. To traffic signal timing, which results in less fuel consumption
find out the algorithm performance with different cycle lengths and smaller travel time. In reverse, when the communication
and communication ranges, we conducted more simulations. range is large, the travel time within the communication range
The traffic demand of the northbound/ southbound movements becomes long. There is enough time to adjust the signal timing.
is 300 vehicle/(hour·lane) and that of the eastbound/westbound In this case, the short cycle length results in large loss time of
movements is 400 vehicle/ (hour·lane). The cycle length varies the traffic signal, which causes degraded performance. Based
from 30s to 90s and the communication range varies from on the results and analysis above, we can conclude that the
400m to 1200m. The fuel consumption and travel time in the wider the communication range, the longer the optimal traffic
whole trip were calculated. signal cycle length.
Fig. 6 shows the average fuel consumption of CTV in the
scenarios with different cycle length and communication range
conditions. Fig 7 shows the average travel time. It can be seen D. Impact of Traffic Demand
from these two figures that a wider range of communication To further test the proposed cooperative method in different
leads to a better performance of fuel economy and transporta- traffic conditions, we conducted simulations with different
tion efficiency in general. However, for a given cycle length, levels of traffic demands.
the fuel consumption and travel time decrease with the increase We set three groups of simulation cases:
of the communication range more sharply when the range is Group A: the traffic demands of each entrance lane are the
small and less so after the range reaches 600m. For example, same. It consists of four cases with the traffic demands of 100,
when the communication range is extended from 400m to 200, 300, and 400 vehicles/hour in each•lane, respectively.
600m with the cycle length of 60s, the fuel consumption Group B: the traffic demands of south/north-bound entrance
 11

Fig. 10. Simulation results of Group B

Fig. 11. Simulation results of Group C

lanes are different from those of east/west-bound entrance The performances of CTV, CV, and ASC in each group of
lanes. It consists of four cases with the south/north-bound simulation cases are shown in Fig. 9∼Fig. 11.
traffic demands of 50, 150, 250, and 350 vehicles/hour•in each From the Fig 9 ∼ 10, we can see that vehicles with all
lane, and the east/west-bound traffic demands of 150, 250, 350, three algorithms consume less fuel and spent shorter time
and 450 vehicles/hour•in each lane, respectively. for the trips in low traffic demands than those in heavy
Group C: the traffic demands of each entrance lane are traffic demands. Additionally, vehicles with the CTV algorithm
equal, and first increasing and then decreasing with time. always consumed less fuel and spent less time during the
The communication range of the three groups of simulation trips compared with VC and ASC. Additionally, the CTV
cases is set as 600m and the cycle length is 60s. The time and VC algorithms can significantly decrease the number of
of each simulation in Group A and B is three hours, and the vehicle stops, which results in short travel time and less fuel
total simulation time of Group C is six hours. We calculated consumption. Furthermore, as demands become heavier, the
the fuel consumption, travel time for the entire trip of 600m, improvements of VC/CTV over ASC decreases.
and the ratio of stopped vehicles. Here, the speed threshold for Fig 11(a) shows the total traffic demand variation with time
stop definition is set as 3 m/s, i.e., when the speed of a vehicle in the simulation of Group C. From the simulation results
is lower than 3 m/s, it is categorized as the stopped vehicle. in Fig 11(b), Fig 11(c) and Fig 11(d), we can also draw
conclusions that the proposed cooperative method can improve and human-driven vehicles may co-exist for a long time. In
the traffic efficiency and vehicle economy in the environment future work, the human-driven vehicles must be considered
of varying traffic demand. in the cooperative optimization method, and the algorithm
Note that the vehicles in ASC cannot predict the traffic performance with different CAV penetration should be studied.
signal phasing and timing while the vehicles in CTV and Secondly, the enumeration method for solving the traffic
VC can receive this information from V2X communication. optimization model is brute force, and may be computationally
Therefore, the vehicles in CTV and VC can adjust their speed demanding. Future research should focus on developing more
to avoid meeting the red light and stopping. Consequently, the rigorous methods to solve the problem. Thirdly, the current
stop rates of CTV and VC in the simulation is significantly method assumes a rather restrictive intersection geometry and
smaller than that of ASC. lane group scenarios (i.e, two lanes for each approaching
direction), which should be relaxed in future studies. Last
but not least, coordination of multiple intersections has the
VI. C ONCLUSION AND F UTURE W ORK
potential to further improve the transportation efficiency and
In this paper, we proposed a cooperative method for traffic vehicle fuel economy, which is worth studying in future
signal optimization and vehicle speed control to improve the research.
transportation efficiency and vehicle fuel economy both in
macro traffic level and micro vehicle level. First, we built a
R EFERENCES
mathematic model of cooperative optimization between traffic
signal and vehicle speed, which consists of traffic optimization [1] US Energy Information Administration, “Annual Energy Outlook 2017
and vehicle optimal control. In the roadside traffic optimiza- with Projections to 2050,” Jan. 2017.
tion, the traffic signal timing and vehicles’ arrival time were [2] D. Schrank, B. Eisele, T. Lomax, and J. Bark, “2015 Urban Mobility
Scorecard,” Texas A&M Transportation Institute, Aug. 2015.
optimized cycle by cycle with the aim to reduce the total [3] R. P. Roess, E. S. Prassas, and W. R. McShane, Traffic Engineering, 3rd
vehicle travel times. Then the vehicles’ optimal arrival times ed, Englewood Cliffs, NJ, USA: Prentice-Hall, 2004.
were sent to corresponding vehicles. In onboard vehicle opti- [4] N. H. Gartner, “Opac: a demand-responsive strategy for traffic signal
control,” Transp. Res. Rec., vol. 906, pp. 75-81, 1983.
mal control, the engine power, brake force, speed profiles and [5] P. Mirchandani, and L. Head, "A real-time traffic signal control system:
trajectories of an approaching vehicle were optimized, consid- architecture, algorithms, and analysis," Transp. Res. Part C: Emerging
ering the constraints of vehicle dynamics, arrival time, and car Techn., vol. 9, no. 6, pp. 415-432, 2001.
[6] D. I. Robertson, and R. D. Bretherton, "Optimizing networks of traffic
following safety with the aim to reduce fuel consumption of signals in real time: The SCOOT method," IEEE Trans. Veh. Technol.,
the vehicle. The enumeration method and the pseudospectral vol. 40, no. 1, pp. 11-15, 1991.
method were used to solve the traffic optimization problem [7] Y. Ren, Y. Wang, G. Yu, H. Liu, and L. Xiao, "An Adaptive Signal
Control Scheme to Prevent Intersection Traffic Blockage," IEEE Trans.
and vehicle optimal control problem, respectively. Finally, we Intell. Transp. Syst., pp. 1-10, 2016.
conducted numerical experiments in simulations to test the [8] H. Chen, C. Zuo, & Y. Yuan, “Control Strategy Research of Engine
proposed cooperative method, and compared the performance Smart Start/Stop System for a Micro Car,” SAE Technical Paper, 2013.
of the method with two benchmark algorithms with different [9] J. Wang, Y. Bian, B. Xu, H. Qin, M. Hu, and B. Huang, "V2I-
based startup assistance system at signalized intersections," Advances
traffic demands, cycle lengths, and V2I communication ranges. in Mechanical Engineering, vol. 7, no. 8, pp. 1-11, 2015.
The simulation results showed that the cooperative method [10] S. E. Li, S. Xu, X. Huang, B. Cheng, and H. Peng, "Eco-Departure
can effectively adjust the traffic signal timing according to the of Connected Vehicles With V2X Communication at Signalized Inter-
sections," IEEE Trans. Veh. Technol., vol. 64, no. 12, pp. 5439-5449,
real-time traffic condition and simultaneously produce optimal 2015.
vehicle trajectory/speed profiles, which results in the improved [11] S. E. Li, S. Xu, X. Huang, B. Cheng, & H. Peng, "Eco-Departure of
transportation efficiency and vehicle fuel economy. In addition, Connected Vehicles With V2X Communication at Signalized Intersec-
tions," IEEE Trans. Veh. Technol., vol. 64, no. 12, pp. 5439-5449, 2015.
the cycle length and communication range have significant [12] J. Wang, S. Li, Y. Zheng, & X. Lu, "Longitudinal collision mitigation via
effects on the performance of the algorithm. Specifically, a coordinated braking of multiple vehicles using model predictive control,"
wider communication range works better with a longer cycle Integrated Computer-Aided Engineering, vol. 22, pp. 171-185, 2015.
[13] J. Lee, B. Park, and I. Yun, "Cumulative Travel-Time Responsive
length, and the algorithm performs the best with the cycle Real-Time Intersection Control Algorithm in the Connected Vehicle
length of 60s and communication range of about 800m in the Environment," J. Transp. Engineering, vol. 139, no. 10, Oct. 2013.
studied scenarios. Furthermore, the simulations with different [14] C. Priemer, and F. Bernhard, "A decentralized adaptive traffic signal
control using V2I communication data," in 12th Int. IEEE Conf. Intell.
traffic demands demonstrated that the proposed algorithm Transp. Syst., St. Louis, 2009, pp. 765-770.
improves the transportation efficiency and fuel economy in [15] N. Goodall, B. Smith, and B. Park, "Traffic signal control with connected
the stable and varying traffic volumes. vehicles," Transp. Res. Rec., J. Transp. Res. Board, vol. 2381, pp. 65-72,
2013.
Note that the proposed cooperative method is a general
[16] J. Zhao, W. Li, J. Wang, X. Ban, “Dynamic Traffic Signal Timing
idea for the intersection management of CAVs and it can be Optimization Strategy Incorporating Various Vehicle Fuel Consumption
applied in not only the 4-leg and 2-lane intersections but also Characteristics,” IEEE Trans. Veh. Technol., vol. 65, no. 6, pp. 3874-
other kinds of intersections. To extend this method to other 3887.
[17] Y. Feng, K. L. Head, S. Khoshmagham, and M. Zamanipour, “A
intersections, we only need to re-formulate the green phase real-time adaptive signal control in a connected vehicle environment,”
constraints (4∼11) and the signal timing constraints (12∼14). Transp. Res. C: Emerging Techn., vol. 55, pp. 460-473, 2015.
The proposed method can be extended in a few important [18] B Asadi, and A Vahidi, “Predictive cruise control: Utilizing upcoming
traffic signal information for improving fuel economy and reducing trip
directions. First, the method assumes 100% penetration of time,”IEEE Trans. Control Syst. Technol., vol. 19, no 3, pp. 707-714,
CAVs. Before the complete deployment of CAVs, the CAVs Apr. 2011.
 13

[19] Q. Jin, G. Wu, K. Boriboonsomsin, and M. J. Barth, "Power-Based Biao Xu received the B.E. degree from Tsinghua 1
Optimal Longitudinal Control for a Connected Eco-Driving System," University, Beijing, China, in 2013. From 2016 to 2
IEEE Trans. Intell. Transp. Syst., vol. 17, no. 10, pp. 2900-2910, 2016. 2017, he was a visiting scholar with University 3
[20] X Wu, X He, G Yu, A Harmandayan, and Y Wang, “Energy-optimal of Washington, Seattle, WA, USA. He is currently 4
speed control for electric vehicles on signalized arterials,” IEEE Trans. working toward the Ph.D. degree in the Department 5
Intell. Transp. Syst., vol. 16, no 5, pp. 2786-2796, Apr. 2015. of Automotive Engineering, Tsinghua University, 6
[21] B. Xu, F. Zhang, J. Wang, and K. Li, "B&B Algorithm-Based Green Beijing, China. 7
Light Optimal Speed Advisory Applied to Contiguous Intersections " in His active research interests include connected 8
15th COTA Int. Conf. of Transp. Prof., Beijing, pp. 363-375, 2015. and automated vehicles, vehicle control, and V2I 9
[22] X. He, H. X. Liu, and X. Liu, "Optimal vehicle speed trajectory on cooperation. Mr. Xu was the recipient of Best Paper 10
a signalized arterial with consideration of queue," Transp. Res. C: Award in 14th Intelligent Transportation Systems 11
Emerging Techn., vol. 61, pp. 106-120, 2015. Asia-Pacific Forum in 2015, and Best Paper Award in 2017 IEEE Intelligent 12
[23] B. HomChaudhuri, A. Vahidi, and P. Pisu, "Fast Model Predictive Vehicle Symposium. 13
Control-Based Fuel Efficient Control Strategy for a Group of Connected
Vehicles in Urban Road Conditions," IEEE Trans. Control Syst. Technol.,
vol. 25, no. 2, pp. 760-767, 2017.
[24] ERTICO-ITS Europe, “eCoMove cooperative mobility systems and
services for energy efficiency,” 2013. URL http://www. ecomove-
project.eu/assets/Uploads/Publications/eCoMoveBrochuresecured.pdf.
[25] Z. Li, L. Elefteriadou, and S. Ranka, "Signal control optimization for Xuegang Jeff Ban received the B.E. and M.S. 14
automated vehicles at isolated signalized intersections," Transp. Res. degrees in automotive engineering from Tsinghua 15
Part C: Emerging Techn., vol. 49, pp. 1-18, 2014. University, Beijing, China, in 1997 and 2000, respec- 16
[26] K. Dresner and P. Stone, "A Multiagent Approach to Autonomous tively, and the M.S. degree in computer sciences and 17
Intersection Management," J. Artif. Intell. Res., vol. 31, pp. 591-656, the Ph.D. degree in transportation engineering from 18
2008. the University of Wisconsin –Madison, Madison, 19
[27] J. Lee and B. Park, "Development and Evaluation of a Cooperative WI, USA, in 2005. He is currently an Associate 20
Vehicle Intersection Control Algorithm Under the Connected Vehicles Professor with the Department of Civil and Envi- 21
Environment," IEEE Trans. Intell. Transp. Syst., vol. 13, no. 1, pp. 81- ronmental Engineering, University of Washington, 22
90, 2012. Seattle, WA, USA. 23
[28] R. Tachet, P. Santi, S. Sobolevsky, L. I. Reyes-Castro, E. Frazzoli, His research interests include transportation net- 24
D. Helbing, C. Rattl, “Revisiting Street Intersections Using Slot-Based work modeling and simulation, urban traffic system modeling and operations, 25
Systems,” PloS One, vol. 11, no. 3, pp. e0149607, 2016. intelligent transportation systems, and connected and automated vehicles. He 26
[29] K. Yang, S. I. Guler, and M. Menendez, "Isolated intersection control is the author of over 100 papers in peer-reviewed journals and conferences. Dr. 27
for various levels of vehicle technology: Conventional, connected, and Ban serves as an Associate Editor for the Journal of Intelligent Transportation 28
automated vehicles," Transp. Res. C: Emerging Techn., vol. 72, pp. 109- Systems, Networks, and Spatial Economics, as well as on the editorial 29
129, 2016. board of Transportation Research, Parts B and C. He was the recipient of 30
[30] Q. He, K. Head, and J. Ding, "Heuristic algorithm for priority traffic 2011 CAREER Award from the US National Science Foundation, the New 31
signal control," Transp. Res. Rec., J. Transp. Res. Board, vol. 2259, pp. Faculty Award from the Council of University Transportation Centers and the 32
1-7, 2011. American Road and Transportation Builders Association in 2012, and Best 33
[31] J. Wang, L. Zhang, D. Zhang, and K. Li, "An adaptive longitudinal Paper Award in 2017 IEEE Intelligent Vehicle Symposium. 34
driving assistance system based on driver characteristics," IEEE Trans.
Intell. Transp. Syst., vol. 14, no. 1, pp. 1-12, 2013.
[32] H. A. Rakha, K.Ahn, K. Moran, B. Saerens, and E. Van den Bulck,
“Virginia tech comprehensive power-based fuel consumption model:
model development and testing,” Transp. Res. D: Transp. Environ., vol.
16, no 7, pp. 492-503, Mar. 2012.
[33] H. Jiang, J. Hu, S. An, M. Wang, and B. B. Park, “Eco approaching at an Yougang Bian received the B.E. degree from 35
isolated signalized intersection under partially connected and automated Tsinghua University, Beijing, China, in 2014. He is 36
vehicles environment,” Transp. Res. C: Emerging Techn., vol. 79, pp. currently working toward the Ph.D. degree in the 37
290-307, 2017. Department of Automotive Engineering at Tsinghua 38
[34] G. Elnagar, M. A. Kazemi, and M. Razzaghi, “The pseudospectral University, Beijing, China. His research interests 39
Legendre method for discretizing optimal control problems,”IEEE Trans. include distributed control of dynamical systems 40
Automat. Control, vol. 40, no. 10, pp. 1793–1796, Oct. 1995. and its applications to connected and automated 41
[35] S. Xu, S. E. Li, K. Deng, S. Li, and B. Cheng, “A Unified Pseudospectral vehicles. 42
Computational Framework for Optimal Control of Road Vehicles,” Mr. Bian received the Excellent Undergraduate 43
IEEE/ASME Trans. Mechatronics, vol. 20, no 4, pp.1499-1510, Aug. Dissertation Award of Tsinghua University in 2014 44
2015. and the Best Paper Award at the 2017 IEEE Intelli- 45
[36] P. E.Gill, W. Murray, and M. A. Saunders, "SNOPT: An SQP algorithm gent Vehicles Symposium. 46
for large-scale constrained optimization," SIAM Rev., vol. 47, no. 1, pp.
99-131, Feb. 2005.
[37] http://www.carsim.com/

Wan Li received the B.S. degree in traffic engineer- 47

ing from Sun Yat-Sen University, Guangzhou, China, 48
in 2012 and the M.S. degree in civil engineering 49
from Louisiana State University, Baton Rouge, LA, 50
USA, in 2014. She is currently working toward the 51
Ph.D. degree in civil engineering from University of 52
Washington, Seattle, WA, USA. 53
Her current research interests include connected/ 54
autonomous vehicles and transportation big data 55
analytics. 56
57 Jianqiang Wang received the B. Tech. and
58 M.S. degrees from Jilin University of Technol-
59 ogy, Changchun, China, in 1994 and 1997, respec-
60 tively, and the Ph.D. degree from Jilin University,
61 Changchun, China, in 2002. He is currently a Profes-
62 sor with the Department of Automotive Engineering,
63 Tsinghua University, Beijing, China.
64 His active research interests include intelligent
65 vehicles, driving assistance systems, and driver
66 behavior. He is the author or coauthor of over 40
67 journal papers, and he is the co-holder of 30 patent
68 applications. Dr. Wang has engaged in over ten sponsored projects, and he
69 was the recipient of Best Paper Award in 2014 IEEE Intelligent Vehicle
70 Symposium, Best Paper Award in 14th ITS Asia Pacific Forum, Best Paper
71 Award in 2017 IEEE Intelligent Vehicle Symposium, Distinguished Young
72 Scientists of NSF China (2016), and New Century Excellent Talents (2008).

73 Shengbo Eben Li received the M.S. and Ph.D.

74 degrees from Tsinghua University in 2006 and 2009.
75 He worked at Stanford University, University of
76 Michigan, and University of California Berkeley. He
77 is currently the associate professor in Department of
78 Automotive Engineering at Tsinghua University.
79 His active research interests include intelligent and
80 connected vehicles, learning-based driver assistance,
81 distributed control and optimal estimation, electrified
82 powertrain management, etc. He is the author of
83 over 100 peer-reviewed journal/conference papers,
84 and the co-inventor of over 20 Chinese patents. He is now the IEEE
85 senior member, and serves as the TPC member of IEEE Intelligent Vehicle
86 Symposium, ISC member of FAST-zero 2017 in Japan, Associated editor
87 of IEEE ITSM and IEEE Trans ITS, etc. Dr. Li was the recipient of Best
88 Paper Award in 2014 IEEE Intelligent Transportation System Symposium,
89 Best Paper Award in 14th ITS Asia Pacific Forum, National Award for
90 Technological Invention in China (2013), Excellent Young Scholar of NSF
91 China (2016), Young Professorship of Changjiang Scholar Program (2016).

92 Keqiang Li received the B. Tech. degree from

93 Tsinghua University, Beijing, China, in 1985, and
94 the M.S. and Ph.D. degrees from Chongqing Univer-
95 sity, Chongqing, China, in 1988 and 1995, respec-
96 tively. He is a Professor in automotive engineering
97 with Tsinghua University.
98 His research interests include vehicle dynamics,
99 control for driver-assistance systems, and hybrid
100 electrical vehicles. He has authored over 90 papers
101 and is a co-inventor on 12 patents in China and
102 Japan. Prof. Li is a Senior Member of the Society of
103 Automotive Engineers of China and is on the editorial boards of International
104 Journal of Intelligent Transportation Systems Research and International Jour-
105 nal of Vehicle Autonomous Systems. He has the recipient of the Changjiang
106 Scholar Program Professor Award and of some awards from public agencies
107 and academic institutions of China.