Transportation Research Part B 112 (2018) 89–112

Contents lists available at ScienceDirect

Transportation Research Part B

journal homepage: www.elsevier.com/locate/trb

Integrated optimization of traffic signals and vehicle

trajectories at isolated urban intersections
Chunhui Yu a,c, Yiheng Feng b, Henry X. Liu b,c,∗, Wanjing Ma a, Xiaoguang Yang a
a
Key Laboratory of Road and Traffic Engineering of the Ministry of Education, Tongji University, 4800 Cao’an Road, Shanghai, PR China
b
University of Michigan Transportation Research Institute (UMTRI), 2901 Baxter Rd, Ann Arbor, MI, 48109, USA
c
Department of Civil and Environmental Engineering, University of Michigan, 2320 G.G. Brown, 2350 Hayward Street, Ann Arbor, MI, USA

a r t i c l e i n f o a b s t r a c t

Article history: Existing traffic signal control systems only allocate green time to different phases to
Received 20 July 2017 avoid conflicting vehicle movements. With advances in connected and automated vehi-
Revised 13 April 2018
cle (CAV) technologies, CAV trajectories not only provide more information than existing
Accepted 14 April 2018
infrastructure-based detection systems, but also can be controlled to further improve mo-
Available online 24 April 2018
bility and sustainability. This paper presents a mixed integer linear programming (MILP)
Keywords: model to optimize vehicle trajectories and traffic signals in a unified framework at iso-
Connected and automated vehicle lated signalized intersections in a CAV environment. A new planning horizon strategy is
Integrated optimization applied to conduct the optimization. All vehicle movements such as left-turning, right-
Vehicle trajectory planning turning and through are considered. Phase sequences, green start and duration of each
Mixed integer linear programming model phase, and cycle lengths are optimized together with vehicle lane-changing behaviors and
Isolated urban intersection vehicle arrival times for delay minimization. Vehicles are split into platoons and are guar-
anteed to pass through the intersection at desired speeds and avoid stops at stop bars.
Exact vehicle trajectories are determined based on optimized vehicle arrival times. For the
trajectory planning of platoon leading vehicles, an optimal control model is implemented
to minimize fuel consumption/emission. For following vehicles in a platoon, Newell’s car-
following model is applied. Simulation results validate the advantages of the proposed con-
trol method over vehicle-actuated control in terms of intersection capacity, vehicle delays,
and CO2 emissions. A sensitivity analysis is conducted to show the potential benefits of a
short minimum green duration as well as the impacts of no-changing zones on the opti-
mality of the proposed model.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction

With increasing traffic demand, vehicles suffer from severe traffic congestions, which cause environmental problems and
economic losses (Koonce et al., 2008). Intersections are usually considered the bottlenecks for traffic flows in urban trans-
portation networks. As reported in National Transportation Operations Coalition (NTOC) (2012), delays at traffic signals on
major roadways were estimated to be 295 million vehicle-hours. Improving traffic signal operations is supposed to have a

∗
Corresponding author at: Department of Civil and Environmental Engineering, University of Michigan, 2320 G.G. Brown, 2350 Hayward Street, Ann
Arbor, MI, USA.
E-mail addresses: 13ych@tongji.edu.cn (C. Yu), yhfeng@umich.edu (Y. Feng), henryliu@umich.edu (H.X. Liu), mawanjing@tongji.edu.cn (W. Ma),
yangxg@tongji.edu.cn (X. Yang).

https://doi.org/10.1016/j.trb.2018.04.007
0191-2615/© 2018 Elsevier Ltd. All rights reserved.
90 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

significant impact on the efficiency of transportation systems in dense urban areas, potentially more effective than any other
operational measures in the traffic engineering toolkit (NTOC, 2012).
Traditional traffic control strategies include fixed-time, vehicle-actuated, and adaptive signal control. Fixed-time control
utilizes historical traffic data to calibrate pre-set signal parameters such as phase sequences, cycle lengths, and green splits.
Vehicle-actuated and adaptive signal control apply infrastructure-based detectors to collect real-time traffic data and adjust
signal timings according to time-variant traffic demand. Abundant studies have been dedicated to traditional signal control
including stage-based (Allsop, 1981; Webster, 1958), group-based (phase-based) (Heydecker, 1992; Silcock, 1997), and lane-
based approaches (Wong and Heydecker, 2011; Wong and Wong, 2003a, b). In the stage-based approach, compatible traffic
movements are grouped and move together in the same time span, called a stage. Green time is then allocated to each
stage. In the group-based approach, there is no need to specify a stage structure. Green time is directly allocated to each
traffic movement. In the lane-based approach, signals can be optimized together with lane markings to further ameliorate
traffic condition.
Thanks to the advances of connected and autonomous vehicle (CAV) technologies, communications between vehicles
(V2V) and between vehicles and infrastructure (V2I) provide a new source of data for traffic control. This bilateral com-
munication renders it possible to convey traffic information (e.g., signal timings or speed advisory) from intersections to
vehicles in real time for trajectory planning as well as collect detailed vehicle trajectory data (e.g., speeds and locations) for
signal control. The new technologies have the potential to prompt new opportunities for traffic control in a more effective
way.
One research area tried to incorporate real-time CAV trajectory information into signal optimization, in addition to
infrastructure-based detector data. Starting with a simplified two-approach intersection, Ilgin Guler et al. (2014) proposed an
algorithm by combining information from CAVs and loop detectors to optimize the sequences of vehicles discharging from
the intersection to minimize vehicle delays and stops. The time a CAV entered the zone of interest and the distance from
the intersection were recorded to estimate virtual arrival time and queue length. Feng et al. (2015) applied the standard
North American National Electrical Manufacturing Association (NEMA) dual-ring, eight-phase structure for signal optimiza-
tion using CAV trajectories. The trajectory information of connected vehicles was also utilized to estimate the locations and
speeds of unequipped vehicles. The phase sequence and duration within each barrier group were formulated in a two-
level optimization problem. Dynamic programming (DP) was adopted to solve the problem. He et al. (2012) investigated a
platoon-based arterial signal optimization model for different travel modes on the basis of real-time platoon request data
and signal status. A mixed-integer linear program (MILP) was formulated to solve the problem with flexible cycle lengths
and offsets. In addition to delays in the objective function, other performance metrics have also been explored such as
weighted cumulative waiting time (WCWT) (Datesh et al., 2011) and cumulative travel time (CTT) (Lee and Park, 2012).
Another research area assumes that signal timings are fixed and known. Signal information is used to optimize CAV
trajectories by controlling vehicle speeds or acceleration rates to reduce fuel/energy consumption, emission, delay and so
on. Typically, an optimal control problem is formulated regarding locations and speeds as state variables and accelera-
tion\deceleration as control variables (Kamal et al., 2013). However, it is generally difficult to solve an optimal control
model efficiently with complex objective and constraint formulations. One direct approach is to discretize time and state
space (Miyatake et al., 2011). The optimal control problem is then converted into a multistage decision process so that DP
can be applied. Another effective approach is to divide a vehicle trajectory into several segments with constant accelera-
tion/deceleration to reduce computational burden. He et al. (2015) and Wu et al. (2015) built approximation models with
three trajectory segments. Vehicles were supposed to accelerate/decelerate to an optimal cruising speed; then travel at this
speed; and finally accelerate/decelerate to a final speed when passing through the intersection. Therefore, the optimal con-
trol problem was transformed into a non-linear optimization problem with much fewer decision variables. Nevertheless,
Wan et al. (2016) argued that the three-segment trajectory was impractical because it was uncomfortable to the occu-
pants and potentially disruptive to the traffic. Instead, a two-segment trajectory was more implementable although it might
be sub-optimal. A vehicle either accelerated with the maximum torque or decelerated with the engine off to a constant
speed and then cruised past the intersection. Further, Rakha and Kamalanathsharma (2011) and Kamalanathsharma and
Rakha (2013) took the downstream of an intersection, where a vehicle restored its speed, into consideration to include fuel
consumption objective in the formulation. Besides analytical approaches, Wang et al. (2014a) proposed a fast numerical so-
lution algorithm based on Pontryagin’s Minimum Principle (PMP) to solve an optimal control problem for different control
objectives. The algorithm was then extended to a corporative control problem (Wang et al., 2014b).
To the best of our knowledge, there are limited studies on the integrated optimization of traffic signals and vehicle
trajectories in a unified framework. Malakorn and Byungkyu (2010) developed a cooperative system where a vehicle tra-
jectory included an accelerating/decelerating segment and a cruising segment. Signals could switch dynamically between
serving major and minor streets. Optimal arrival time of each vehicle was identified, which then determined the accelera-
tion/deceleration rate within the first segment. In contrast, Li et al. (2014) assumed a generic optimal vehicle trajectory of
four segments. The first and third segments had constant acceleration/deceleration rates, while the second and the fourth
segments had constant speeds. Signals were optimized by enumerating all feasible signal plans. However, models of both
studies were implemented at simplified intersections with two single-lane through approaches, which is unrealistic, espe-
cially in an urban traffic network. The previous signal optimization algorithms can hardly handle complex phase structures,
and the optimality of trajectory segmentation was not guaranteed. Moreover, lane changing behaviors are not considered
due to simplified settings of intersection geometry.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 91

Fig. 1. An intersection with four arms.

Different from signal-based traffic control, novel “signal free” intersections were proposed where traffic signals were
removed on the basis of 100% fully CAVs. Typically, vehicle trajectories are organized at an intersection by slot-based algo-
rithms. A notable slot-based system consists of an intersection manager and vehicle agents (Dresner and Stone, 2004, 2007,
2008). An intersection is divided into a grid of several tiles which can only be occupied by one vehicle at one time. Vehicle
agents send requests to the intersection manager to reserve space-time within the intersection. The intersection manager
then determines whether to accept or reject the requests according to a control policy. Usually the “first come, first served”
(FCFS) rule is applied. Nevertheless, vehicles are not guaranteed to avoid stops at stop bars. To improve efficiency, planning
techniques were applied to optimize vehicle arrival times and speeds passing the intersection but the choice of vehicle tra-
jectories seemed arbitrary (Au and Stone, 2010). Further, a batch strategy with adaptive vehicle platooning was developed
for high traffic demand (Tachet et al., 2016). Actually, the slot-based control focuses on vehicle management within an in-
tersection area and does not consider vehicle trajectories before arrival of the intersection. For example, vehicle trajectories
are assumed to be fixed to track reservation distance. In essence, the slot-based control does not ensure the optimality in
terms of vehicle platooning.
To address these research deficiencies, this paper presents an MILP model to optimize traffic signals and vehicle trajec-
tories at isolated urban intersections in a unified framework. Phase sequences, green start and duration of each phase, and
cycle lengths are optimized together with vehicle lane-changing behaviors and vehicle arrival times in the MILP model. Ve-
hicles are guaranteed to pass through an intersection at desired speeds and avoid stops at stop bars. A new planning horizon
strategy is applied to conduct the optimization. Platoons in each lane are identified based on the optimization results. Exact
vehicle trajectories are then generated by optimal control models and car-following models. The trajectory of each platoon
leading vehicle is optimized by an optimal control model with the objective to minimize fuel consumption and emission.
Lower and upper bounds of arrival time constraints are imposed to the MILP model in order to generate feasible trajecto-
ries. Clearance time is considered to eliminate conflictions between incompatible vehicle movements within an intersection
area. The proposed MILP model is similar with slot-based “signal free” models if the minimum green time constraint is re-
moved. However, the proposed model optimizes vehicle trajectories at the upstream of an intersection instead of managing
trajectories within an intersection area.
The remainder of this paper is structured as follows. Section 2 presents the problem and the key notations. Section 3 gen-
erates the trajectories of vehicles by lane based on vehicle arrival times at stop bars as well as provides arrival time con-
straints for the MILP model. In this way, the trajectory generation of vehicles in one lane is equivalent to the determination
of arrival times. Section 4 introduces the MILP model in detail that optimize signals, vehicle arrival times, and vehicle lane
choices. Section 5 conducts numerical studies and sensitivity analysis. Finally, conclusions and recommendations are deliv-
ered in Section 6.

2. Notations and problem description

2.1. Notations

Before presenting the model, two types of notations including parameters and control variables are introduced.

2.1.1. Parameters
Several types of parameters are defined in this section including geometric parameters, traffic parameters, vehicle tra-
jectory parameters, and signal parameters. The geometric parameters which are based on the layout of an intersection are
assumed to be given. Fig. 1 shows the details of one arm of a typical four-arm intersection as an example. Vehicle trajec-
tories and signals are updated each time optimization is conducted, denoted as time t0 . Vehicles that pass a stop bar in
92 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

Table 1
Notations and parameters.

General notations

i: Arm index
k: Approach lane index in each arm
(i, j): Traffic flow from arm i to arm j, which is also called phase
ω: Vehicle index
t0 : Time when an optimization process starts, s
M: A sufficiently big number

Geometrical parameters

I: Set of arms at the intersection

ζki, j : 1, if lane k in arm i is used by traffic flow (i, j); 0, otherwise
Li : Control zone in arm i in which vehicle trajectories are to be controlled, m
Lip : No-changing zone in arm i in which vehicles keep their previously optimized trajectories, m
Ki : Set of approach lanes in arm i; Ki = {1, …, |Ki |}, where |Ki | is the size of Ki
Kω : Set of permitted lanes for vehicle ω, which are the dedicated lanes for the movement direction of vehicle ω
K0 : Set of lanes used by traffic flows having right of way at time t0 , K0 = {k|ζki, j = 1, k ∈ Ki , (i, j ) ∈ 0 }

Traffic parameters

: Set of
all traffic flows at the intersection
ic : Set of
incompatible traffic flow pairs that cannot have right of way at the same time
0 : Set of
compatible traffic flows that are receiving green at time t0
p : Set of
the traffic flows whose green ended before time t0
i : Set of
vehicles in arm i at time t0
ip : Set vehicles in the no-changing zone in arm i at time t0 , ip ⊂ i
of
ω : Set vehicles ahead of vehicle ω ∈ i at time t0 , whose permitted lane set overlaps with that of ω; that is,
of

ω = {ω |Kω ∩ Kω = ∅, xω0 < xω0 , ω ∈ i}
Vehicle trajectory parameters

xω0: Distance between vehicle ω and its ahead stop bar at time t0 , m
vω0 : Speed of vehicle ω at time t0 , m/s
δ  ωk : 1, if vehicle ω is in lane k at time t0 ; 0, otherwise
ω
t f : Previously optimized arrival time of vehicle ω, s
teω : Generating time of vehicle ω, s
tlω : Time of the previous lane changing behavior of vehicle ω, s
tlmin : Minimum time interval between two lane changing behaviors of a vehicle, s
vωf : Desired speed of vehicle ω passing through the intersection, m/s
vmax : Speed limit, m/s
τ ω: Time displacement of vehicle ω in Newell’s car-following model, s
dω : Space displacement of vehicle ω in Newell’s car-following model, m
aL : Absolute value of the maximum comfortable deceleration rate, m/s2
aU : Value of the maximum comfortable acceleration rate, m/s2
α1 : Weight of total vehicle delay in the objective

Signal parameters

N: Number of signal cycles in the signal optimization horizon

d: Tolerance of solution quality degradation, s
α2 : Weight of cycle lengths in the objective
ts : Start time of the signal optimization horizon, s
θi,pj : Green start for inactive traffic flow (i, j) ∈ p in the current/first cycle. Inactive flow means the green time for this traffic flow was
already terminated by time t0 , s
φi,pj : Green duration for inactive traffic flow (i, j) ∈ p in the current/first cycle, s
θi,0j : Green start for active traffic flow (i, j) ∈ 0 in the current/first cycle, s
gi,min
j
: Minimum green duration for traffic flow (i, j), s
π i,j, l, m : Clearance time between incompatible traffic flows, defined as the earliest time when traffic flow (l, m) has right of way after traffic
flow (i, j) terminates, s
zi,n,sj,l,m : Relative time between green starts measured from traffic flow (i, j) to (l, m) in nth cycle, s
zi,n,ej,l,m : Relative time between green ends measured from traffic flow (i, j) to (l, m) in nth cycle, s

the same lane in the same cycle are regarded as one platoon. The first vehicle in a platoon is the leading vehicle and the
remaining vehicles in the same platoon are following vehicles. Main parameters are summarized in Table 1.

2.1.2. Control variables

Two types of control variables are defined. Let V = (T, S) be the set of all variables, where T is the subset of vehicle
trajectory variables; and S is the subset of signal variables. The control variables are updated after each optimization process.
Main control variables are summarized in Table 2.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 93

Table 2
Control variables.

Vehicle trajectory variables

δkω : 1, if vehicle ω is in lane k; 0, otherwise
λω : 1, if vehicle ω maintains its previously optimized trajectory; 0, otherwise
γ ω: 0, if the trajectory of vehicle ω is not affected by its all preceding vehicles and is to be updated; 1, otherwise
μ :ω 0, if vehicle ω decides to change lanes; 1, otherwise
ω ,ω 
η : 0, if vehicles ω and ω are in the same lane; 1, otherwise
xω (t): Distance between vehicle ω and the ahead stop bar at time t, m
lω (t): Travelled distance by vehicle ω from time t0 to t, m
vω (t): Speed of vehicle ω at time t, m/s
aω (t): Acceleration/deceleration rate of vehicle ω at time t, m/s2
tω
f
: Arrival time of vehicle ω at the intersection, s
tω : Travel time of vehicle ω, s
βnω : 1, if vehicle ω passes through the intersection in nth cycle; 0, otherwise
Signal variables
θi,nj : Green start for traffic flow (i, j) in nth cycle, s
φi,n j : Green duration for traffic flow (i, j) in nth cycle, s
ni,k : Green start in lane k of arm i in nth cycle, s
ni,k : Green duration in lane k of arm i in nth cycle, s
ni, j,l,m : 1, if the green start of traffic flow (l, m) immediately follows that of traffic flow (i, j) in nth cycle; 0, otherwise
Cn : Cycle length of nth cycle, s

Fig. 2. Model framework.

2.2. Problem description

For a typical intersection, there are three vehicle movements (i.e., left-tuning, through, and right-turning) in each arm.
Each movement has a different desired speed to pass through the intersection for safety concerns. As shown in Fig. 1, in
each arm, the approach lane index is incremented from the left most lane. Each approach lane is dedicated to one vehicle
movement. The distance between a vehicle and the stop bar at time t is denoted as xω (t). Li is the control zone in arm i in
p
which vehicle trajectories can be optimized. Li is the no-changing zone in which vehicles keep their previously optimized
trajectories. That is, only the trajectories of vehicles that are outside the no-changing zone will be updated over time. The
no-changing zone is designed to reduce computational burden but at the cost of optimality.
The task is to integrate vehicle trajectory planning into traffic signal optimization in a unified framework to minimize
vehicle delays. The model framework is shown in Fig. 2. The trajectory planning serves for two purposes: 1) build the
relationships between vehicle arrival times in the same lane, which helps build the arrival time constraints (e.g., upper and
lower bounds) in the MILP model; 2) generate trajectories for vehicles in the same lane given their arrival times, which are
supposed to be optimized in the MILP model. Note that no exact vehicle trajectories or platooning are needed in building the
constraints of arrival times in the MILP model. The outputs of the optimization model include signals, vehicle lane choices,
and vehicle arrival times. Platoons, i.e., the vehicles that pass the stop bar in the same lane in the same cycle, are identified
based on the optimization results. However, only vehicle lane choices and vehicle arrival times are needed in planning exact
trajectories (i.e., acceleration, speed, and location profiles) for vehicles in each lane. Platooning is not necessarily needed
when generating exact vehicle trajectories.
94 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

3. Vehicle trajectory planning

This section builds the relationships between arrival times and trajectories for vehicles in the same lane. An optimal
control model is introduced to generate optimal trajectories for platoon leading vehicles with the objective of minimizing
fuel consumption/emission given arrival times. The Newell’s car-following models are used to capture the trajectories of
following vehicles. In this way, the identification of vehicle trajectories is equivalent to determining the arrival times. To
guarantee the feasibility of the optimal control models, this section also helps build the constraints of vehicle arrival times
for the MILP model.

3.1. Trajectories of platoon leading vehicles

Vehicles that pass through an intersection in the same lane in the same cycle are regarded as one platoon. The first
vehicle in a platoon is a platoon leading vehicle. Its trajectory is planned with the objective of reducing fuel consump-
tion/emission. The initial state at time t0 and the final state at the intersection (i.e., location and speed) of each platoon
leading vehicle ω are known. The trajectory control strategy of vehicle ω is formulated as the solution to the following
optimal control problem (P1). An optimal trajectory can be identified given the value of vehicle arrival time t ω
f
. It should be
bounded to guarantee the feasibility of P1.
P1:
 tω
f
min
ω
|aω (t )|dt (1)
a (t ) t0


l˙ω (t ) = vω (t )
(2)
v˙ ω (t ) = aω (t )

l ω (t0 ) = 0
(3)
vω (t0 ) = vω0
 ω ω
l t f = xω
  0 (4)
vω t ωf = vωf

0 ≤ vω (t ) ≤ vmax , t0 ≤ t ≤ t ω
f (5)

−aL ≤ aω (t ) ≤ aU , t0 ≤ t ≤ t ω
f
(6)
tω
f
fixed

where lω (t) is the distance travelled by vehicle ω from time t0 to time t. It is different from xω (t) that is the distance
between vehicle ω and the stop bar at time t. The integration over the absolute value of the acceleration rate is used as
the objective function in Eq. (1) for fuel consumption/emission minimization. Although fuel consumption/emission may be
related to both vehicle speeds and acceleration rates (Jiang et al., 2017; Wang et al., 2014a), Eq. (1) is used for simplification,
which is a common practice in optimal control theory Naidu, 2002). The use of this simplified objective function has also
been validated in Feng et al. (2018). Eq. (2) describes vehicle dynamics. Eqs. (3) and ((4) are the constraints of initial and
final vehicle states. Eqs. (5) and (6) are the bounds of vehicle speeds and acceleration rates. The solution to P1 is the
acceleration profile of leading vehicle ω.
Eq. (5) is the constraint of a state variable (i.e., vehicle speed) in P1, which makes the optimal control problem difficult
to solve. To address the problem, we classify the condition into two cases depending on whether the upper bound vmax in
Eq. (5) will be violated. Fig. 3 shows the trajectories in the two cases with minimum travel time. The possible maximum
speed in Fig. 3 is obtained by accelerating from the initial state (0, vω 0
) and decelerating to the final state (xω0
, vωf ). In
Fig. 3(a), the constraint Eq. (5) is never violated and it can be removed from the optimal control problem (P1). In Fig. 3(b),
the constraint Eq. (5) may be active and then a trajectory segment of cruising is included.
Case 1:
If the following condition is true, then the constraint on the upper bound of speed vmax in Eq. (5) can be released. That
means in the generated vehicle trajectory, the vehicle speed never reaches vmax .
 2  2
(vmax )2 − vω0 (vmax )2 − vωf
+ > xω
0 (7)
2aU 2aL
A simplified optimal control problem can be solved analytically. Only the final solutions are shown here. The readers can
refer to Feng et al. (2018) for detailed derivation.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 95

Fig. 3. Illustration of two cases: (a) Case 1, and (b) Case 2.

Denote vω ω
aL and vaU to be the lower and upper of vehicle speeds when the vehicle enters the control zone in each arm
as
  ω 2
vωaL = 2aL xω
0
+ vf (8)

  ω 2
 ω 2 vf
vωaU = v f − 2aU xω0 , if xω
0 ≤ 2aU (9)
0, otherwise
Vehicle ω is controllable when vω ω ω
aL ≥ v0 ≥ vaU which is supposed to be satisfied. According to the proposed control strategy,
vehicle ω will maintain controllable until arriving at the intersection.
Denote tLω and tUω as
⎛

 ω 2  2 ⎞
v1 − vω0 v1 − vωf 2aU aL xω + aL v0 + aU vωf
tLω = + ⎝v1 = 0
⎠ (10)
aU aL aL + aU

⎧   2
⎪  ω 2
⎪
⎪ ω vf
and vω 2xω
vωf aL
⎪
⎨+∞, if x0 > 2aU 0 < a −
0 L aU
⎛  ⎞
tUω =  2 (11)
⎪ vω0 − v1 vωf − v1 ⎝ aU (vω
0 )
2
+aL vωf −2aU aL xω
⎪
⎪ 0
⎠, otherwise
⎪
⎩ aL + aU v1 = aL +aU

where tLω and tUω are the lower and upper bounds of travel times to reach the intersection to make P1 feasible con-
sidering vehicle dynamics limits (refer to Feng et al. (2018) for detailed derivation). Note that vω ω
aU and tU are piecewise
ω ω
functions due to the constraint v (t) ≥ 0 in Eq. (5). Therefore, travel time t of the leading vehicle ω can vary within the
following bounds:
tLω ≤ t ω ≤ tUω (12)
The corresponding lower and upper bound of vehicle arrival times at the intersection are thus further specified as
t0 + tLω ≤ t ω ω ω
f = t0 + t ≤ t0 + tU (13)
The optimal trajectory of a leading vehicle is determined when tω or t ω
f
is identified (refer to Appendix A for details).
Case 2:
If the following condition holds, then the upper bound speed vmax may be reached in the generated trajectory:
 2  2
(vmax )2 − vω0 (vmax )2 − vωf
+ ≤ xω
0 (14)
2aU 2aL
The difference between Case 2 and Case 1 is the calculation of the minimum travel time, denoted as t  ω
L
instead:
t  ωL = t1 + t2 + t3 (15)
where
vmax − vω0
t1 = (16)
aU
96 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

Fig. 4. Possible trajectories of a following vehicle.

vmax − vωf
t2 = (17)
aL
 2
(vmax )2 − (vω0 ) (vmax )2 − vωf
2

xω
0 − 2aU
− 2aL
t3 = (18)
vmax
Eq. (16) is the acceleration time from speed vω0
to vmax with the maximum acceleration rate aU . Eq. (17) is the decel-
eration time from speed vmax to vωf with the maximum deceleration rate aL . Eq. (18) is the cruising time at speed vmax .
Therefore, the travel time tω of leading vehicle ω in Case 2 is bounded by
t  ωL ≤ t ω ≤ tUω (19)
where tUω is calculated in Eq. (11) in Case 1. The arrival time is thus bounded by:
t0 + t  ω ω ω ω
L ≤ t f = t0 + t ≤ t0 + tU (20)
The optimal trajectory can also be generated according to Appendix A when tω or t ω
f
is determined.

3.2. Trajectories of following vehicles

Generally, a following vehicle has two types of trajectories as shown in Fig. 4. If following vehicle ω is far away from its
preceding vehicle ω , its trajectory is not affected by its preceding vehicle (Type 1). Otherwise, the trajectory is presented
in Type 2, which is modeled by a car-following model, when following vehicle ω catches up with its preceding vehicle ω  .
Case 1: Trajectory Type 1
If vehicle ω follows trajectory Type 1, it is supposed to drive as fast as possible to minimize the delay. The trajectory is
the solution of the following optimal control problem:
P2:
 tω
f
min
ω
1dt = t ω
f − t0 (21)
a (t ) t0

l˙ω (t ) = vω (t )
(22)
v˙ ω (t ) = aω (t )

l ω (t0 ) = 0
(23)
vω (t0 ) = vω0
 ω ω
l t f = xω
  0 (24)
vω t ωf = vωf
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 97

0 ≤ vω (t ) ≤ vmax , t0 ≤ t ≤ t ω
f (25)

−aL ≤ aω (t ) ≤ aU , t0 ≤ t ≤ t ω
f
(26)
tω
f
fixed

The constraints of P2 are the same as those of P1. The objective Eq. (21) is the minimization of the travel time (i.e.,
delay). Actually, the solution to P2 is a special case of P1 in which the given arrival time t ω
f
is the lower bound. Therefore,
the optimal travel time tω can be expressed as
  2
(vmax )2 − (vω0 ) (vmax )2 − vωf
2

tLω , if + > xω
t ω = 2aU 2aL 0 (27)
t  ωL , otherwise
where tLω and t  ω
L
are calculated in Eqs. (10) and (15). The arrival time at the intersection is

tω
f = t0 + t
ω (28)

Case 2: Trajectory Type 2

For a vehicle ω following trajectory Type 2, Newell’s linear car-following model (Newell, 2002) is applied. It is described
as
 

xω (t + t − τ ω ) + dω ,
xω (t + t ) = max (29)
xω (t ) − xU

where t is the time step; xU is the upper bound of the travel distance considering the maximum speed and acceleration
rate:
⎧
⎨vω (t )t + 1 aU t 2 , if t ≤ t 
xU = 2 (30)
⎩vω t t  + 1 a t  2 + v t − t  , otherwise
() U max
2
where t = (vmax − vω (t))/aU . Therefore, the stationary time headway between a following vehicle and its preceding vehicle
is
dω
hω = τ ω + ω (31)
v f

The arrival time of following vehicle ω is related to that of its preceding vehicle ω :

tω ω
f = tf + h
ω (32)

3.3. Trajectory generation procedure

Sections 3.1 and 3.2 describe different types of vehicle trajectories for platoon leading vehicles and following vehicles
in the same lane given their arrival times at the intersection. However, a platoon leading vehicle may become a following
vehicle over time and vice versa. In our trajectory generation procedure, it is not necessary to know whether a vehicle is a
leading vehicle or a following vehicle. The trajectories of platoon leading vehicles and following vehicles can be generated
in one procedure. The only information needed is the distribution of vehicles across lanes and their arrival times optimized
from the MILP model, which will be introduced in the following section.
Given the optimized arrival time t ω f
of vehicle ω in one lane and the current time t0 (or the optimized travel time
t ω = t ωf − t0 ), the vehicle trajectory type is identified according to the following steps.

Step 1: If the arrival time remains the same (i.e., t ω

f
= t  ωf ), then vehicle ω maintains its planned trajectory that is opti-
mized in the previous optimization process and go to Step 4. Otherwise, go to the next step.

Step 2: If the arrival time of vehicle ω and its preceding vehicle ω satisfies t ω f
= tωf
+ hω , then vehicle ω is a following
vehicle and follows trajectory Type 2 in Section 3.2 and go to Step 4. Otherwise, go to the next step.
Step 3: Vehicle ω may be a platoon leading vehicle in Section 3.1 or a following vehicle that follows trajectory Type 1
in Section 3.2. The trajectories in both cases can be determined according to the solutions to P1. Readers can refer to
Appendix A for detailed solutions and Feng et al. (2018) for detailed derivation. Go to the next step.
Step 4: The planned trajectory of vehicle ω over the following time is generated.
98 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

4. Formulations of optimization model

This section presents the optimization model for vehicle arrival times and traffic signal parameters (i.e., phase sequences,
green starts, green durations, and cycle lengths) considering lane-changing behaviors. The objective is vehicle delay mini-
mization. Vehicle trajectory planning is integrated by optimizing vehicle arrival times and lane choices, which determines
vehicle delays. The optimization is conducted based on a planning horizon strategy. Basic assumptions, vehicle trajectory
constraints, signal constraints, and the objective function are introduced sequentially.

4.1. Assumptions

The following assumptions are made to formulate the optimization problem:

a) Vehicles are homogeneous. The desired speed when passing the intersection is only related to its movement direction.
b) All vehicles are controllable.
c) Vehicles and the signal controller can communicate information in real time within the control zone.
d) Vehicles are in permitted lanes when entering the control zone in each arm. The permitted lanes are the dedicated lanes
for a certain vehicle movement.
e) Lane changing behaviors are assumed to be completed instantly.

4.2. Vehicle trajectory constraints

Vehicle trajectory constraints deal with trajectory variables to address the choice of occupied lanes, lane changing behav-
iors, and vehicle arrival times at the intersection.

4.2.1. Permitted occupied lanes

Each vehicle in arm i occupies only one lane which is specified by

δkω = 1, ∀ω ∈ i ; i ∈ I (33)
k∈Ki

The occupied lane must be in the set of permitted lanes for vehicle ω:

δkω = 1, ∀ω ∈ i ; i ∈ I (34)
k∈Kω

4.2.2. Possible target lanes

It is assumed that each vehicle ω in each arm i can only select a lane adjacent to the current one as the target lane if
vehicle ω is contemplating changing lanes:
(
min k +1,|Ki | )
δkω = 1, ∀ω ∈ i ; i ∈ I (35)
k=max (k −1,1)

where k is the index of the currently occupied lane of vehicle ω at t0 , and it satisfies:
δkω = 1 (36)
Further, vehicle ω is supposed to maintain its current lane if no preceding vehicles exist:
  
1 − δkω ≤ I{ω |δ  ω =1, xω

< xω } ω , ∀ ω ∈ i ; i ∈ I (37)
k 0 0
ω  ∈ ω
where IA (x) is an indicator function. IA (x) = 1, if x ∈ A; IA (x) = 0, otherwise. The right-hand side of Eq. (37) is zero if vehicle
ω has no preceding vehicles. This makes δkω = 1 indicating that vehicle ω keeps its currently occupied lane.
On the other hand, vehicle ω is allowed to move to an adjacent lane k (i.e., δkω = 1) only if the safety distance between
vehicle ω and the vehicles in lane k is guaranteed:
  2 
 ω −
 ( vω0 )
2 vω0

 
xω
0 − x0
ω ω ω
d + v0 τ + 2aL − 2aL ≥ − 1 − δkω M
(38)

if xω ω
0 ≤ x0
  2 
 ω ω
 ω ω ω
vω0

( vω )
2  
x0 − x0 − d + v0 τ + 2aL
− 20aL ≥ − 1 − δkω M
(39)

if xω
0 > x0
ω

∀ω ∈ {ω |δ  ωk = 1, ω ∈
i }; ω ∈ i ; i ∈ I, where max (k − 1, 1) ≤ k ≤ min (k + 1, |Ki |) and k = k . The safety distance in
Eqs. (38) and (39) is calculated so that a collision can be avoided when both vehicles decelerate with the maximum decel-
eration rate.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 99

4.2.3. Lane changing behaviors

The lane changing decision variable μω of vehicle ω is constrained by
−M(1 − μω ) ≤ δ  ω ω ω
k − δk ≤ M (1 − μ ), ∀k ∈ Kω ; ω ∈ i ; i ∈ I (40)
 
μω ≥ M δ  ωk + δkω − 2 + 1, ∀k ∈ Kω ; ω ∈ i ; i ∈ I (41)

When vehicle ω changes lanes, there must be a lane k that satisfies δ  ω

k
= δkω , which means that μω = 0 by Eq. (40). On the
other hand, when vehicle ω maintains its original lane k (i.e., ∃k ∈ Kω that satisfies δ  ω
k
= δkω = 1), then μω = 1 by Eq. (41).
Further, the minimum time interval between two lane changing behaviors is set for safety concerns:
t0 − tlω − tlmin ≥ −Mμω , ∀ω ∈ i ; i ∈ I (42)
If t0 − tlω ≥ tlmin , μω can be either one (i.e., no lane changing) or zero (i.e., lane changing). Otherwise, μω can only be one.

4.2.4. Gap acceptance conditions for lane changing

A vehicle can move into a lane only when the spacing between this vehicle and those vehicles in the target lane is
sufficient. That is, after performing lane changing, the spacing between the current vehicle and its preceding and following
vehicles in the target lane should both be sufficient, even if the vehicle in the target lane is decelerating (Yeo et al., 2008):
 
xω ω ω ω
0 − x (t0 − τ ) ≥ d − Mη
ω,ω − Mμω , ∀ω  ∈  ; ω ∈  ; i ∈ I
ω i (43)
  
xω ω ω ω
0 − x (t0 − τ ) ≥ d − Mη
ω,ω − Mμω , ∀ω  ∈  ; ω ∈  ; i ∈ I
ω i (44)
 ω 2  
ω 2
 v0 v0 
xω ω ω ω ω
0 − x0 ≥ d + v0 τ + − − Mηω,ω − Mμω
2aL 2aL (45)
∀ω  ∈ ω ; ω ∈ i ; i ∈ I
 ω 2  
 2
ω
v0 vω0  
xω ω ω ω
0 − x0 ≥ d + v0 τ + − − Mηω,ω − Mμω
2aL 2aL (46)
∀ω  ∈ ω ; ω ∈ i ; i ∈ I

xω (t0 − τ ω ) is used in Eqs. (43) and (44) for consistence with Eq. (29) in the car-following model. Eqs. (43)–(46) are effective
 
when vehicles ω and ω are to be in the same lane (i.e., ηω,ω = 0) after either one changes lanes (i.e., μω = 0 or μω = 0).

ηω,ω is constrained by the following Eqs. (47) and (48):
   
δkω − ηω,ω ≤ δkω ≤ δkω + ηω,ω , ∀k ∈ Kω ∩ Kω ; ω ∈ ω ; ω ∈ i ; i ∈ I (47)
 
ηω,ω ≤ 2 − δkω − δkω , ∀k ∈ Kω ∩ Kω ; ω ∈ ω ; ω ∈ i ; i ∈ I (48)
 
When vehicles ω and ω are in the same lane (ηω,ω = 0), Eq. (47) guarantees that δkω = δkω , ∀k ∈ Kω ∩ Kω . On the other
 
hand, if there is a lane k satisfying δkω = δkω = 1, Eq. (48) guarantees that ηω,ω = 0. Therefore, Eqs. (47) and (48) make
 
ηω,ω = 0 equivalent to δkω = δkω , ∀k ∈ Kω ∩ Kω .

4.2.5. Lower and upper bounds of vehicle arrival times

Section 3 builds the relationship between vehicle trajectories and vehicle arrival times at a stop bar for platoon leading
vehicles and following vehicles in the same lane. To make the optimal control models feasible, upper and lower bounds of
vehicle arrival times are given. Note that a platoon leading vehicle may become a following vehicle over time and vice versa.
Instead of setting different travel time bounds for leading and following vehicles respectively, we formulate the bounds that
apply to all vehicles.
The lower bound is
 2  2
(vmax )2 − vω0 (vmax )2 − vωf
tω ≥ t0 + tLω − Mλω , if + > xω
f 2aU 2aL 0 (49)
∀ ω ∈ i −  p
i
;i ∈I
 2  2
(vmax )2 − vω0 (vmax )2 − vωf
tω ≥ t0 + tLω − Mλω , if + ≤ xω
f 2aU 2aL 0 (50)
∀ ω ∈ i −  p
i
;i ∈I
where tLω and tLω are calculated in Eqs. (10) and (15), respectively. Eqs. (49) and (50) indicate that the constraints of
minimum arrival time hold for both platoon leading vehicles and following vehicles. The lower bounds are calculated based
100 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

on the optimal control models with allowable acceleration rates. Newell’s linear car-following model is, however, used to
capture the trajectories of following vehicles. To take the inconsistence into consideration, λω is used in Eqs. (49) and (50).
λω = 1 if vehicle ω keeps its previously optimized trajectory. Then Eqs. (49) and (50) are inactive.
Additionally, vehicle ω cannot overtake its preceding vehicle in the same lane no matter it is a platoon leading vehicle
or a following vehicle:
 
tω ω ω
f ≥ t f + h − Mη
ω,ω , ∀ω  ∈  ; ω ∈  −  p ; i ∈ I
ω i i
(51)
The safety constraint Eq. (51) is borrowed from the car-following model Eq. (32). Eq. (51) is effective only when vehicles

ω and ω are to pass through the intersection in the same lane (i.e., ηω,ω = 0).
The upper bound of arrival time is:
tω ω ω
f ≤ t0 + tU + Mγ , ∀ω ∈ i − i ; i ∈ I
p
(52)
where tUω is calculated in Eq. (11). Similar with λω in Eqs. (49) and (50), γ ω is used in Eq. (52).
γ ω = 0 if the trajectory
of vehicle ω is not affected by all of the preceding vehicles and the trajectory is going to be updated. λω and γ ω are
constrained as
    
1 − ηω,ω − ρ ω,ω + Mλω ≥ γ ω , ∀ω ∈ i − ip ; i ∈ I (53)
ω  ∈ ω
  
tω ω ω
f − t f − h ≤ Mη
ω,ω + Mρ ω,ω , ∀ω  ∈  , ω ∈  −  p ; i ∈ I
ω i i
(54)

 
ρ ω,ω ≤ 1 − ηω,ω , ∀ω ∈ ω , ω ∈ i − ip ; i ∈ I (55)

−M(1 − λω ) ≤ t ω ω ω
f − t f ≤ M (1 − λ ), ∀ω ∈ i − i ; i ∈ I
p
(56)

where ρ ω ,ω 
is an introduced binary variable. Eqs. (54) and (55) ensure that ρ ω ,ω 
= 1 when the trajectory of vehicle ω is not
 
affected by its preceding vehicle ω (i.e., ηω,ω = 0 and t ωf
> tωf
+ hω ). Eqs. (53) and (56) indicate that the upper bound con-
  
straint (52) is only effective (γ ω = 0) if vehicle ω is not affected by all the preceding vehicles ( ω ∈ω (1 − ηω,ω − ρ ω,ω ) =
0) and its trajectory is going to be updated (λω = 0).

4.2.6. No-changing zones

Vehicles in no-changing zones keep their previously optimized trajectories. That is, the trajectories of these vehicles will
not be updated with varying traffic condition. The benefits of doing so are to reduce computational burden but at the cost
of optimality. The constraints are specified by
δkω = δ  ωk , ∀k ∈ Ki ; ω ∈ ip ; i ∈ I (57)

tω ω
f = t f , ∀ ω ∈ i ; i ∈ I
p
(58)
Eqs. (57) and (58) guarantee no lateral (i.e., lane-changing behaviors) and longitudinal trajectory (i.e., arrival time at stop
bars) changes, respectively. According to the trajectory generation procedure in Section 3.3, maintaining the previously op-
timized arrival time is equal to maintaining the previously planned trajectories.

4.3. Signal constraints

Signal constraints deal with signal parameters in an optimization horizon. Vehicle trajectory planning is integrated into
signal optimization with the aid of vehicle arrival times and lane choices. Phase sequences, green start and duration of each
phase, and cycle lengths are optimized.

4.3.1. Lane signal settings

If the traffic flow from arm i to arm j is permitted in several lanes (ζk = 1), then vehicles in these lanes must receive
i, j

identical signal indication:

   
−M 1 − ζki, j ≤ ni,k − θi,nj ≤ M 1 − ζki, j
(59)
∀n = 1, . . . , N; k ∈ Ki ; (i, j ) ∈ ; i ∈ I
   
−M 1 − ζki, j ≤ ni,k − φi,n j ≤ M 1 − ζki, j
(60)
∀n = 1, . . . , N; k ∈ Ki ; (i, j ) ∈ ; i ∈ I
where θi,nj and φi,n j are the green start and green duration for traffic flow (i, j) in the nth cycle; ni,k and ni,k are the green
start and green duration in lane k of arm i in the nth cycle. These constraints ensure that ni,k = θi,nj and ni,k = φi,n j , i.e.,
identical signal settings for all permitted lanes k.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 101

4.3.2. Green start time

Since phase sequences are not determined in advance, the green start time of each phase in every cycle is confined in
one cycle length:
n−1
 
n
 
ts + C n ≤ θi,nj ≤ ts + C n , ∀(i, j ) ∈ ; n = 1, . . . , N (61)
n =1 n =1

However, the green start time is known if the phase was already started or terminated before time t0 in the first cycle
as shown in Eqs. (62) and (63), respectively:
θi,nj = θi,0j , ∀n = 1; (i, j ) ∈ 0 (62)

θi,nj = θi,pj , ∀n = 1; (i, j ) ∈  p (63)

The green start times for the phases that will receive green after time t0 in the first cycle are constrained in Eq. (64):
θi,nj ≥ t0 , ∀n = 1; (i, j ) ∈  − 0 −  p (64)

4.3.3. Duration of green

For safety concerns, a minimum value of green duration is set in Eq. (65):
gi,min
j
≤ φi,n j ≤ C n , ∀(i, j ) ∈ ; n = 1, . . . , N (65)
The green durations for the phases whose greens were terminated before time t0 in the first cycle are fixed:
φi,n j = φi,pj , ∀n = 1; (i, j ) ∈  p (66)
The green durations for the phases that are green at time t0 in the first cycle are constrained:
φi,n j ≥ t0 − θi,nj , ∀n = 1; (i, j ) ∈ 0 (67)

4.3.4. Green end time

The proposed control strategies are able to generate compact vehicle platoons. At most one platoon can pass through the
intersection in one lane during one cycle. The green start and end times for each phase should be in the same cycle:
n−1
 
n
 
ts + C n ≤ θi,nj + φi,n j ≤ ts + C n , ∀(i, j ) ∈ ; n = 1, . . . , N (68)
n =1 n =1

4.3.5. Cycle length

Consider the start time ts of the optimization horizon and the current time t0 , the cycle length of the first cycle is
constrained in Eq. (69):
C n ≥ t0 − ts , n = 1 (69)

4.3.6. Phase sequence

Traffic flows (i, j) and (l, m) are incompatible if they cannot have right of way at the same time. The phase sequence that
avoids incompatible traffic flows in each cycle are constrained as:
ni, j,l,m + nl,m,i, j = 1, ∀((i, j ), (l, m ) ) ∈ ic ; n = 1, . . . , N (70)

4.3.7. Clearance time

For a pair of incompatible traffic flows, the following constraints are specified to ensure the clearance time:
 
θi,nj + φi,n j + πi, j,l,m ≤ θl,m
n
+ M 1 − ni, j,l,m
(71)
∀((i, j ), (l, m ) ) ∈ ic ; n = 1, . . . , N
 
θl,m
n
+ φl,mn
+ πl,m,i, j ≤ θi,nj + M 1 − nl,m,i, j
(72)
∀((i, j ), (l, m ) ) ∈ ic ; n = 1, . . . , N
If traffic flow (l, m) follows (i, j) in the nth cycle, i.e., ni, j,l,m = 1, constraint Eq. (71) is effective. Otherwise, constraint
Eq. (72) is effective.
Moreover, the clearance time between the traffic flows in subsequent cycles should also be guaranteed:
θi,nj + φi,n j + πi, j,l,m ≤ θl,m
n+1
, ∀(i, j ) ∈ ; (l, m ) ∈ ; n = 1, . . . , N − 1 (73)
Similarly, the clearance time between the first traffic flow in the first cycle and the last traffic flow in the Nth cycle should
also be guaranteed:

N
θi,Nj + φi,Nj + πi, j,l,m ≤ θl,m
1
+ C n , ∀(i, j ) ∈ ; (l, m ) ∈  (74)
n=1
102 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

4.3.8. Vehicle arrival times at stop bars

Due to vehicle trajectory planning, vehicles in the same lane are turned into compacted platoons to pass through the
intersection within green intervals [ni,k , ni,k + ni,k ]. Queues at stop bars are avoided for full utilization of intersection ca-
pacity. A binary variable βnω is defined to distribute vehicles to different cycles. Let βnω = 1 if vehicle ω is supposed to pass
through the intersection in the nth cycle; and βnω = 0, if in other cycles. The constraints of arrival time t ω f
are expressed as
 
ni,k − M 1 − δkω − M(1 − βnω ) ≤ t ωf
 
≤ ni,k + ni,k + M 1 − δkω + M(1 − βnω ) (75)
∀n = 1, . . . , N; k ∈ Kω ; ω ∈ i ; i ∈ I

Eq. (75) is effective when vehicle ω is supposed to pass the stop bar in lane k (δkω = 1) in the nth cycle (βnω = 1). Further,
vehicle ω can choose only one of the N cycles in the signal optimization horizon:


N
βnω = 1, ∀ω ∈ i ; i ∈ I (76)
n=1

4.3.9. Other signal constraints

In practice, there may be some constraints on the relative timing of green starts and ends for different traffic flows.
Define zi,n,sj,l,m and zi,n,ej,l,m as the relative time for green starts and ends measured from traffic flow (i, j) to (l, m) in each cycle,
respectively. The constraints are specified as

θi,nj + zi,n,sj,l,m = θl,m

n
, ∀(i, j ) ∈ ; (l, m ) ∈ ; n = 1, . . . , N (77)

θi,nj + φi,n j + zi,n,j,l,m

e
= θl,m
n
+ φl,m
n
, ∀(i, j ) ∈ ; (l, m ) ∈ ; n = 1, . . . , N (78)

These constraints are mainly used to make sure that two signal groups start or end simultaneously (i.e., zi,n,sj,l,m = 0 or
zi,n,ej,l,m = 0) for practical considerations.

4.4. Optimization models

The primary objective of the proposed optimization model is the minimization of vehicle travel delays. The travel delay
of each vehicle is defined as the difference between the actual travel time and the free flow travel time. It is noticed that
there may be multiple solutions that have the same total vehicle delay but with different cycle lengths. Since we have no
constraints of maximum cycle lengths or maximum phase green times to make the model more flexible, a secondary objec-
tive of cycle length minimization is added. It increases the frequency of switching right of way and potentially decreases the
delays of the incoming vehicles in the future. The proposed model is a hierarchical multi-objective optimization model. Total
vehicle delay is first minimized and then cycle lengths are minimized for the solutions with the same minimum delay. The
most common approach to such a model is the weighted sum method (Marler and Arora, 2004). As a result, the objective
function is formulated as
P3:

  Li
 
N
min α1 tω ω
f − te − + α2 Cn (79)
( )
V = T̄ ,S
i ∈ I ω ∈ i
vmax n=1

where T̄ is a subset of the trajectory variable set T in Table 2 (i.e., T̄ = T − {xω (t ), l ω (t ), vω (t ), aω (t )}); α 1 and α 2 are
weighting parameters and α 1  α 2 > 0. The constraints are Eqs. (33)–(78). t ω f
− teω is the actual travel time of vehicle ω.
ω ω
Note that t f − te also includes the waiting time before entering the control zone when there is a queue at the boundary
of a control zone. Li /vmax is the free flow travel time. The objection function and all constraints are linear. Therefore, the
optimization model (P3) is an MILP problem that can be solved by many existing algorithms or commercial solvers.
 
In the MILP model, the number of effective binary variables is i∈I ω∈i −ip (|Ki | + |ω | + N + 3). The number of

i∈I (|i − i | + 2N |Ki | ) + N (2|| + |ic | + 1). The number of effective trajectory con-
p
effective continuous variables is
 
straints is bounded by i∈I ω∈ − p (10 + |i | + 2|Kω | + (7 + |Ki |)|ω |). And the number of effective signal constraints is
   i i

i∈I (2N |||Ki | ) + ω∈i (1 + N |Kω | ) + (3N + 1 )|| + |0 | + | p | + (3N − 1 )|ic | + 3N || . The complexity of the MILP
2
i∈I
model mainly depends on the vehicle number in the control zone. The numbers of binary and continuous variables exhibit
quadratic and linear growth with respect to the vehicle number |i | in each arm in the worst cases, respectively. The con-
straint number shows quadratic growth with respect to |i | in the worst cases as well.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 103

Fig. 5. Illustration of planning horizon.

4.5. Impacts of weighting parameters on solution quality

Generally, P3 is not equivalent to the original hierarchical multi-objective optimization model. Intuitively, cycle length
minimization in P3 urges vehicles to pass through the intersection within less time and thus leads to less delay. However,
in theory, P3 may produce a solution with worse quality in terms of total vehicle delay. Denote d1 and C1n as the minimum
total vehicle delay and the length of the nth cycle given by the original hierarchical multi-objective optimization model.
Denote d2 and C2n as the total vehicle delay and the cycle length of any solution given by P3. We have


N 
N
α1 d1 + α2 C1n = α1 d2 + α2 C2n (80)
n=1 n=1
N n
N n
where d2 ≥ d1 and n=1 C2 ≤ n=1 C1 . Then we get
 

N 
N
α1 (d2 − d1 ) = α2 C1n − C2n (81)
n=1 n=1
N n
N n
If d2 = d1 and n=1 C1 = n=1 C2 , then the solution to P3 is also the solutions to the original model. Otherwise, we have
 
α2 N 
N
d2 − d1 = Cn − C n ≤ d (82)
α1 n=1 1 n=1 2
where d is the tolerance of solution quality degradation. To guarantee Eq. (82), α 1 /α 2 is expected to satisfy the following
Eq. (83) in a conservative way:
N
α1 n
n=1 C1
≥ (83)
α2 d

Although N n=1 C1 is unknown, its upper bound can be easily estimated by solving a special case of P3 with α 1 = 1 and
n

α 2 = 0, i.e., only minimizing total vehicle delay. The resultant signals are divided into two parts by the passing time of the
last vehicle. The 2nd part, i.e., the signals after the passing time of the last vehicle, may be unrealistic because of too large
cycle lengths. But it has no impacts on vehicle delays. The length of this part can then be recalculated based on minimum
N n
green durations. The upper bound is the total length of the 1st part and the 2nd part and is used to replace n=1 C1 in
Eq. (83) as the guide to the selection of α 1 and α 2 in each optimization.

4.6. Planning horizon procedure

Previous studies use time (in seconds) as the length of the planning horizon with a fixed number of cycles (Feng et al.,
2015), which is different in this study. The cycle number N in the planning horizon, as shown in Fig. 5, depends on the
number of vehicles considered in the optimization. Note that cycle lengths are optimized over time and, therefore, the
total time of the planning horizon may vary over time as well. Based on the constraints in Section 4.3.8, N needs to be
large enough so that all vehicles are planned to pass through the intersection in the N cycles. A smaller N may render the
optimization model infeasible while a larger N increases computational burden. As a result, we choose the smallest N that
makes the model feasible. Note that the choice of N has no impacts on solution optimality under the condition that the
MILP model is feasible. The planning horizon procedure is shown in Fig. 6, which follows:

Step 1: Initialize horizon start time ts = 0 and cycle number N = 1 at initial time t0 = 0.
Step 2: Collect information of vehicles in the control zone at time t0 .
Step 3: Solve the special case of the MILP model P3 with α 1 = 1 and α 2 = 0.
Step 4: If the model is infeasible, then update the cycle number N = N + 1 and go to Step 3. Otherwise, go to the next
step.
Step 5: Select α 1 and α 2 as explained in Section 4.5 so that α 1 /α 2 is large enough.
Step 6: Solve the MILP model P3.
104 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

Fig. 6. Planning horizon procedure.

Table 3
Basic traffic demand and volume/capacity ratios.

Traffic demand in pcu/h (v/c) To Arm

From Arm 1 2 3 4

1 – 200 (0.38) 400 (0.35) 100 (0.21)

2 150 (0.32) – 150 (0.29) 200 (0.35)
3 380 (0.33) 150 (0.32) – 180 (0.35)
4 100 (0.19) 200 (0.35) 100 (0.21) –

Step 7: Update the signals and the planned vehicle trajectories in the control zone according to the optimization results.
Step 8: Record the optimized length C1 of the first cycle.
Step 9: Update time t0 = t0 + t, where t is the time step.
Step 10: If t0 reaches the final simulation time, then end the process. Otherwise, go to the next step.
Step 11: If t0 ≥ ts + C1 , then update ts = ts + C1 . Go to Step 2.

5. Numerical examples

5.1. Experimental data

To evaluate the proposed models, a typical four-arm intersection with all directions of movements is applied. The lane
markings of the intersection are shown in Fig. 7. Right-turning vehicles are not controlled by traffic signals but they are
controlled to arrive at the intersection at a desired speed. The length Li of the control zone in each arm is 300 m which is
the reliable communication range of Dedicated Short-Range Communications (DSRC) (Emmelmann et al., 2010). The length
p
Li of the no-changing zone in each arm is 50 m.
The basic traffic demand and the volume/capacity (v/c) ratios are shown in Table 3. The v/c ratios are calculated with the
assumed green duration of 26 s for each phase and a cycle length of 120 s. The saturation flow in each lane is determined
by the time headway in Eq. (31). Vehicles are generated according to Poisson distribution, which is a common practice for
traffic control at isolated intersections (Jiang et al., 2017; Li et al., 2014). Speed limit vmax is 15 m/s. The desired speeds vωf of
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 105

Fig. 7. Lane markings of a four-arm intersection.

Table 4
Throughput.

Demand factor Throughput (veh)

Actuated Control CAV-based Control Increase (%)

0.6 289.60 291.20 0.55

1.0 446.00 449.20 0.72
2.0 686.00 691.50 0.80
3.0 764.60 824.80 7.87
4.0 769.70 922.10 19.80

left-turning, though, and right-turning vehicles passing through the intersection are 10 m/s, 13 m/s, and 8 m/s, respectively.
Vehicles enter the control zone at the speed of 13 m/s instead of the speed limit in order to test more cases in the trajec-
tory generation in Appendix A. The time displacement τ ω and space displacement dω in the car-following model are 0.9 s
(assuming quick reaction of CAVs) and 6 m. The minimum time interval between two lane changing behaviors of a vehicle is
5 s. The absolute values of the maximum comfortable acceleration and deceleration rates (aU and aL ) are 2 m/s2 and 4 m/s2 .
The clearance time π i,j, l, m between incompatible traffic flows (e.g., the traffic flow from arm 1 to arm 2 and the traffic
i, j
flow from arm 2 to arm 3) is 4 s. The minimum green time gmin is 6 s. The tolerance d of solution quality degradation is
3 s. The weighting parameters α 1 and α 2 in the objective function Eq. (79) are 300 and 1, which satisfy Eq. (83) in each
optimization. To investigate the environmental impacts, the CO2 emission model in Frey et al., (2002) is employed.
The optimization models are written in C# and solved using Gurobi 7.5.1 (Gurobi Optimization Inc., 2017). All the exper-
iments are performed in a desktop computer with an Intel 3.6 GHz CPU and 16 GB memory. An upper limit of 1.5 s is set for
real-time application. A sub-optimal solution produced by the solver will be accepted if the solving time exceeds the time
limit.

5.2. Results and discussions

Vehicle-actuated control is applied in the simulation as the benchmark for comparison with the proposed control
method, denoted as CAV-based control. In actuated control, the maximum green durations for arm 1 and arm 3 are 30 s,
and the maximum green duration for arm 2 and arm 4 are 20 s. The minimum green duration is 4 s which is the optimal
value by trial and error. The unit extension time is 2 s. The time of each simulation scenario is 1200 s and the simulation
time step is 1 s. Average values of throughput, vehicle delays, and CO2 emissions of ten different random seeds are recorded
and shown in Tables 4–6. Five levels of traffic demand are tested, which are the product of the basic demand and a de-
mand factor. The demand with the factors from 0.6 to 2.0 is under-saturated and the demand with factors 3.0 and 4.0 are
over-saturated. The under-saturated and over-saturated traffic condition is observed with actuated control.
Table 4 shows that CAV-based control improves intersection capacity. The throughput increase under CAV-based con-
trol is insignificant with under-saturated traffic demand. This is consistent with our intuition because demand is below
intersection capacity under both control methods. When the demand factor further increases to 3.0 and 4.0, intersection
capacity is reached under actuated control because the throughput almost remains the same. In contrast, the throughput
106 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

Table 5
Average delay.

Demand factor Average Delay (s/veh)

Actuated Control CAV-based Control Decrease (%)

0.6 14.65 8.62 41.16

1.0 15.84 11.33 28.47
2.0 18.32 13.59 25.82
3.0 61.71 15.40 75.04
4.0 98.28 16.34 83.37

Table 6
Average CO2 emissions.

Demand factor Average CO2 Emissions (g/veh)

Actuated Control CAV-based Control Decrease (%)

0.6 123.96 114.73 7.45

1.0 126.19 123.51 2.12
2.0 134.98 132.38 1.93
3.0 212.65 133.08 37.42
4.0 267.07 135.39 49.31

under CAV-based control keeps increasing noticeably. This indicates that intersection capacity under CAV-based control is
higher compared with actuated control.
Table 5 shows the significant decrease of vehicle delays when CAV-based control is applied, which can reach ∼40% under
low traffic demand and ∼80% under high demand. The benefits are mainly due to improved intersection capacity as well as
the more efficient use of green time at the intersection. Vehicle trajectories are optimized so that all vehicles pass through
the intersection at high desired speeds without stops. Thus, no vehicle queues are generated at stop bars, either. As a
result, the green start-up lost time is eliminated and more vehicles can pass the intersection during the same green interval
compared with actuated control. One interesting observation is that the delay reduction under CAV-based control decreases
first and then rises as demand increases. Under under-saturated demand, the benefits are more remarkable with lower
demand which indicates strong flow uncertainty. Under over-saturated demand, the delay under actuated control rises more
significantly as demand increases. Because the intersection capacity under CAV-based control is much higher than that under
actuated control.
Table 6 shows the comparison of CO2 emissions. The results are similar to those in Table 5. CAV-based control outper-
forms actuated control under both under- and over-saturated demand. This is intuitive because the trajectories of platoon
leading vehicles are planned with the aim of reducing fuel consumption/emission. Following vehicles also have smoother
trajectories since they do not make complete stops at stop bars. But the benefits decrease with increasing traffic volumes
with under-saturated demand. The reason is that vehicles are likely to accelerate or decelerate more frequently under CAV-
based control when traffic condition varies. As shown in Fig. 8, vehicles decelerate only once before arriving at the in-
tersection under actuated control while vehicles keep adjusting their speeds according to varying traffic condition under
CAV-based control. However, the benefits are significant under over-saturated demand. Because intersection capacity is im-
proved under CAV-based control and thus vehicles experience much less delays. This indicates that vehicles spend much
less time traversing the intersection. In this way, CO2 emissions are greatly reduced.

5.3. Sensitivity analysis

5.3.1. Minimum green time

Minimum green time is designed to guarantee safety. It may be reduced when every vehicle is controllable and the
pedestrian flow can be ignored (e.g., intersections with overpasses for pedestrians). In that case, the proposed model is
likely to bring more benefits. To this end, a sensitivity analysis of minimum green time is performed to explore the poten-
tial benefits. The experiments show that a smaller minimum green duration results in significantly smaller vehicle delays
and lower CO2 emissions under CAV-based control, as shown in Fig. 9. When the minimum green time decreases from 6 s
to 1 s, the CO2 emissions and average vehicle delay under CAV-based control both decrease notably compared with actuated
control. For example, the decreases of CO2 emissions and vehicle delays reach ∼16% and ∼70%, respectively, when the mini-
mum green duration is 1 s. This is reasonable because a smaller minimum green duration helps avoid green waste time due
to flow fluctuations, especially under low demand. A small green duration can be used for efficiency, similar with the con-
sideration of skipping a phase in some signalization approaches. In current practice, the minimum green duration is usually
set to at least 5 seconds. However, it can be reduced to a smaller value under CAV-based control since no driver reaction
time and human errors need to be considered.
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 107

Fig. 8. Trajectories of left-turning vehicles in arm 1 as an example: (a) actuated control, and (b) CAV-based control.

18 70

CO Emissions
16 2 65
Delay

14 60
Dec reas e of CO Emis s ions (%)

Decrease of Average Delay (%)

12 55

10 50
2

8 45

6 40

4 35

2 30

0 25
6 5 4 3 2 1
Minimum Green (s)

Fig. 9. Decrease of CO2 emissions and vehicle delays with decreasing minimum green under CAV-based control compared with actuated control. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

5.3.2. Length of no-changing zones

The purpose of setting no-changing zones is to reduce computational burden at the cost of optimality. Vehicles within
no-changing zones keep their previously planned trajectories. Thus, the number of vehicles considered in the proposed MILP
model is reduced. Fig. 10 shows the impacts of the length of no-changing zones on the performance of CAV-based control.
Four demand levels from low to high are tested with the demand factors from 1.0 to 4.0. Fig. 10(a) shows that the length of
no-changing zones has no significant impacts on the throughput at all demand levels. This is because the demand has not
reached the intersection capacity, as discussed in Section 5.2. Fig. 10(b) shows that vehicle delays increase with increasing
length of no-changing zones. The delay increase is more significant with a longer length of no-changing zones under high
demand (demand factors 3 and 4). Fig. 10(c) shows similar results to Fig. 10(b) but in terms of CO2 emissions. Therefore,
the setting of no-changing zones does not have much negative impact on the performance of the model if traffic demand is
not high and the zone length is not too long.
108 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

Fig. 10. Impacts of the length of no-changing zones: (a) throughput, (b) vehicle delays, and (c) CO2 emissions.

6. Conclusions and recommendations

This research develops an MILP model that optimizes vehicle trajectories and traffic signals in a unified framework at
isolated urban intersections with multiple phases and all directions of vehicle movements. Phase sequences, green start and
duration of each phase, and cycle lengths are optimized together with vehicle lane-changing behaviors and arrival times for
vehicle delay minimization. Vehicle trajectory planning is integrated into signal optimization via the lane choice and the
arrival time of each vehicle at stop bars. A planning horizon strategy is applied to conduct the optimization. Vehicles are
grouped into platoons and are guaranteed to pass through the intersection at desired speeds to avoid stops at stop bars. The
trajectories of vehicles in each lane are planned based on optimized vehicle arrival times. Simulation results show that the
proposed CAV-based control outperforms vehicle-actuated control in terms of intersection capacity, vehicle delay, and CO2
emissions. A sensitivity analysis is conducted to show the potential benefits of a small minimum green duration as well as
the impacts of no-changing zones on the optimality of the proposed control.
It is assumed that all vehicles are controllable which may restrict the implementation of the proposed model. To relax
this assumption for lower market penetration of CAVs, more constraints including the choice of platoon leading vehicles
and the consideration of queues at stop bars are necessary. A car-following model is used to capture following vehicle
trajectories, which provides the potential for extending the current work to mixed traffic condition. The proposed model lays
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 109

the foundation for implementing more complicated and realistic traffic conditions, where the un-controllable vehicles can
be described by the car-following model. For simplicity, homogeneous vehicles are assumed and vehicles in the same lane
have same desired speed upon arriving at the intersection. It is interesting to extend the model to heterogeneous vehicles,
for example, in the case with shared lanes. For real-world implementation, the uncertainties in vehicle driving behaviors
and traffic environment cannot be ignored. Furthermore, we plan to expand this model to optimize vehicle trajectories and
signal timings at a network level. The combination of the proposed model and the reservation-based models is also worth
investigating.

Acknowledgments

This research was partially funded by the US Department of Energy (EERE Award No. DE-EE0 0 07212) and China Scholar-
ship Council (No. 201506260033). The views presented in this paper are those of the authors alone.

Appendix A. Trajectory generation for leading vehicles

Given the optimized arrival time t ω f

of vehicle ω and the current time t0 (or the optimized travel time t ω = t ω f
− t0 ),
the optimal trajectory of ω is identified by solving P1. The solutions are among six scenarios. Please refer to Feng et al.
(2018) for detailed derivation.
Scenario 1: t ω = tLω
As shown in constraint Eq. (12), the minimum travel time of vehicle ω is tLω that is calculated in Eq. (10). This scenario
only happens in Case 1 in Section 3.1. When t ω = tLω , vehicle ω has to accelerate with aU to speed vω (t1ω ) at location
xω (t1ω ) at time t1ω and then it decelerates with − aL until reaching the intersection with speed vωf at time t ω
f
. t1ω , xω (t1ω ), and
vω (t1ω ) are calculated as
vωf − vω0 + aL t ω
t1ω = t0 + (84)
aU + aL

vω (t1ω ) = vω0 + aU (t1ω − t0 ) (85)

 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (86)
2
The notations are explained in Tables 1 and 2. Note that only the deceleration segment exists if vω
0
= vωf + aL (t ω
f
− t0 ).
Scenario 2: tLω < t ω < tOL
ω
ω is calculated as
The travel time tOL
⎧  ω 2
⎪
⎪ xω v0 − vωf
⎪
⎪ 0 ω ω
⎨ vω + 2vω aL , if v0 ≥ v f
ω =
tOL
0 0
 ω 2 (87)
⎪
⎪ xω v0 − vωf
⎪
⎪ 0
⎩ vω + 2vω a , otherwise
f f U

In this scenario, vehicle ω accelerates with aU to speed vω (t1ω ) at location xω (t1ω ) at time t1ω , maintains this speed until
time t2ω at location xω (t2ω ), and decelerates with − aL to speed vωf at the intersection at time t ω
f
.
Travel time (t1ω − t0 ) is the root of the following equation:
 
aL + aU vωf − vω0
aU t 2 − tω
f − t0 + aU t
2aL aL
  2 
vω0 − vωf  
+ xω
0 + − tω
f − t0 vω
0 =0 (88)
2aL

As t2ω ≥ t1ω , it can be proved that (t1ω − t0 ) takes the smaller root and t2ω is calculated as
 
vω0 + aU t1ω − t0 − vωf
t2ω = t ω
f − (89)
aL
vω (t1ω ), xω (t1ω ), and xω (t2ω ) are calculated as
vω (t1ω ) = vω (t2ω ) = vω0 + aU (t1ω − t0 ) (90)
 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (91)
2
110 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

xω (t2ω ) = xω (t1ω ) − vω (t1ω )(t2ω − t1ω ) (92)

Scenario 3: t ω = t ω OL
ω
(v −v ) ω ω 2
Scenario 3–1: vω ω = x0 + 0 f )
≥ vωf (tOL
0 vω
0
2vω a
0 L
Vehicle ω drives with the initial speed vω0
until time t1ω at location xω (t1ω ), and then decelerates with − aL to speed vωf at
the intersection at time t ω
f
. Note that if vω
0
= vωf , only the segment with the constant speed exists. t1ω , vω (t1ω ), and xω (t1ω )
are calculated as
vω0 − vωf
t1ω = t ω
f − (93)
aL

vω (t1ω ) = vω0 (94)

xω (t1ω ) = xω ω ω
0 − v0 (t1 − t0 ) (95)
ω (vω −vω )2
Scenario 3–2: vω ω = x0 + 0 f )
< vωf (tOL
0 vωf 2vω aU
f
Vehicle ω first accelerates with aU to speed vωf at location xω (t1ω ) at time t1ω , and then drives at the constant speed vωf
until arriving at the intersection at time t ω
f
. t1ω , vω (t1ω ), and xω (t1ω ) are calculated as

vωf − vω0
t1ω = t0 + (96)
aU

vω (t1ω ) = vωf (97)

 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (98)
2
ω < t ω < t ω
Scenario 4: tOL OU
ω is calculated as
The travel time tOU
⎧  ω 2
⎪
⎪ xω v0 − vωf
⎪
⎪ 0 ω ω
⎨ vω − 2vω aL , if v0 ≥ v f
ω
tOU =
f f
(99)
⎪  ω 2
⎪
⎪ x ω v − vωf
⎪
⎩ ω−0 0
, otherwise
v0 2vω
0 U
a

xω (vω −vω )2
Scenario 4–1: vω 0
≥ vωf (tOU
ω = 0 − 0 f )
vωf 2vω aL
f
In this scenario, vehicle ω decelerates with − aL to speed vω (t1ω ) at location xω (t1ω ) at time t1ω , maintains the constant
speed vω (t1ω ) until time t2ω at location xω (t2ω ), and decelerates again to speed vωf upon arriving at the intersection at time
tω
f
. Note that if vω
0
= vωf , only the segment with the constant speed exists. t1ω , vω (t1ω ), xω (t1ω ), t2ω , vω (t2ω ), and xω (t2ω ) are
calculated as
 ω 2  ω  2
2aL xω
0 −
v0 + v f
vω (t1ω ) = vω (t2ω ) =  ω   (100)
2 aL t f − t0 − vω
0
+ vωf
 
vω0 − v t1ω ω
t1ω = t0 + (101)
aL
 
vω t1ω − vωf
t2ω = t ω
f − (102)
aL
 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (103)
2

xω (t2ω ) = xω (t1ω ) − vω (t1ω )(t2ω − t1ω ) (104)

ω (vω −vω )2
Scenario 4–2: vω ω = x0 − 0 f )
< vωf (tOU
0 vω
0
2vω a
0 U
 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112 111

In this scenario, vehicle ω accelerates with aU to speed vω (t1ω ) at location xω (t1ω ) at time t1ω , maintains the constant
speed vω (t1ω ) until time t2ω at location xω (t2ω ), and accelerates again to speed vωf upon arriving at the intersection at time
tω
f
. t1ω , vω (t1ω ), xω (t1ω ), t2ω , vω (t2ω ), and xω (t2ω ) are calculated as
 ω 2  ω  2
2aU xω
0 +
v0 − v f
vω (t1ω ) = vω (t2ω ) =  ω   (105)
2 aU t f − t0 + vω
0
− vωf
 
vω t1ω − vω0
t1ω = t0 + (106)
aU
 
vωf − vω t1ω
t2ω = t ω
f − (107)
aU
 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (108)
2

xω (t2ω ) = xω (t1ω ) − vω (t1ω )(t2ω − t1ω ) (109)

Scenario 5: t ω = tOU
ω
(v −v ) ω ω ω 2
Scenario 5–1: vω ω = x0 − 0 f )
≥ vωf (tOU
0 vωf 2vω aL
f
In this scenario, vehicle ω first decelerates with − aL to speed vωf at location xω (t1ω ) at time t1ω , and then drives at the
constant speed vωf until arriving at the intersection at time t ω
f
. Note that if vω
0
= vωf , only the segment with the constant
speed exists. t1ω , vω (t1ω ), and xω (t1ω ) are calculated as
vω0 − vωf
t1ω = t0 + (110)
aL

vω (t1ω ) = vωf (111)

 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (112)
2
(v −v ) ω ω ω 2
Scenario 5–2: vω ω = x0 − 0 f )
< vωf (tOU
0 vω
0
2vω a
0 U
In this scenario, vehicle ω drives at the initial speed vω0
until time t1ω at location xω (t1ω ), and then accelerates with aU to
speed v f upon arriving at the intersection at time t f . t1 , vω (t1ω ), and xω (t1ω ) are calculated as
ω ω ω

vωf − vω0
t1ω = t ω
f − (113)
aU

vω (t1ω ) = vω0 (114)

xω (t1ω ) = xω ω ω
0 − v0 (t1 − t0 ) (115)
Scenario 6: t ω > tOU ω
The travel time tOU ω is calculated in the Eq. (99). In this scenario, vehicle ω decelerates with − a to speed vω (t ω ) at
L 1
location xω (t1ω ) at time t1ω , maintains this speed until time t2ω at location xω (t2ω ), and accelerates with aU to speed vωf upon
arriving at the intersection at time t ω f
.
Travel time (t1ω − t0 ) is the root of the following equation:
aU + aL
v − v  
− tω
tf 0
aL t 2 + f − t0 aL t
2aU aU
   
  vω − v0 ω 2
+ vω0 t ωf − t0 + f − xω
0 =0 (116)
2aU

As t2ω ≥ t1ω , it can be proved that (t1ω − t0 ) takes the smaller root and t2ω is calculated as
 
vωf − vω0 + aL t1ω − t0
t2ω = t ω
f − (117)
aU
112 C. Yu et al. / Transportation Research Part B 112 (2018) 89–112

vω (t1ω ) = vω (t2ω ) = vω0 − aL (t1ω − t0 ) (118)

 
vω0 + vω t1ω
xω (t1ω ) = xω
0 − (t1ω − t0 ) (119)
2

xω (t2ω ) = xω (t1ω ) − vω (t1ω )(t2ω − t1ω ) (120)

References

Allsop, R.E., 1981. Computer Program Sigset For Calculating Delay-Minimising Traffic Signal Timings – Description and Manual for Users. University College
London, Transport Studies Group, London.
Au, T.C., Stone, P., 2010. Motion planning algorithms for autonomous intersection management bridging the gap between task and motion planning. 2010
AAAI Workshops.
Datesh, J., Scherer, W.T., Smith, B.L., 2011. Using K-means clustering to improve traffic signal efficacy in an intellidrive environment. In: 2011 IEEE Forum on
Integrated and Sustainable Transportation System (FISTS). IEEE, Vienna, pp. 122–127.
Dresner, K., Stone, P., 2004. Multiagent traffic management: a reservation-based intersection control mechanism. In: Proceedings of the Third International
Joint Conference on Autonomous Agents and Multiagent Systems. IEEE Computer Society, New York, pp. 530–537.
Dresner, K., Stone, P., 2007. Sharing the road: autonomous vehicles meet human drivers. In: The 20th International Joint Conference on Artificial Intelligence,
pp. 1263–1268.
Dresner, K., Stone, P., 2008. A multiagent approach to autonomous intersection management. J. Artif. Intell. Res. 31, 591–656.
Emmelmann, M., Bochow, B., Kellum, C., 2010. Vehicular Networking: Automotive Applications and Beyond. John Wiley & Sons.
Feng, Y., Head, K.L., Khoshmagham, S., Zamanipour, M., 2015. A real-time adaptive signal control in a connected vehicle environment. Transp. Res. 55,
460–473.
Feng, Y., Yu, C., Liu, H.X., 2018. Spatial and temporal intersection control in a connected and automated vehicle environment. Transp. Res. Part C 89,
364–383.
Frey, H., Unal, A., Chen, J., Li, S., Xuan, C., 2002. Methodology For Developing Modal Emission Rates For Epa’s Multi-Scale Motor Vehicle & Equipment
Emission System. Ann Arbor, Michigan: US Environmental Protection Agency. North Carolina State University for Office of Transportation and Air Quality,
U.S. Environmental Protection Agency, Ann Arbor, MI.
Gurobi Optimization Inc., 2017. Gurobi Optimizer Reference Manual.
He, Q., Head, K.L., Ding, J., 2012. PAMSCOD: platoon-based arterial multi-modal signal control with online data. Transp. Res. Part C 20, 164–184.
He, X., Liu, H.X., Liu, X., 2015. Optimal Vehicle speed trajectory on a signalized arterial with consideration of queue. Transp. Res. Part C 61, 106–120.
Heydecker, B.G., 1992. Sequencing of Traffic Signals. Clarendon Press, Oxford.
Ilgin Guler, S., Menendez, M., Meier, L., 2014. Using connected vehicle technology to improve the efficiency of intersections. Trans. Res. Part C 46, 121–131.
Jiang, H., Hu, J., An, S., Wang, M., Park, B.B., 2017. Eco approaching at an isolated signalized intersection under partially connected and automated vehicles
environment. Transp. Res. Part C 79, 290–307.
Kamal, M.A.S., Mukai, M., Murata, J., Kawabe, T., 2013. Model predictive control of vehicles on urban roads for improved fuel economy. IEEE Trans. Control
Syst. Technol. 21, 831–841.
Kamalanathsharma, R.K., Rakha, H.A., 2013. Multi-stage dynamic programming algorithm for eco-speed control at traffic signalized intersections. In: 16th
International IEEE Conference on Intelligent Transportation Systems (ITSC 2013). IEEE, The Hague, pp. 2094–2099.
Koonce, P., Rodegerdts, L., Lee, K., Quayle, S., Beaird, S., Braud, C., Bonneson, J., Tarnoff, P., Urbanik, T., 2008. Traffic Signal Timing Manual, Final Report ed.
Federal Highway Administration, Washington, DC, USA.
Lee, J., Park, B., 2012. Development and evaluation of a cooperative vehicle intersection control algorithm under the connected vehicles environment. IEEE
Trans. Intell. Transp. Syst. 13, 81–90.
Li, Z., Elefteriadou, L., Ranka, S., 2014. Signal control optimization for automated vehicles at isolated signalized intersections. Transp. Res. Part C Technol.
49, 1–18.
Malakorn, K.J., Byungkyu, P., 2010. Assessment of mobility, energy, and environment impacts of intellidrive-based cooperative adaptive cruise control and
intelligent traffic signal control. In: 2010 IEEE International Symposium on Sustainable Systems and Technology (ISSST). IEEE, Arlington, VA, pp. 1–6.
Marler, R.T., Arora, J.S., 2004. Survey of multi-objective optimization methods for engineering. Struct. Multidiscipl. Optim. 26, 369–395.
Miyatake, M., Kuriyama, M., Takeda, Y., 2011. Theoretical study on eco-driving technique for an electric vehicle considering traffic signals. In: 2011 IEEE
Ninth International Conference on Power Electronics and Drive Systems (PEDS), pp. 733–738.
Naidu, D.S., 2002. Optimal Control Systems. CRC Press, Inc., U.S.
National Transportation Operations Coalition (NTOC), 2012. 2012 National Traffic Signal Report Card, Technical Report, Washington, DC.
Newell, G.F., 2002. A simplified car-following theory: a lower order model. Transp. Res. Part B 36, 195–205.
Rakha, H., Kamalanathsharma, R.K., 2011. Eco-driving at signalized intersections using V2i communication. In: 14th International IEEE Conference on Intel-
ligent Transportation Systems (ITSC). IEEE, Washington, DC, pp. 341–346.
Silcock, J.P., 1997. Designing signal-controlled junctions for group-based operation. Transp. Res. Part A 31, 157–173.
Tachet, R., Santi, P., Sobolevsky, S., Reyes-Castro, L.I., Frazzoli, E., Helbing, D., Ratti, C., 2016. Revisiting street intersections using slot-based systems. PLoS
ONE 11, 1–9. doi:10.1371/journal.pone.0149607.
Wan, N., Vahidi, A., Luckow, A., 2016. Optimal speed advisory for connected vehicles in arterial roads and the impact on mixed traffic. Transp. Res. Part C
69, 548–563.
Wang, M., Daamen, W., Hoogendoorn, S.P., van Arem, B., 2014a. Rolling horizon control framework for driver assistance systems. part i: mathematical
formulation and non-cooperative systems. Transp. Res. Part C 40, 271–289.
Wang, M., Daamen, W., Hoogendoorn, S.P., van Arem, B., 2014b. Rolling horizon control framework for driver assistance systems. part ii: cooperative sensing
and cooperative control. Transp. Res. Part C 40, 290–311.
Webster, F.V., 1958. Traffic Signal Settings. Her Majesty’s Stationery Office, London, England.
Wong, C.K., Heydecker, B.G., 2011. Optimal allocation of turns to lanes at an isolated signal-controlled junction. Transp. Res. Part B 45, 667–681.
Wong, C.K., Wong, S.C., 2003a. A lane-based optimization method for minimizing delay at isolated signal-controlled junctions. J. Math. Model. Algorithms
2, 379–406.
Wong, C.K., Wong, S.C., 2003b. Lane-based optimization of signal timings for isolated junctions. Transp. Res. Part B 37, 63–84.
Wu, X., He, X., Yu, G., Harmandayan, A., Wang, Y., 2015. Energy-optimal speed control for electric vehicles on signalized arterials. IEEE Transp. Intell. Transp.
Syst. 16, 2786–2796.
Yeo, H., Skabardonis, A., Halkias, J., Colyar, J., Alexiadis, V., 2008. Oversaturated freeway flow algorithm for use in next generation simulation. Transp. Res.
Record 2088, 68–79.