IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO.

3, SEPTEMBER 2018 287

Mixed-Integer Linear Programming for Optimal

Scheduling of Autonomous Vehicle
Intersection Crossing
Seyed Alireza Fayazi and Ardalan Vahidi

Abstract—We propose an urban traffic management scheme for search direction has focused on improving intersection flow by
an all connected vehicle environment. If all the vehicles are au- optimizing timing of traditional traffic signals informed by uni-
tonomous, for example, in smart city projects or future’s dense directional communication from connected vehicles [5], [6].
city centers, then such an environment does not need a physical
traffic signal. Instead, an intersection control server processes data One can expect higher energy efficiency and intersection flow
streams from approaching vehicles, periodically solves an opti- with bi-directional vehicle-signal communication where signals
mization problem, and assigns to each vehicle an optimal arrival adjust their timings and vehicles their speeds [7]. Autonomous
time that ensures safety while significantly reducing number of cars can further benefit from traffic signal information because
stops and intersection delays. The scheduling problem is formu- they not only process the incoming information rather effort-
lated as a mixed-integer linear program (MILP), and is solved
by IBM CPLEX optimization package. The optimization outputs lessly but also can precisely control their speed and arrival time
(scheduled access/arrival times) are sent to all approaching vehi- at a green light. The situation can get even better with 100%
cles. The autonomous vehicles adjust their speed accordingly by penetration of autonomous vehicles since a physical traffic light
a proposed trajectory planning algorithm with the aim of access- is not needed anymore as shown in concept papers [8]–[14].
ing the intersection at their scheduled times. A customized traffic Also because autonomous cars have much faster reaction times
microsimulation environment is developed to determine the poten-
tials of the proposed solution in comparison to two baseline sce- than human driven cars, the intersection controller can rapidly
narios. In addition, the proposed MILP-based intersection control switch between phases [15].
scheme is modified and simulated for a mixed traffic consisting of Some of the benefits of eliminating traffic signals in an all
autonomous and human-controlled vehicles, all connected through autonomous vehicle environment is discussed in [9] and demon-
a wireless communication to the intersection controller of a signal- strated by interesting simulation results in a recent publication
ized intersection.
[11]. However, optimal scheduling of vehicle arrivals at such
Index Terms—Intelligent transportation systems, connected and intersections remains an open problem. Our paper attempts to
autonomous vehicles, intersection traffic management, mixed in- address the gap in problem formulation by formalizing this
teger linear program, trajectory planning, traffic simulation and
modeling.
scheduling problem as a Mixed Integer Linear Program (MILP)
and shows its benefits in microsimulations. The proposed MILP-
based controller receives information such as location and speed
I. INTRODUCTION from each subscribing vehicle and suggests optimal access times
HILE traffic signals ensure safety of conflicting move- to individual vehicles. The vehicles adjust their speed by a tra-
W ments at intersections, they also cause much delay,
wasted fuel, and tailpipe emissions. Frequent stops and goes
jectory planning algorithm that each execute locally in order to
reach the intersection at their assigned access times. The access
induced by a series of traffic lights often frustrates passen- times are computed periodically on a central server by solv-
gers. However, recent studies have shown that vehicle to signal ing a MILP. The objective of the optimization is smoothing the
connectivity can improve this situation. Several publications traffic flow and minimizing the intersection delay, while ensur-
have focused on uni-directional traffic signal to vehicle com- ing intersection safety and considering each vehicle’s desired
munication for guiding connected vehicles to arrive at green velocity.
which increases their energy efficiency [1]–[4]. Another re- Preliminary results of our MILP-based intersection control
were presented in [16] and this journal paper expands on those
preliminary results with improved and more detailed formula-
Manuscript received June 28, 2017; revised November 2, 2017 and February
13, 2018; accepted February 17, 2018. Date of publication June 1, 2018; date
tion. This paper also provides a modified version of our MILP-
of current version August 23, 2018. This work was supported by a research based intersection control scheme with a physical traffic light
award from the BMW Information Technology Research Center in Greenville, to be applied to a mix of autonomous and human-controlled
SC, USA. (Corresponding author: Seyed Alireza Fayazi.)
The authors are with the Department of Mechanical Engineering, Clem-
connected vehicles.
son University, Clemson, SC 29634 USA (e-mail:, [email protected]; This paper is organized as follows: after a broad literature
[email protected]). review in Section II, we introduce the scheduling problem in
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
simple words in Section III. The notations used in this paper
Digital Object Identifier 10.1109/TIV.2018.2843163 are explained in Section IV. Problem formulation is described
2379-8858 © 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.

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
288 IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO. 3, SEPTEMBER 2018

in Section V followed by conversion to a MILP in Section VI. proposing an imaginary access area around the intersection, we
In Section VII, the nature of the MILP solution is shown in a managed to have a formulation based only on time of arrivals
simplified case study problem. In Section VIII, our proposed to that area. Then we converted our vehicle arrival scheduling
trajectory-planning algorithm to guide the autonomous vehicles problem to a mixed-integer linear program (MILP). The opti-
for a timely arrival at intersection is explained. Microsimula- mization problem is solved using IBM CPLEX solver and the
tion testbeds, computation load, and results are described in corresponding outputs (scheduled access/arrival times) are sent
Sections IX–XI, respectively. The modified intersection control to all approaching vehicles in our customized microsimulation
scheme for mixed traffic conditions is provided in Section XII, environment. A trajectory-planning algorithm is also presented
followed by conclusions. that guides individual vehicles for a timely arrival at access area.
Mixed-integer linear program (MILP) has been used in var-
ious path planning applications with collision avoidance, such
II. LITERATURE REVIEW
as airplanes and autonomous vehicles. In the field of vehicle-
The coordination and optimal timing of traffic signals are intersection coordination, Zhu et al. [20] used a lane-based traf-
by nature complex problems and backed by years of research in fic flow model to optimize the total travel time. The output
traffic engineering and operations research. In recent years, there of their optimization is the traffic flow from one lane to an-
has been a great deal of attention paid to intersection control for other through the conflict point model without considering the
connected vehicles. Towards the most related work, Raravi et al. travel times and velocities that are desirable to individual ve-
[17] determined the merge sequence in which vehicles cross the hicles. The vehicles’ capability to meet the scheduled time of
intersection region by formulating an optimization problem with arrival is not specifically addressed by the authors in [20] and
constraints to ensure safety. The formulation proposed in [17] is no microsimulations are provided. Our paper, however, provides
nonlinear, and, as a result, Matlab optimization toolbox fmincon microsimulations considering the interaction between vehicles
is used. The fmincon tool may only give local solutions [18] and and ensures that the vehicles can meet their scheduled arrival
does not guarantee a global optimum [17]. Lee et al. [19] also times by a trajectory-planning algorithm.
employed optimal control for a cooperative vehicle intersection Also in the area of phase and timing optimization for stan-
control. In this work, the trajectories of any two conflicting ve- dard two-phase or eight-phase intersection controllers, a set of
hicles are modified to minimize the overlap of trajectories in the MILP formulations have been proposed in the literature. Most
intersection area. This would not always provide a feasible solu- of these formulations can be applied to mixed traffic consist-
tion because of the complexity of the optimization formulation ing of autonomous and human-controlled vehicles. They either
(the objective function and constraints are nonlinear) [15], [20]. use off-line historical traffic data or assume a one-way com-
For this reason, in [19], a combination of active-set, interior munication where connected vehicles report information to an
point, and genetic algorithms is used as backup which adds to intersection controller but their travel trajectory cannot be con-
the execution time of their intersection control. trolled. For instance, He et al. [5] used probe vehicles’ on-line
Some other works, such as [21] and [22], used job-scheduling information to identify pseudo-platoons and found an optimal
techniques. Colombo et al. [21] view the time interval that each signal plan using MILP. Little’s MILP formulation [27], and a
vehicle spends in an intersection as the length of the job to recent work in [7], solves the bandwidth maximization problem,
be executed. Similarly, Xie et al. [22] view clusters in the ag- and assigns optimal offsets to the standard traffic signals in a
gregate flow representation of different routes as the jobs to two-way arterial. Other works in this area use model predictive
be scheduled; where clusters are a basic representation of a control for traffic signal control problem and formulate it as a
vehicle or group of vehicles [23]. They use an approximate dy- MILP problem [6], [28], [29].
namic programing procedure, called Controlled Optimization of
Phases [24], to obtain a near optimal solution. Also Ahn et al.
[25] translated the intersection collision avoidance to a job-shop III. PROBLEM STATEMENT
scheduling problem assuming first-order dynamics for the vehi- We seek an intersection controller that coordinates and har-
cles. In the field of multi-agent systems, the solution provided by monizes the flow of the approaching connected vehicles. The
Dresner et al. in [8] and [9] is based on a reservation paradigm controller resides on a computational server and receives infor-
which allows the vehicles to reserve a block in space-time in an mation of all subscribing vehicles and then schedules the inter-
intersection. The solution is not optimal in the sense that it is a section access time for each vehicle regularly. The scheduled
First Come, First Serve approach, and a reservation is rejected if access times (arrival times) are sent to all subscribing vehicles
any part of the requested space-time block has been previously so that they can adjust their speed accordingly. The challenge
reserved or occupied by another vehicle. A detailed review of co- is to find appropriate access times that ensure safety, passenger
operative intersection management systems can be found in [26]. comfort, and smoother traffic flow.
This paper proposes an optimization-based approach for in- We ignored all the left and right turns to simplify the
tersection traffic management. The challenge is to provide ve- presentation of ideas. We assume a two-phase/four-movement
hicles with travel recommendations that ensure energy effi- intersection. As shown in Fig. 1(a), Phase X (φX = {X , X })
ciency, safety (collision-free passage through intersection), and corresponds to a set of two traffic movements: (1) south-bound
smoother traffic flow (less intersection delay and number of denoted by dark letter X or X ; (2) north-bound denoted by light
stops). Incorporating all these goals into the intersection con- letter X or X . Similarly, Phase O (φO = {O , O }) corresponds
trol algorithm will complicate the formulation. However, by to a set of two traffic movements: (1) west-bound denoted by
Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
FAYAZI AND VAHIDI: MILP FOR OPTIMAL SCHEDULING OF AUTONOMOUS VEHICLE INTERSECTION CROSSING 289

TABLE I
NOTATIONS USED TO EXPRESS INTERSECTION ATTRIBUTES

Fig. 1. (a) Simplified two-phase/four-movement intersection. (b) An example

of 9 vehicles approaching an intersection.

TABLE II
NOTATIONS USED TO EXPRESS CONNECTED VEHICLES’ ATTRIBUTES

Fig. 2. An example of scheduling vehicle arrivals at an intersection (case study

r cv3 can speed-up so the southbound platoon on Phase X
of Fig. 1(b)). The depicted green/red timing is applicable to all-autonomous does not slow down or stop at red.
intersections with no physical traffic light only. r cv8 can speed-up to catch up with Phase X green light.
r cv7 and cv9 adjusts their speed to build a westbound pla-
toon on Phase O.
dark letter O or O ; (2) east-bound denoted by light letter O or
O .
IV. DEFINITIONS AND NOTATIONS
Fig. 1(b) demonstrates a sample scheduling problem for 9
vehicles. The remaining distances from the current position of Achieving an optimal solution such as the example shown
these vehicles to the intersection access point (stop position) in Fig. 2, requires to first formalize the problem objective as
are indicated on the y-axis of Fig. 2. Note that the y-axis values well as the constraints on vehicle movements. Before providing
are the remaining distances on four different movements which the proposed formulation, we first introduce the notation and
are projected on one single axis. As shown on this y-axis and their definitions in this section. Tables I and II summarize the
also in Fig. 1(b), five connected vehicles denoted by cv2 , cv4 , notations used in this paper.
cv5 , cv6 , and cv8 are traveling on Phase X and four vehicles cv1 , Connected Vehicles: For each intersection, we will assume a
cv3 , cv7 , and cv9 are traveling on Phase O. An intersection con- subscription process by which the approaching connected vehi-
troller is sought that is capable of scheduling the vehicle arrivals cles send subscription requests to the intersection control server
as shown on the horizontal axis of Fig. 2, as an example. It is and announce their presence, phase/movement at the intersec-
assumed that the vehicles follow the travel trajectories shown tion, and intended time of arrival. We set the subscription dis-
by dashed lines in Fig. 2 in order to make their scheduled arrival tance to 500 m from the intersection center. An unsubscribe
times. The provided sample solutions for both the scheduled ar- message is later sent from each individual vehicle to the inter-
rival times and the travel trajectories were plotted schematically. section controller server at the time the vehicle clears the inter-
The solution assigns virtual green splits to Phase X and O (see section. Thus, data is exchanged only during the subscription
green/red timings in Fig. 2) in such a way that intersection stops period, and the intersection occupancy can be tracked by the
are avoided if at all possible. The solution assumes that: intersection controller. We represent the list of all subscribed
r cv2 can slow down to pass the intersection with the south- connected vehicles as CV = {cvi }ni=1 where n is the size of
bound platoon on Phase X. CV . The list of connected vehicles is sorted by distance to the

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
290 IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO. 3, SEPTEMBER 2018

and explained in Table II. Please note that in this paper, we

assume that all vehicles prefer to travel at the average velocity
vavg = 56.3 kph (35 mph); and as a result, their distance divided
by vavg approximately yields their desired access times with
respect to current time (t0 = 0 sec).
If a vehicle has successfully followed the intersection con-
troller commands, it reports its previously assigned access time
as its new desired access time in each communication with the
intersection controller. In other words:

Fig. 3. A schematic of different regions of the proposed intersection. taccess,des,i (l) = taccess,i (l − 1) (3)
where taccess,des,i (l) is the desired access time at the lth execu-
tion of the vehicle controller of cvi ; and taccess,i (l − 1) is the
intersection where cv1 is the closest vehicle to the intersection most recently updated access time previously communicated to
at the time of subscription. The length of the vehicle is denoted the vehicle cvi . This is intended to minimize the deceleration
by L, and is taken to be 5.0 meter. and acceleration required each time the vehicle receives a new
Intersection: It is assumed that the intersection is a square assigned access time which is more fuel efficient.
with width W =10 m. As shown in Fig. 1(a), we consider a
two-phase intersection consisting of Phase X and Phase O as V. PROBLEM FORMULATION
φ = {φX , φO }. Each phase includes a set of non-conflicting
movements and M = {O , O , X , X } is the set of movements A. Objective
used in this paper. (see Table I). It is assumed that all intersecting The objective of increasing intersection throughput will be
roads have the same speed limit denoted by vm ax . formalized here as an optimization problem. The optimization
Time Instances: For each vehicle approaching an intersection, goal is to find the sequence and times of arrival for the vehicles
we are interested in the following time instances: (1) time when such that the maximum (latest) access time assigned to the
the front of the vehicle enters the intersection area at the stop- subscribed vehicles is minimized and any potential collision is
bar; (2) time when the rear of the vehicle exits the intersection prevented. Furthermore, in defining the objective function we
area; (3) time when the front of the vehicle reaches an access take into account the desired arrival time of the vehicles in such
distance from the intersection. As shown in Fig. 3, these time a way that vehicles would not face extreme delay or expedition
instances are denoted by tenter , texit , and taccess , respectively. In compared to their desired arrival times.
this figure, the intersection area and the access area are shown We formulate the main goal as to find the optimal sequence
by a shaded area and a solid box, respectively. Border of the and time of arrivals (taccess ) for the subscribed vehicles such that
access area is defined by daccess that is the estimated stopping the difference between the current time (t0 ) and the expected
distance of a vehicle in case of a safety concern and is calculated arrival time of the last vehicle passing the intersection in a given
as a function of the road average speed vavg : time window is minimized:
2
vavg J1 = taccess,j − t0
daccess = tres vavg − (1)
2adec,m ax s.t. taccess,j = max({taccess,1 , ..., taccess,n }) (4)
where tres = 0.5 sec is assumed to be the response time of an Minimizing the aforementioned objective could force the ve-
autonomous vehicle [11], and adec,m ax = −4 (m/s2 ) is the max- hicles to travel near the speed limit against their preference. To
imum deceleration considered for passenger cars in emergency avoid such a scenario, we define a cost on the difference between
braking. We obtain daccess ≈ 38 m by setting vavg = 56.3 kph assigned and desired access times for all vehicles:
(35 mph).

n
It should be emphasized that intersection access time (taccess ) J2 = |taccess,i − taccess,des,i | (5)
for each vehicle is the time the vehicle enters the access area; all i=1
other vehicles in the opposing movement must access the access The total cost function is then:
area at a sufficiently later time. If the intersection is not yet
cleared for safe passage of a vehicle in the opposing movement J = w1 J1 + w2 J2 (6)
then the access area provides enough stopping distance to avoid where w1 and w2 are penalty weights. We hypothesize that
a collision. The vehicles can be notified of the presence of an this optimization will result in reduced fuel consumption and
opposing vehicle at the intersection not only by their on-board intersection delay, even though these factors are not explicitly
sensors but also by the intersection controller. incorporated into the objective function.
Vehicle Attributes: The attributes of each vehicle cvi ∈ CV
(1 ≤ i ≤ n) that is subscribed to the intended intersection con- B. Constraints
troller are described by:
Several constraints are imposed to ensure safety. The main
cvi = mi , φi , di , vi , taccess,des,i  (2) challenge is expressing the constraints as a function of access

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
FAYAZI AND VAHIDI: MILP FOR OPTIMAL SCHEDULING OF AUTONOMOUS VEHICLE INTERSECTION CROSSING 291

vehicle pass and clear the access point. As an example, for a

vehicle of length L = 5 meter, stopping right√at access point,
and accelerating with ai = 2 m/s2 , this time is 2a iL
a i = 2.24 sec
which is consistent with our simulations. For vehicles stopped
further in the queue, this time is shorter than 2.24 sec. As a
result, for a standstill vehicle we set tgap1 to the time it would
take the vehicle body, with length of L, to completely pass over
Fig. 4. The earliest possible access time based on speed limit and maximum the access point or to 1 sec whichever is larger. For a very slow
accelerations.
moving leading vehicle, we set tgap1 to the time it would take
a following vehicle to decelerate and maintain a minimum safe
distance and time headway.
times so that a linear constrained optimization problem can be Please note that tgap1 defines a lower bound on the headway
derived at the end. In this study, we are assuming that all ve- between vehicles at an access point only. It means that: i) The
hicles have same length, weight, and acceleration/deceleration actual headway at access point is a decision variable determined
capabilities. In a real-world condition, these information needs by the intersection controller and can be larger than tgap1 after
to be communicated at the time that vehicles send subscription each intersection controller execution; and ii) The actual head-
requests to the intersection controller. way between two vehicles on their way to access area can be
1) Speed Limit and Maximum Acceleration: For each ve- different from tgap1 . For example, if the spacing (a function
hicle cvi , we should consider the speed limit requirement of time headway) to the vehicle in front is getting dangerously
vi ≤ vm ax as well as the maximum acceleration constraint short, a car-following subroutine that is incorporated into each
ai ≤ aacc, m ax ; where vi and ai are the velocity and acceler- vehicle trajectory control takes control of the vehicle, and does
ation of the vehicle, vm ax is set based on the speed limit of the not let the spacing between vehicles go below a certain thresh-
road, and aacc, m ax is set to +3 m/s2 . We introduce taccess, m in,i old based on their velocities. The car-following subroutine is
as the earliest time that cvi can access the intersection, if it trav- explained shortly in Section VIII.
els at maximum acceleration and maximum speed possible. This 3) Safety Gap on Conflicting Movements: Two vehicles trav-
earliest possible access time is calculated as explained in Fig. 4. eling on conflicting movements also need to be separated by a
Then we rephrase our aforementioned speed and acceleration safety gap. This time gap, if selected properly, guarantees that
constraints as: a vehicle can only enter the access area after all conflicting ve-
hicles have left the intersection area. Considering two vehicles
taccess,i ≥ taccess, m in,i (7) cvj and cvk that are on different phases of φj ∈ φ and φk ∈ φ
2) Safety Gap on the Same Movement: Two consecutive ve- (φj = φk ), there are four possible situations with just enough
hicles that are traveling on the same movement (e.g., east bound) safety gap between the vehicles. These situations are shown in
should be separated by a safety gap (headway) that is denoted by a space-time diagram in Fig. 5 by projecting the vehicles tra-
tgap1 in this paper. This time gap is independent of the vehicles’ jectories on one single distance axis. It can be concluded from
speed [11] (except at very low speeds), and is the minimum this figure that the following constraints cover all the possible
following time gap to avoid a rear end collision. The computer situations; here ∨ is the OR operator:
control in autonomous vehicles will eliminate human reaction
times; and assuming a very small communication delay, a time
taccess,j − texit,k ≥ 0
headway of 0.2 seconds may be sufficient for safe automatic
vehicle following [30]. Adding a safety factor, it is suggested ∨
in [11] that a h = 1 sec headway provides a reasonable upper
taccess,k − texit,j ≥ 0
bound for the response time of an autonomous vehicle. To en-
force the safety gap of tgap1 = h = 1 sec at access point, we ∀cvj , cvk ∈ CV, φj , φk ∈ φ, φj = φk . (9)
add the following constraint on any two consecutive vehicles
traveling on the same movement:
Although (9) mandates at least 0 sec gap between the access-
taccess,j − taccess,k ≥ tgap1 ing and exiting time instances of the first and second vehicles,
it does not explicitly state the gap needed between the access-
∀cvj , cvk ∈ CV, dj ≥ d k , mj , mk ∈ M, m j = mk .
ing times of those vehicles. We are specifically interested in
(8)
the time gap between access timestamps so that, at the end, we
However, a 1 sec headway is not sufficient at very low speeds can formulate the optimization problem based on access times
such as when discharging from a queue. As observed in our sim- only. For this reason, we define texit = taccess + Δttravel where
ulations, when discharging from a queue, the headway between Δttravel is the travel time between access point and exit point
the first vehicle and the second vehicle accessing the access of a vehicle. To simplify, we assume the intersection is a square
point can be as large as 2.3 sec. This time gap can be signif- and the two intersecting roads have the same speed limits. As
icant for trucks or other longer vehicles [31] and is limited to a result, the travel time does not depend on the phase. By sub-
the time period that front bumper and rear bumper of the lead stituting texit = taccess + Δttravel into constraint (9), we can

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
292 IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO. 3, SEPTEMBER 2018

Fig. 5. Possible scenarios of two vehicles passing an intersection with just enough safety gap between them. (a), (d) cvk accessing right after cv j exiting. (b),
(c) cv j accessing right after cv k exiting.

equalities in such a way that if one equation holds true then the
other equations are always redundant. The most widely known
method to handle this disjunctions is the big-M method that,
in our application, requires a binary variable B and a constant
Mbig [32]. For each set of constraints in (10) for cvj and cvk , we
add one artificial binary variable Bi (1 ≤ i ≤ # of constraints)
to take care of the discontinuity as:
taccess,j − taccess,k + Mbig Bi ≥ tgap2
Fig. 6. The longest possible travel time between the access and exit points. taccess,k − taccess,j + Mbig (1 − Bi ) ≥ tgap2
∀cvj , cvk ∈ CV, φj , φk ∈ φ, φj = φk , Bi binary
rephrase this constraint as: (11)

taccess,j − taccess,k ≥ tgap2 where Bi can be either 0 or 1, and Mbig is a large enough
number. If Bi =0 then the first equation of the above constraint
∨ holds true if taccess,j − taccess,k ≥ tgap2 and the second equa-
taccess,k − taccess,j ≥ tgap2 tion (taccess,k − taccess,j ≥ (tgap2 − Mbig )) is redundant and al-
ways holds true if Mbig is large enough. If Bi = 1 then the
∀cvj , cvk ∈ CV, φj , φk ∈ φ, φj = φk . (10) first equation (taccess,j − taccess,k ≥ (tgap2 − Mbig )) is redun-
where tgap2 = Δttravel is the safety gap we need between access dant and always holds true if Mbig is large enough, and the
times and is equal to the time period that a vehicle needs to second equation holds true if taccess,k − taccess,j ≥ tgap2 .
first pass the access area, then pass the intersection area, and It is possible to predict how large Mbig must be because ei-
finally exit the intersection completely. The longest travel time ther of the redundant equations discussed above needs to be al-
a vehicle could take is when it is stopped behind the access area ways fulfilled; and this requires that Mbig ≥ tgap2 + taccess,j −
and accelerates at its assigned access time as shown in Fig. 6(a). taccess,k . Considering the fact that tgap2 is small, and can be
We set an average acceleration of 2 m/s2 and obtain Δttravel neglected compared to Mbig , a lower bound to Mbig is equal
= 7.3 sec as calculated in Fig. 6(b) as the longest travel time. to the latest possible access time with respect to current time.
Consequently, we set tgap2 = 7.5 sec in our simulations as a As a worst-case scenario, a vehicle would travel the whole sub-
conservative value. With the price of reducing the intersection scription distance (e.g., 2 km) at a very low speed (e.g., 20 kph)
throughput, this conservative safe gap will not only allow safe and would stop at access point waiting for a long time (e.g., 500
left and right turns with respect to conflicting traffic but also sec); that leads to a 860 sec interval before the vehicle can access
reduces the possibility of collisions in worst case scenarios. the intersection. Although Mbig can be 860 sec, we set Mbig to
a large enough constant just to consider all possible scenarios
such as longer subscription distances and longer waiting delays.
VI. HANDLING OF REMOVABLE DISCONTINUITIES
We set Mbig to 2000 in our formulation.
We plan to solve the problem proposed in the previous sec-
tion by linear programming. As a result, any discontinuity and B. Discontinuity in Cost Function J1
disjunction in the formulation needs to be removed first. There
The cost function J1 , introduced in (4), is discontinuous in
are three disjunctions found in the previous section:
the sense that it includes a decision variable (the largest access
time). One solution, as used in our previous work [16] and [33],
A. Discontinuity in Constraint is to always assign the largest (last) access time to the furthest
Constraint (10) has discontinuity and is not linear because it subscribed vehicle that is cvn , considering the fact that the list
includes the OR logic operator (∨). The goal here is to convert of all subscribed vehicles (CV ) is sorted by distance to the
this constraint into an AND-combination of two or more in- intersection. We then rephrased the optimization objective and

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
FAYAZI AND VAHIDI: MILP FOR OPTIMAL SCHEDULING OF AUTONOMOUS VEHICLE INTERSECTION CROSSING 293

constraint as given in (12) where taccess,n is the access time

assigned to the cvn and should be larger than or equal to the
access times of all other vehicles.
J1 = taccess,n − t0
s.t. taccess,n ≥ ({taccess,1 , ..., taccess , n − 1}) (12)
However, in this manuscript, we restate (4) as
J1 =max({taccess,1 − t0 , ..., taccess,n − t0 }) which is a min-
max problem. As a result, we can replace J1 with a new so-
called slack variable Δtaccess , latest and n extra constraints as
follows:
J1 = Δtaccess,latest
s.t. Δtaccess,latest ≥ (taccess,i − t0 ) ∀i ∈ {1, ..., n} (13)

C. Discontinuity in Cost Function J2

The absolute value signs cannot be included in a lin-
ear programming formulation; as a result, the cost func-
tion J2 , introduced in (5), needs to be restated. We restate
J2 by adding a new so-called slack variable Δtaccess,abs,i =
|taccess,i − taccess,des,i |. Then, considering the fact that |x| =
max{x, −x} for any real number x, we can add two constraints
of Δtaccess,abs,i ≥ (taccess,i − taccess,des,i ) and Δtaccess,abs,i ≥
−(taccess,i − taccess,des,i ) in order to ensure that our added slack
variable is equal to |taccess,i − taccess,des,i |. In summary, the re-
stated cost function and the added constraints are as follows:
n
J2 = Δtaccess,abs,i
i=1

s.t. Δtaccess,abs,i ≥ (taccess,i − taccess,des,i )

Δtaccess,abs,i ≥ −(taccess,i − taccess,des,i ) (14)

VII. MILP CASE STUDY

Fig. 7. Scheduling the arrivals of vehicles (case study of Fig. 1(b)) using the
The linear objective function/constraints and mixed-integer proposed MILP model, solved by IBM’s CPLEX optimization package (y-axis:
variables in the optimal solution make our problem a Mixed In- projected remaining distance to access area, x-axis: assigned access times). (a)
teger Linear Program (MILP) for which efficient methods exist. All weight given to intersection throughput improvement. (b) All weight given
to satisfying the desired speeds of all vehicles. (c) 50% − 50% compromised
Our detailed microsimulation results with measure of effective- solution. (d) 80% − 20% compromised solution.
ness (MOE) analysis will be given later in Section XI; however,
this section employs a simplified case study only to observe an
example numerical output of the proposed formulation. and w2 = 100% in Fig. 7(b), all the weight is given to satisfy-
To solve this MILP problem in this section, we use IBM’s ing the desired speeds of all vehicles (J2 , cost function (14));
CPLEX optimization package. We simulate the same example as and, as a result, the intersection clearance time was increased
previously shown in Fig. 2 including n = 9 connected vehicles by 13 sec compared with Fig. 7(a). Two compromised solutions
cvi (1 ≤ i ≤ 9). We set the speed limit to vm ax = 72.4 kph are also demonstrated in Fig. 7(c) and (d) for comparison pur-
(45 mph), and average arterial road speed to vavg = 56.3 kph poses. Because of the updated cost function introduced in (13),
(35 mph). We assume that the current state of all the vehicles the intersection clearance time of these solutions are improved
is available: they are all traveling at vavg and their distance to compared to our results previously presented in [16]. For our
access area are [690, 750, 780, 900, 990, 1080, 1170, 1230, microsimulation, we set equal weights as w1 = 50% and w2 =
1290] meters, respectively. 50%.
Fig. 7(a)–(d), demonstrates the optimal solutions to the afore- A feature of our formulation is that overlapped arrival-times
mentioned problem found by MILP. As shown in Fig. 7(a), by can be assigned to two vehicles that are traveling on different
setting w1 = 100% and w2 = 0%, all the weight is given to parallel traffic movements (e.g., cv4 and cv2 in Figs. 2 and 7(b)–
intersection throughput improvement (J1 , cost function (13)), (d)). This also holds true for two opposing vehicles both making
and, as a result, some vehicles end up traveling near the speed left or right turns because they are not considered conflicting
limit (solid red lines). On the other hand, by setting w1 = 0% movements and they can be served at the same time. However, a

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
294 IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO. 3, SEPTEMBER 2018

Fig. 9. Latest possible access time based on the minimum cruising speed
(v m in ) and desired deceleration (a d e c ).
Fig. 8. Trajectory-planning (access-time-tracking subroutine) for vehicles so
that they can pass the MILP-controlled intersection with no stop.

the vehicle’s velocity at current time. We set aacc = 2 m/s2

safety gap of tgap2 is needed to separate a left- or right-turning as the desired acceleration of vehicles, only if it satisfies
vehicle from its conflicting traffic movements. Most important the condition of 2 m/s2 ≥ 2(d iΔ−vt 2i Δ t) to avoid taking the
feature that can also be identified in the Fig. 7(a), is that since square root of negative number in (15); otherwise, we set
tgap2 > tgap1 , platoon formations will be encouraged, otherwise aacc = 2(d iΔ−vt 2i Δ t) .
lone vehicles have to clear the intersection with the longer safe r Deceleration: If taccess,i > t0 + d i , the vehicle would be
vi
gap tgap2 with respect to opposing movements which reduces early at access point if it maintains its current speed. Thus,
intersection capacity. the vehicle needs to decelerate to a cruising speed (vcruise,i )
as shown in Fig. 8(c). Knowing the desired deceleration
VIII. AUTONOMOUS TRAJECTORY-PLANNING adec , the cruising speed is calculated by:
In order to guide the autonomous vehicles, a trajectory-
vcruise,i = vi + adec Δt
planning algorithm needs to run locally on individual vehicles’
  
control system. Our trajectory-planning engine has two sub- 1
routines: i) an access-time-tracking subroutine that computes a + 2adec adec Δt2 + vi Δt − di (16)
2
feasible trajectory for the vehicle in such a way that it can reach
and pass the access area at its assigned access time; and, ii) We set adec = −2m/s2 as the desired deceleration of
a car-following subroutine that constrains the vehicle’s speed vehicles only if it satisfies the condition of −2 m/s2 ≤
to obtain a safe space headway to the vehicle in front, if nec- 2(d i −v i Δ t)
Δ t2 to avoid taking the square root of negative num-
essary. Note that in this paper, to simplify the model of an
ber in (16); otherwise, we set adec = 2(d iΔ−vt 2i Δ t) .
autonomous vehicle, we assume there is no obstacle to avoid
and lane-changing is not allowed.
B. Planning a Stop at Queue or Access-Point
A. Planning an Uninterrupted Passage of a Vehicle The aforementioned equations do not always have solutions.
In other words, it is not always possible for a vehicle to pass
The basic fact to be considered before explaining the access-
without stopping. If the assigned access time is far away in time
time-tracking algorithm is that if a vehicle’s remaining travel
such that even a very slow-moving vehicle reaches the inter-
time to the intersection access area at current speed is equal
section before that time, then the vehicle should be prepared to
to the remaining time to its assigned access time, then, obvi-
stop. After stopping at an access point or at a queue behind ac-
ously, the vehicle doesn’t need to adjust its speed. As shown in
cess point, the vehicle waits for its reserved access time to come
Fig. 8(a), this means that the vehicle can continue driving at its
before it starts to proceed to the access area and the intersection.
current speed if Δt = taccess,i − t0 = dv ii , where di and vi are
To formulate this into our trajectory-planning algorithm, we
the current distance and velocity at current time t0 , and taccess,i
introduce taccess, m ax,i as the latest time that the vehicle can po-
is the access time assigned to the vehicle by the intersection
tentially access the intersection if it travels at minimum cruising
controller. Most often, however, acceleration or deceleration is
speed (vm in ) and average deceleration. Then, the vehicle needs
necessary to adjust the vehicle speed as explained below:
r Acceleration: If taccess,i < t0 + d i , the vehicle needs to to be prepared for a complete stop if:
vi
accelerate to a cruising speed in order to make it to the ac- taccess,i ≥ taccess, m ax,i (17)
cess point at its assigned timestamp. As shown in Fig. 8(b),
where taccess, m ax,i is locally computed by the trajectory-
the vehicle should keep that speed (vcruise,i ) till it reaches
planning engine of each vehicle as explained in Fig. 9.
the access area. Knowing the desired acceleration aacc , the
cruising speed is calculated by:
 C. Car-Following Subroutine
 
1 There is a close interaction between the car-following and
vcruise,i = vi +aacc Δt− 2aacc aacc Δt2 +vi Δt−di
2 access-time-tracking subroutines. As an example, let j − 1 and
(15) j be a pair of vehicles, with j following j − 1. If j − 1 is
where Δt is the remaining time to the assigned access time, moving slower and not obeying intersection controller com-
di is the remaining distance to the access point, and vi is mands then it may interfere with j. This interference does not

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
FAYAZI AND VAHIDI: MILP FOR OPTIMAL SCHEDULING OF AUTONOMOUS VEHICLE INTERSECTION CROSSING 295

the access area and waits at the access point to receive a feasible
access time.

D. Managing Access Time Infeasibility

If an unwanted change or disturbance happens locally to a
vehicle (e.g., because of a slow-moving vehicle ahead), a previ-
ously assigned access time may become infeasible, not allowing
the vehicle to pass the intersection. Infeasibility can be identified
when a vehicle cannot reach the access area at its assigned times-
tamp even if it finds the opportunity to travel at its maximum
acceleration and speed possible. In other words, an assigned
access time of taccess,i is infeasible if:
Fig. 10. Scheduling of two non-conflicting vehicles cv j and cv j −1 (vehicle
j following j − 1) when (a) both follow the intersection commands, (b) vehicle
j − 1 doesn’t follow the intersection controller commands between 10 sec and taccess,i < taccess, m in,i (18)
60 sec of the simulation.

where taccess, m in,i is previously explained in Fig. 4.

With the intersection controller, the car following and access-
time-tracking subroutines all running algorithms continuously,
the vehicles are capable of making real-time adjustments even
in the infeasible situation. If no feasible access time is avail-
able, the vehicle is commanded to travel at its desired veloc-
ity (free flow) which may be constrained by the car follow-
ing subroutine if there is a slow-moving vehicle ahead. This
continues till the next execution of the intersection controller
where it assigns a new feasible access time to the vehicle and
also updates the access times of other vehicles accordingly. If
no appropriate access time is assigned by the intersection con-
troller well ahead before the access area, the trajectory-planning
engine commands the vehicle for a complete stop at access
Fig. 11. Scheduling of two conflicting vehicles cv j and cv k when (a) both point or at a queue behind access point. The stopped vehicle
follow the intersection commands, (b) vehicle k doesn’t follow the intersection
controller commands between 10 sec and 60 sec of the simulation. then needs to wait till the next execution of the intersection
controller.
It should be emphasized that it is probable that the distur-
cause collision because vehicle j’s trajectory-planning engine bance (e.g., a slow-moving vehicle) fades away before (18)
has a car-following subroutine that takes the control of the ve- warns an infeasible situation. In this case, the access-time-
hicle and avoids collision if the spacing between j and j − 1 tracking subroutine takes the control back and recommends an
is getting dangerously short. The access-time-tracking subrou- updated travel trajectory to the vehicle by rerunning its algo-
tine of j takes the control back from the car-following if j − 1 rithms (Section VIII-A). The updated trajectory probably rec-
accelerates increasing the inter-vehicle gap. ommends larger acceleration and speed to compensate the time
To illustrate this, Fig. 10(a) shows the travel trajectories and wasted.
access times of these vehicles if they both obey the intersection
controller commands. Fig. 10(b), on the other hand, shows a
simulation where j − 1 slows down to 24.1 kph (15 mph) at 10 IX. BENCHMARKING TESTBEDS
sec and doesn’t obey the intersection controller commands until In order to conduct comprehensive evaluations in this paper,
60 sec of the simulation. we implemented two simulation testbeds with pre-timed traf-
It should be emphasized here that no vehicle is allowed to fic signal control systems. Autonomous vehicles approaching
cross the intersection (access area) at a time different from the a pre-timed traffic signal control provide a baseline testbed,
access time assigned to the vehicle. This avoids side impacts against which we can benchmark and compare our MILP-based
while the car-following subroutine prevents rear-end collisions. intersection controller. The signal phase and timing for the
Fig. 11 demonstrates a two-vehicle case study simulation with benchmark pre-timed traffic signals were obtained off-line from
similar conditions of Fig. 10 but for vehicles on two conflicting SYNCHRO (Trafficware 2011) optimization program. This op-
phases. In Fig. 11(b), the slow moving vehicle (cvk ) receives timized signal phase and timing were then used in the microsim-
an updated access time commanding it to pass after the other ulation to model the current state of the traffic light (cycle
conflicting vehicle (cvj ) and avoiding a side impact. If this time = 100 sec, green split = 44.5 sec, and yellow interval
updated access time is not sent to cvk , this vehicle does not pass = 3.5 sec).

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
296 IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO. 3, SEPTEMBER 2018

Fig. 13. Screenshots of simulation testbeds. Testbed A: Pre-timed signalized

intersection with no communication. Testbed B: Pre-timed signalized
intersection with unidirectional communication and speed-advisory. Testbed
Fig. 12. The graphics shows how our modified speed adisory algorithm finds MILP: MILP-controlled intersection with bidirectional communication.
the feasible velocity intervals in order to avvoid stopping at red, if possible.

X. COMPUTATIONAL LOAD
A. Testbed A: Camera-Based Trajectory-Planning at In microsimulations of our MILP-based intersection control
Pre-Timed Signalized Intersection scheme, the MILP problem was solved by IBM’s CPLEX opti-
In this testbed, we assume that there is no communication mization package running on Intel Core [email protected] GHz Windows
between the autonomous vehicles and the traffic signal devices, 7 laptop with 8 GB of RAM. By tuning this package for per-
and the vehicles can observe the current state of the traffic signals formance improvement and by reducing preprocessing compu-
ahead only with their on-board cameras. In our simulations, as tational load, we were able to achieve an average intersection
soon as a vehicle is within the range of its imaginary camera controller execution time of 120 ms for an average number of
(300 m), the current state of the simulated traffic light is fed into 50 subscribed vehicles over one hour of simulation. The in-
the vehicle’s trajectory-planner. tersection controller execution time varied between 28 ms and
2400 ms for 50 subscribed vehicles. It should be emphasized
that MILP is running by the intersection controller every 4 sec;
B. Testbed B: Communication-Based Trajectory-Planning at as a result, the system is able to adapt its recommendations to
Pre-Timed Signalized Intersection unwanted interference or delays, e.g., change in speed because
of a slow-moving vehicle ahead.
We also consider another benchmark algorithm where all au- The aforementioned execution time includes the MILP
tonomous vehicles are assumed to be able to receive the deter- solver execution time plus the time needed for preprocessing
ministic future state of traffic signals via unidirectional wireless the probe vehicle data and expressing the problem in canonical
communications when they are within the range of 500 m from form of: minimizing J subject to A.x ≤ b. In this form, x is the
the intersection. As soon as this information is received, the vector of decision variables to be determined defined as x =
trajectory of each individual autonomous vehicle, in free flow, (taccess,1 , ..., taccess,n , Δtaccess,abs,1 , ..., Δtaccess,abs,n , B1 , ...,
is planned based on the speed advisory algorithm proposed by Bm , Δtaccess,latest ); where n is the number of all subscribed
our group in [1], [2]. This algorithm is modified and its de- vehicles, and m is the number of artificial binary variables of
tails are not included here in consideration of the length of (11). Also, A is a matrix and b is a vector of coefficients that
t the paper. together represent the constraints and bounds that appeared in
Nevertheless, an example is given in Fig. 12 to clarify the (7), (8), (11), (13), and (14).
general features of our modified speed advisory system. Fig. 12
shows the feasible speed intervals that a vehicle traveling at a
XI. SIMULATION RESULTS
speed of vi can follow without stopping at red. These speed
intervals are limited to the speed limit vm ax (e.g., 45 mph) and In this section, we try to determine the potential in a MILP
the minimum cruising speed possible vm in (e.g., 20 mph). In controlled intersection. Towards this objective, in our simula-
this paper, we add a constant buffer Δt = 4 sec after each start- tion, we assume that instantaneous vehicle information is avail-
of-green so that a vehicle would never be advised to cross the able to the intersection controller, no disturbance input exists
stop-bar at red. Among the available speed intervals shown in in the simulated traffic, there is no obstacle to avoid and lane-
Fig. 12, we choose a target speed as close as possible to the changing is not allowed, all vehicles prefer to travel at the av-
desired average speed vavg . erage velocity, and they all have same length and acceleration
At this testbed, we are assuming that the autonomous vehicles capabilities. The microsimulation was implemented using Java.
have information about the instantaneous queue size when they The screenshots of the implemented simulation environments
are within 300 m of the intersection. If a queue exists then the for three testbeds are shown in Fig. 13.
vertical axis of Fig. 12 represents the distance to the rear end Each simulated testbed includes an intersection with four
of the queue. In this case, the safety buffer Δt compensates, in legs, each 500 m long. The simulated vehicles arrive based on a
part, for the queue dissipation time. probabilistic generation method: a negative Exponential dis-

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
FAYAZI AND VAHIDI: MILP FOR OPTIMAL SCHEDULING OF AUTONOMOUS VEHICLE INTERSECTION CROSSING 297

TABLE III
THE SIMULATION RESULTS, AND THE OVERALL PERFORMANCE IMPROVEMENTS ACHIEVED BY MILP-BASED INTERSECTION CONTROLLER WHEN
ALL VEHICLES ARE CONNECTED AND AUTONOMOUS

1
All MOEs are reported for one intersection with four approaches (each approach is 500 m long).
2
The MILP results were obtained by setting w 1 = 50% and w 2 = 50%.
3
Travel time over a distance of 500 m to the intersection.

tribution [34] was used for 750 vehicles/hour for all four ap-
proaches. The vehicles’ arrival pattern is recorded and replayed
for each testbed. In this way, the same arrival pattern was repli-
cated for each testbed. Three simulations were conducted for
Testbeds A, B, and MILP. The average and maximum speeds
were set to vavg = 56.3 kph (35 mph) and vm ax = 72.4 kph
(45 mph). For our communication-based testbeds (testbed B
and MILP), the subscription range is 500 m.
The MOE results of the simulations are given in Table III,
Fig. 14. The green/red timing output of the modified intersection control, ap-
where the performance improvements achieved by our MILP- plicable to mixed traffic of non-autonomous and autonomous connected vehicles
based intersection controller are also provided compared to at a signalized intersection.
Testbeds A and B. The results of our MILP-controlled inter-
section were obtained by setting w1 = 50% and w2 = 50%
in (6); however, it may be possible to tune these parameters Assuming all the vehicles are connected to the intersection con-
for improved results. The MOEs studied in our simulations and troller, we make the following modifications:
given in Table III are: (1) the intersection total number of stops, r Physical traffic signals are added and simulated at the ac-
(2) the intersection total stopped delay, (3) the average stopped cess points. The status of the traffic light (green, yellow,
delay per stopped vehicle, and (4) the average travel time per and red) controls the traffic flow to the access area at each
vehicle. As shown in Table III, by using our MILP-based control of four approaches to the intersection. It is possible to
at Testbed MILP, the intersection delay and stops were signifi- have the traffic lights located at the stop-bars, controlling
cantly reduced compared to pre-timed intersection benchmarks; the traffic flow to the intersection area instead. However,
and travel times were not compromised. These improvements this requires major modification to our intersection control
were obtained although the road capacities were reduced in scheme.
Testbed MILP. The capacity is reduced because the queue start- r All autonomous vehicles travel based on their assigned
ing point for Testbed MILP is the access point, while for other access time with the desired velocity of vavg . The non-
testbeds it is the stop bar at intersection. The MILP results are autonomous vehicles desire to travel at a randomly selected
also affected by the conservative safe gap value of tgap2 = 7.5 velocity between vm in and vm ax .
sec; however, this amount of gap reduces the chance of collisions r Both autonomous and non-autonomous vehicles decide
in worst case scenarios. to pass or stop at the access point based on the traffic
By using the speed advisory at Testbed B, the total duration light status only. When autonomous vehicles are in close
and number of stops at intersection is decreased compared with proximity of the traffic light, they ignore their assigned
that of Testbed A; however, the average travel time has not been access times and pass at green or stop at red signal. Thus,
improved (shown in Table III). The reason is that, in Testbed B, an autonomous vehicle has the flexibility to pass at a time
most of the vehicles that manage to pass the intersection without different than its assigned access time as long as the light
stopping have been commanded to travel at slow speeds in order is green.
to avoid a red light. r Considering a slower response time for drivers, the MILP-
based intersection controller assumes tgap1 = 2 sec if a
following vehicle is not autonomous.
XII. MODIFIED DESIGN FOR MIXED TRAFFIC CONDITIONS r The example timing of Fig. 2 is shown again in Fig. 14
In this section, we provide a modified version of our MILP- but in an application with a physical traffic light. A mini-
based intersection control scheme that can be applied to a mixed mum green time of Gm in = 8 sec is incorporated into the
traffic consisting of automated and human-controlled vehicles. signal timing. The larger this green time, the more non-

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
298 IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, VOL. 3, NO. 3, SEPTEMBER 2018

dination scheme provided 19.5% benefit in fuel consumption

compared to our baseline testbed with a pre-timed traffic signal.
The initial proposed algorithm considers a futuristic all au-
tonomous driving environment and eliminates the need for phys-
ical traffic signals (100% penetration rate of equipped vehicles).
This can be utilized in future’s dense city centers where only
autonomous vehicles would be allowed to travel. However, we
also modified the proposed algorithm to be applied to mixed
traffic consisting of autonomous and human-controlled vehi-
cles, all connected to the intersection controller of a signalized
Fig. 15. The change of measure of effectiveness with different mixed traffic
conditions (one hour simulation of signalized intersection). intersection. More modifications are required to make this in-
tersection control scheme applicable to situations with purely
human-driven vehicles that have no wireless connectivity; as
autonomous vehicles find enough time to react to a green an example, a maximum green time needs to be considered
light and cross the access point. As shown in Fig. 14, this for each phase of the traffic light. Other future work includes
green time is split in two, each added to the beginning and simulating left- and right-turning vehicles as well as multi-lane
ending of each phase. The length of the yellow change intersections.
interval is 3.5 sec. The length of clearance or all-red phase
is Rall−red = 7.5 sec based on the discussion previously
ACKNOWLEDGMENT
provided in Section V-B3. As shown in Fig. 14, the inter-
section controller sets tgap2 = Gm in +Rall−red = 15.5 sec. The authors would like to thank Dr. A. Luckow from BMW
Different scenarios with varying the penetration rate of au- Information Technology Research Center.
tonomous vehicles are simulated with the same conditions ex-
plained in Section XI. The results are shown in Fig. 15. The
figure demonstrates how MOEs degrade compared with that of REFERENCES
all-autonomous condition (100% penetration rate). [1] B. Asadi and A. Vahidi, “Predictive use of traffic signal state for fuel
At 100% penetration, during one hour simulation, we ob- saving,” in Proc. 12th IFAC Symp. Transp. Syst., 2009, pp. 484–489.
[2] B. Asadi and A. Vahidi, “Predictive cruise control: Utilizing upcoming
served 685 stops, 1h and 42 min total stopped delay, 9 sec aver- traffic signal information for improving fuel economy and reducing trip
age stopped delay per stopped vehicle, and 47 sec average travel time,” IEEE Trans. Control Syst. Technol., vol. 19, no. 3, pp. 707–714,
time. Please note, because of having a physical traffic signal, we May 2011.
[3] H. Rakha and R. K. Kamalanathsharma, “Eco-driving at signalized inter-
should use our modified intersection control scheme presented sections using V2I communication,” in Proc. IEEE Intell. Transp. Syst.
in this section to simulate a 100% penetration of autonomous Conf., 2011, pp. 341–346.
vehicles. Otherwise, the traffic signal switches rapidly between [4] S. Mandava, K. Boriboonsomsin, and M. Barth, “Arterial velocity plan-
ning based on traffic signal information under light traffic conditions,” in
phases and it doesn’t turn from red to green till the moment Proc. 12th Int. IEEE Conf. Intell. Transp. Syst., 2009, pp. 1–6.
an autonomous vehicle crosses the access point (see Fig. 2). [5] Q. He, K. L. Head, and J. Ding, “PAMSCOD: Platoon-based arterial multi-
That is why the aforementioned results for an all-autonomous modal signal control with online data,” Transp. Res. C, Emerg. Technol.,
vol. 20, no. 1, pp. 164–184, 2012.
environment are different from Table III. [6] M. A. S. Kamal et al., “Traffic signal control of a road network using
MILP in the MPC framework,” Int. J. Intell. Transp. Syst. Res., vol. 13,
no. 2, pp. 107–118, 2015.
XIII. CONCLUSIONS & FUTURE WORK [7] G. D. Nunzio, G. Gomes, C. Canudas-de-Wit, R. Horowitz, and P. Moulin,
“Speed advisory and signal offsets control for arterial bandwidth maxi-
We proposed a novel intersection control scheme for an all mization and energy consumption reduction,” IEEE Trans. Control Syst.
autonomous vehicle environment. Our five key contributions are Technol., vol. 25, no. 3, pp. 875–887, May 2017.
[8] K. Dresner and P. Stone, “Multiagent traffic management: A reservation-
in: i) an intersection control algorithm that anticipates vehicle based intersection control mechanism,” in Proc. 3rd Int. Joint Conf. Auton.
arrivals and facilitates uninterrupted passage of them at intersec- Agents Multiagent Syst., 2004, vol. 2, pp. 530–537.
tions, ii) reducing the vehicle-intersection coordination problem [9] K. Dresner and P. Stone, “A multiagent approach to autonomous intersec-
tion management,” J. Artif. Intell. Res., vol. 31, pp. 591–656, 2008.
to a mixed-integer linear program, iii) the improvements in traf- [10] M. A. S. Kamal, J. Imura, T. Hayakawa, A. Ohata, and K. Aihara, “A
fic flow, compared to [16], by introducing an updated cost func- vehicle-intersection coordination scheme for smooth flows of traffic with-
tion, iv) a trajectory planning algorithm that guides autonomous out using traffic lights,” IEEE Trans. Intell. Transp. Syst., vol. 16, no. 3,
pp. 1136–1147, Jun. 2015.
vehicles for a timely arrival at intersections, and v) developing [11] R. Tachet et al., “Revisiting street intersections using slot-based systems,”
a customized microsimulation test environment in which simu- PLoS One, vol. 11, no. 3, 2016, Art. no. e0149607.
lated vehicles are guided by our trajectory-planning algorithm. [12] Y. J. Zhang, A. A. Malikopoulos, and C. G. Cassandras, “Optimal control
and coordination of connected and automated vehicles at urban traffic
Microsimulation results demonstrated that our linear formula- intersections,” in Proc. IEEE Amer. Control Conf., 2016, pp. 6227–6232.
tion not only significantly reduced the intersection delay and [13] A. de La Fortelle, “Analysis of reservation algorithms for cooperative
stops compared to pre-timed intersection benchmarks, but also planning at intersections,” in Proc. 13th Int. IEEE Conf. Intell. Transp.
Syst., 2010, pp. 445–449.
ensured that no crash occurred and did not compromise the travel [14] L. Li and F.-Y. Wang, “Cooperative driving at blind crossings using in-
times. Our recent work, on implementing a Vehicle-In-the-Loop tervehicle communication,” IEEE Trans. Veh. Technol., vol. 55, no. 6,
(VIL) platform [33], showed that our intersection-vehicle coor- pp. 1712–1724, Nov. 2006.

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.
FAYAZI AND VAHIDI: MILP FOR OPTIMAL SCHEDULING OF AUTONOMOUS VEHICLE INTERSECTION CROSSING 299

[15] S. I. Guler, M. Menendez, and L. Meier, “Using connected vehicle tech- [32] M. Berkelaar, K. Eikland, and P. Notebaert, “Reference guide to open
nology to improve the efficiency of intersections,” Transp. Res. C, Emerg. source (mixed-integer) linear programming system (version 5.1.0.0 dated
Technol., vol. 46, pp. 121–131, 2014. 1 may 2004).” [Online]. Available: http://lpsolve.sourceforge.net/5.1/
[16] S. A. Fayazi, A. Vahidi, and A. Luckow, “Optimal scheduling of au- [33] S. A. Fayazi and A. Vahidi, “Vehicle-in-the-loop (VIL) verification of a
tonomous vehicle arrivals at intelligent intersections via MILP,” in Proc. smart city intersection control scheme for autonomous vehicles,” in Proc.
IEEE Amer. Control Conf., 2017, pp. 4920–4925. IEEE Conf. Control Technol. Appl., 2017, pp. 1575–1580.
[17] G. Raravi et al., “Merge algorithms for intelligent vehicles,” in Next Gen- [34] T. V. Mathew, “Transportation systems engineering,” Cell Transmission
eration Design and Verification Methodologies for Distributed Embedded Models, IIT Bombay, Mumbai, India, 2014.
Control Systems. New York, NY, USA: Springer, 2007, pp. 51–65.
[18] T. Coleman, M. A. Branch, and A. Grace, Optimization Toolbox—For
Use With MATLAB, User’s Guide for MATLAB, vol. 5, 1999.
[19] J. Lee and B. Park, “Development and evaluation of a cooperative vehicle S. Alireza Fayazi received the B.Sc. degree from K.
intersection control algorithm under the connected vehicles environment,” N. Toosi University of Technology, Tehran, Iran, the
IEEE Trans. Intell. Transp. Syst., vol. 13, no. 1, pp. 81–90, Mar. 2012. M.Sc. degree from the University of Tehran, Tehran,
[20] F. Zhu and S. V. Ukkusuri, “A linear programming formulation for au- Iran, both in electrical and electronics engineering,
tonomous intersection control within a dynamic traffic assignment and and the Ph.D. degree in dynamic systems and controls
connected vehicle environment,” Transp. Res. C, Emerg. Technol., vol. 55, (mechanical engineering) from Clemson University,
pp. 363–378, 2015. Clemson, SC, USA. He is currently a Postdoctoral
[21] A. Colombo and D. Del Vecchio, “Efficient algorithms for collision avoid- Fellow with Clemson University. During 2012–2013,
ance at intersections,” in Proc. 15th ACM Int. Conf. Hybrid Syst., Comput. he was a Visiting Researcher with the University of
Control, 2012, pp. 145–154. California, Berkeley and was also working as a Vis-
[22] X.-F. Xie et al., “Schedule-driven intersection control,” Transp. Res. C, iting Researcher with BMW Group Technology Of-
Emerg. Technol., vol. 24, pp. 168–189, 2012. fice, Mountain View, CA, USA. Before joining Clemson University, he was a
[23] A. Hawkes, “Traffic control with connected vehicle routes in SURTRAC,” Research and Development Engineer for three years with Kerman Tablo Cor-
M.S. thesis, Robot. Inst., Carnegie Mellon Univ., Pittsburgh, PA, USA, poration where he worked on discrete control systems and digital control for
Tech. Rep. CMU-RI-TR-16–20, 2016. embedded applications.
[24] S. Sen and K. L. Head, “Controlled optimization of phases at an intersec-
tion,” Transp. Sci., vol. 31, no. 1, pp. 5–17, 1997.
[25] H. Ahn and D. Del Vecchio, “Semi-autonomous intersection collision
avoidance through job-shop scheduling,” in Proc. 19th Int. Conf. Hybrid
Syst., Comput. Control, 2016, pp. 185–194.
[26] L. Chen and C. Englund, “Cooperative intersection management: A sur-
vey,” IEEE Trans. Intell. Transp. Syst., vol. 17, no. 2, pp. 570–586, Ardalan Vahidi received the B.S. and M.Sc. degrees
Feb. 2016. in civil engineering from Sharif University, Tehran,
[27] J. D. Little, “The synchronization of traffic signals by mixed-integer linear Iran, in 1996 and 1998, respectively, the M.Sc. de-
programming,” Oper. Res., vol. 14, no. 4, pp. 568–594, 1966. gree in transportation safety from The George Wash-
[28] M. A. S. Kamal, J. I. Imura, and T. Hayakawa, “Network-wide opti- ington University, Washington, DC, USA, in 2002,
mization of traffic signals using mixed integer programming,” J. Robot. and the Ph.D. degree in mechanical engineering from
Mechatronics, vol. 26, no. 5, pp. 607–615, 2014. the University of Michigan, Ann Arbor, MI, USA,
[29] S. Lin, B. De Schutter, Y. Xi, and H. Hellendoorn, “Fast model predictive in 2005. He is currently a Professor in mechanical
control for urban road networks via MILP,” IEEE Trans. Intell. Transp. engineering with Clemson University, Clemson, SC,
Syst., vol. 12, no. 3, pp. 846–856, Sep. 2011. USA. During 2012–2013, he was a Visiting Scholar
[30] C. Chien and P. Ioannou, “Automatic vehicle-following,” in Proc. IEEE with the University of California, Berkeley. He has
Amer. Control Conf., 1992, pp. 1748–1752. also held Scientific Visiting positions at BMW Technology Office in Califor-
[31] S. A. Fayazi and A. Vahidi, “Crowdsourcing phase and timing of pre-timed nia and at IFP Energies nouvelles in France. His research interests include the
traffic signals in the presence of queues: Algorithms and back-end system intersection of energy, vehicular systems, and automatic control. His recent pub-
architecture,” IEEE Trans. Intell. Transp. Syst., vol. 17, no. 3, pp. 870–881, lications span topics in alternative vehicle powertrains, intelligent transportation
Mar. 2016. systems, and connected and autonomous vehicle technologies.

Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 22,2025 at 13:23:36 UTC from IEEE Xplore. Restrictions apply.