Content uploaded by Shunsuke Aoki
Author content
All content in this area was uploaded by Shunsuke Aoki on Jul 03, 2018
Content may be subject to copyright.
Dynamic Intersections and Self-Driving Vehicles
Shunsuke Aoki and Ragunathan (Raj) Rajkumar
Electrical and Computer Engineering, Carnegie Mellon University
Pittsburgh, PA
Email: shunsuka@andrew.cmu.edu, raj@ece.cmu.edu
Abstract—Connected and automated vehicles are expected
to be at the core of future intelligent transportation systems.
One of the main practical challenges for self-driving vehicles
on public roads is safe cooperation and collaboration among
multiple vehicles when conflicts arise on shared road segments.
Intersections controlled by traffic lights and stop signs are
common examples of such potential conflicts, and cooperative
protocols for such intersections have been studied. On the other
hand, there are many different types of shared road segments.
In this paper, we study Dynamic Intersections that might appear
almost anytime and anywhere on public roads and that might
lead to automobile accidents. We consider how a self-driving
vehicle can safely navigate these dynamic intersections by using
sensor-based perception and inter-vehicle communications. We
present a cooperative protocol for dynamic intersections which
can be used by self-driving vehicles for safely coordinating with
other vehicles. Under our protocol, self-driving vehicles can also
create a vehicular communication-based traffic manager named
Cyber Traffic Light when the area is congested. A cyber traffic
light functions as a self-optimizing traffic light by estimating the
traffic volumes and by wirelessly coordinating among multiple
self-driving vehicles. Our simulation results show that our pro-
tocol has higher traffic throughput, compared to simple traffic
protocols while ensuring safety.
Index Terms—Autonomous vehicles; Intersection management;
Vehicular networks; Intelligent transportation systems
I. INTRODUCTION
With advances in Cyber-Physical System (CPS) technolo-
gies, connected and automated vehicles are becoming in-
creasingly feasible. Many companies are testing self-driving
vehicles on public roads [12], while academic research has
produced the earliest prototypes, such as the CMU self-driving
Cadillac SRX [19].
One of the main practical challenges for self-driving ve-
hicles on public roads is safe cooperation and collaboration
among multiple vehicles when conflicts arise on shared road
segments. Intersections controlled by traffic lights, round-
abouts, and stop signs are common examples of such potential
conflicts. We refer to these road intersections as “stationary
intersections” since they are captured in map databases, are
known a priori, do not move, and last for a very long time.
Cooperative protocols for such intersections have been studied,
for example in [8], [10].
In this paper, we identify and classify Dynamic Intersections
that might appear almost anytime and anywhere on public
roads leading to potential dynamic conflicts that must be safely
resolved. We then consider how a self-driving vehicle can
safely navigate these dynamic intersections using a combina-
tion of sensor-based perception and inter-vehicle communica-
tions. A dynamic intersection (a) represents a shared region on
the road where traffic conflicts can arise, (b) may be temporary,
(a) One-way Alternative Use
(Temporary DI) (b) Turning Left through
Traffic (Unmapped DI)
(c) Single-track Lane or
Narrow Bridge
(Mapped DI) (d) Merge Point
(Mapped DI)
(e) Lane Changing
(Moving DI)
(f) Passing
(Moving DI)
(g) Center Turn Lane
(Variegated DI)
Figure 1: Example of Different Types of Dynamic Intersec-
tions.
(c) may arise dynamically due to the occasional presence of
vehicles, and (d) is not included in a map database (unlike a
multi-way road intersection or a roundabout).
In Figure 1, we present 7 examples of different types of
dynamic intersections: (a) One-way Alternative Use due to
construction; (b) Turning Left through Traffic into a parking
lot or driveway; (c) Single-track Lane or Narrow Bridge; (d)
Merge Point; (e) Lane Changing; (f) Passing on a lane usually
used by oncoming traffic; and (g) Center Turn Lane. In a
dynamic intersection, a road segment must be shared by two or
more vehicles whose travel paths intersect. Other examples of
dynamic intersections occur in large parking lots, in multi-level
parking garages, in areas with disabled or crashed vehicles,
on roads with double parked vehicles blocking a lane, along
damaged roads, or near potholes. Human drivers navigate these
regions using social norms, courtesy, hand gestures, eye con-
tact, and common sense. However, for self-driving vehicles,
to guarantee safety, real-time interaction and negotiation are
essential.
In this paper, we present a cooperative traffic protocol for
safe traversal through many kinds of dynamic intersections.
In this protocol, each self-driving vehicle uses both Vehicle-
to-Vehicle (V2V) communications and internal sensor-based
perception systems. Our protocol avoids vehicle collisions
and possible deadlocks by using V2V communications, and
guarantees safety by using sensor-based perception systems.
Under the protocol, the self-driving vehicles also create a V2V
communication-based traffic manager named Cyber Traffic
Light (CyberTL) when traffic is congested to provide fairness
to vehicles waiting for a long time. The CyberTL works as a
self-optimizing traffic light at a dynamic intersection, by lo-
cally estimating the traffic volume and wirelessly coordinating
among the vehicles. In other words, the CyberTL provides high
traffic throughput around dynamic intersections, while keeping
the worst-case waiting time short.
The contributions of this paper are as follows.
1) We define and classify dynamic intersections that might
appear almost anytime and anywhere.
2) We present a cooperative protocol for guaranteeing
safety at dynamic intersections.
3) We evaluate our traffic protocol using a simulator-
emulator, and demonstrate the superior performance of
our protocol.
The remainder of this paper is organized as follows. Section
II describes our system assumptions, and presents a classifi-
cation of dynamic intersections. Section III presents our co-
operative dynamic intersection protocol along with examples
to illustrate its safety and usage. In Section IV, we evaluate
our protocol and compare its performance against traditional
traffic protocols. Section V discusses previous work related to
our research. Finally, Section VI presents our conclusions and
future work.
II. ASSUMPTIONS
In this section, we present our system architecture and
a classification of dynamic intersections. We also provide a
review of the existing traffic control methods for human-driven
cars.
A. Vehicle Systems
We assume that a self-driving vehicle comprises various
hardware and software components, such as a map database,
a navigation system, a perception system, an autonomous
vehicle controller, a localization capability, and a wireless
communication interface [19].
The map database contains accessible lane information,
any turn restrictions associated with each lane, and the geo-
graphical layout of stationary intersections. The database also
contains meta-information about stationary road intersections
like roundabouts and whether they are controlled by physical
traffic lights or stop signs. However, it does not contain all the
possible collision points and dynamic intersections on public
roads. The navigation system works with the map database
and allows passengers to input, plan, and display a route
to be taken. The perception system senses the surrounding
environment by fusing data from different types of sensors,
such as radar, LIDAR, and/or vision cameras. Sensor fusion
combines the radar and LIDAR data, and makes them more
consistent. The autonomous vehicle controller takes as its input
the route generated by the navigation system, and actuates
the vehicle to traverse the chosen path. In addition, each
vehicle is equipped with a high-accuracy Global Positioning
System (GPS) receiver. This receiver provides the vehicle with
both localization and time-synchronization services. Reason-
able positioning accuracy is important for self-driving vehi-
cle operation with the required level of accuracy potentially
achieved by Differential GPS techniques and/or correction
services such as Real-Time Kinematics (RTK) [14]. Time
synchronization among vehicles is obtained from GPS and its
PPS clock source. Finally, the wireless interface enables the
system to transmit and receive information wirelessly using
one or more mechanisms, such as 4G/5G cellular interfaces,
satellite communications, or an embedded infrastructure. In
this paper, we assume that each self-driving vehicle has an
On-Board Unit (OBU) that supports Dedicated Short-Range
Communications (DSRC) and the Wireless Access in a Vehic-
ular Environment (WAVE) protocol stack [9], [13]. This OBU
is used to exchange data among multiple automated vehicles.
The OBU also performs the functions of a Cyber Traffic Light
(See Section III-D) by sending approval or rejection to the
requests sent by other surrounding vehicles.
To simplify our presentation, we assume that each vehicle
has good GPS reception, and all system components are
functional. We will discuss the safety assurance provided by
our protocol in Section III-E, where we show that packet losses
do not compromise safety.
B. Dynamic Intersections
A dynamic intersection is a road area shared by two or more
vehicles whose travel paths intersect, giving rise to conflict
points. Moreover, this area is not controlled by traffic lights,
stop signs, or roundabouts. We first give the common features
of these dynamic intersections, and then, we classify different
types of dynamic intersections.
At a dynamic intersection, (i) vehicles can enter from
more than one direction; (ii) no physical traffic light, sign,
or roundabout is used to manage traffic; (iii) priorities may
be assigned to approaching directions; and (iv) the map
database does not store the presence of the intersection. Under
the existing traffic system, human drivers navigate dynamic
intersections by following priority rules, by communicating
(a) One-way Alternative Use (b) Turning Left through
Traffic
Figure 2: Dynamic Intersection area and Lane priority.
with each other, by making eye contact or using hand signals,
and/or by accounting for the surrounding environment.
Each dynamic intersection contains at least one collision
point (or conflict point). The collision point is the spatial point
where two vehicles might collide because their travel paths
intersect.
We classify dynamic intersections into at least 5 different
types. These are presented in Figure 1: Temporary Dynamic
Intersection,Unmapped Dynamic Intersection,Mapped
Dynamic Intersection,Moving Dynamic Intersection, and
Variegated Dynamic Intersection. We describe these types
below.
Temporary Dynamic Intersection (Temporary DI): This
road segment is temporarily used by vehicles coming from
two different directions due to construction, traffic control,
or blocking obstacles, as shown in Figure 1-(a). The detailed
information about the intersection might not be included in
the map database.
Unmapped Dynamic Intersection (Unmapped DI): An
Unmapped DI is not stored as a road intersection in the map
database, but there is a possible collision point when vehicles
turn left through oncoming traffic. This type of dynamic
intersection typically appears in front of the entrance of a
parking area, grocery store, restaurant, home, and so on, as
shown in Figure 1-(b).
Mapped Dynamic Intersection (Mapped DI): A Mapped
Dynamic Intersection may be stored in some map databases
as a collision point, such as a Single-track Lane and a Merge
Point, as shown in Figure 1-(c) and (d), respectively. The
collision point is typically not managed by traffic lights.
Moving Dynamic Intersection (Moving DI): A Moving
Dynamic Intersection appears when the vehicle tries to change
lanes, for lane-changing maneuvers and for passing maneu-
vers, as shown in Figure 1-(e) and (f). When a vehicle cannot
start the maneuver, the vehicle might wait for a few seconds
while moving forward and try again. Hence, the spatial area
of the dynamic intersection moves.
Variegated Dynamic Intersection (Variegated DI): A
Variegated Dynamic Intersection represents an area having
multiple dynamic intersections, such as a Center Turn Lane.
A Center Turn Lane might contain multiple dynamic intersec-
tions for turning left through traffic and merging right. Such
an intersection is shown in Figure 1-(g).
Inter-vehicle interaction and negotiation are essential for
self-driving vehicles at dynamic intersections to ensure safety
and enable higher throughput. We recommend the use of
Vehicle-to-Vehicle (V2V) communications and local percep-
tion at these intersections. Vehicle-to-Infrastructure (V2I) com-
munications to handle these intersections is not practical due
to financial and practicality considerations in our view.
In the rest of this paper, without loss of generality, we will
use the dynamic intersection shown in Figure 2 to present,
illustrate, and evaluate our protocol. This helps us simplify
the presentation and adhere to page limits.
In our protocol, we will use the priority assigned to lanes
LHand LL, as shown in Figure 2. This priority assignment
follows the existing traffic rules for human-driven vehicles.
Around dynamic intersections, the vehicles on lane LHhave
precedence, and the vehicles on lane LLhave to yield in the
presence of vehicles on lane LH.
C. Existing Traffic Control Methods
In this section, we discuss existing traffic control methods
for the intersections we see in practice. First, traffic lights are
widely used at stationary road intersections. In this method,
vehicles can enter the intersection when the corresponding
light is green. To enhance traffic throughput, a feasible single
timing plan is often used. In the most common approach in the
US [1], the effective cycle length is estimated, and the cycle
length is split into a green time segment for each phase. The
calculation for the effective cycle length Cand green time g
is described in Eqs. (1) and (2).
C=L
1−CS
RS
(1)
g=CV
CS ×(C−L)(2)
Here, Lis the total lost time, usually taken as the sum of the
inter-green periods. CS represents the critical volume that is
the sum of the critical phase volume for each lane. Also, RS
represents the reference sum of phase flow rates representing
the theoretical maximum value that the intersection could
accommodate. CV is the critical volume for a phase. (Our
Cyber Traffic Light to be described in Section III-D will be
based on this scheme.)
Next, a different traffic management scheme is used in
construction zones. According to the USDOT [2], there are
several different approaches to manage vehicles around a
work zone: (i) Flagger Method, (ii) Temporary Traffic Signal
Method, and (iii) Stop or Yield Control Method. In the flagger
method, qualified human flaggers manage traffic around the
work zone. In the temporary traffic signal method, traffic
signals are temporarily added and used to control traffic
movements. Signal period and timing are often manually set
beforehand. In the stop and yield method, the side that is
closed should yield or stop for oncoming traffic from the
open side. We will compare these schemes against our traffic
protocol in Section IV.
III. OUR CO OP ER ATIV E DYNAMIC INTERSECTION
PROTO CO L
In this section, we present our cooperative vehicle protocol
for dynamic intersections. The protocol is designed to avoid
vehicle collisions at dynamic intersections while maintaining
traffic throughput.
First, we use dynamic vehicle states to guide the behaviors
and the communications to be used by each vehicle. Each
vehicle is in a vehicle state Φstate that dynamically transitions,
and is used to command the vehicle to navigate the dynamic
intersection while accounting for the surrounding environment,
congestion, and waiting time.
Secondly, to assure safety, our protocol uses request-
response-based negotiation around the dynamic intersection.
We assume that there are two neighboring lanes, LHand LL,
with the lane having the dynamic intersection being LH. All
negotiations are initiated from the vehicles on lane LLwith
their requests. The vehicles on lane LHjust respond to the
received requests.
Thirdly, we will use a Cyber Traffic Light (CyberTL)1
which works as a self-planning, self-organizing, and self-
optimizing traffic manager for a dynamic intersection when
congestion arises. Once a CyberTL allocates the green period
for vehicles coming from the two directions, other surrounding
vehicles follow the time allocation and traverse the dynamic
intersection within the allocated green period.
Our protocol uses V2V communications and the perception
systems on each vehicle. All negotiations among multiple self-
driving vehicles is conducted using vehicular communications.
The perception systems are used to improve the assurance of
collision-free traversal through the dynamic intersection.
Self-driving vehicles use Advanced Safety Messages for
V2V communications, and we use the second optional part
of Basic Safety Messages (BSM) for communicating protocol
information [13]. Therefore, Advanced Safety Messages are
broadcast at 10 Hz, the same rate as the BSM.
Unlike human-driven vehicles, self-driving vehicles using
our protocol do not have to stop before a dynamic intersection
unless necessary for safety reasons. Therefore, as will be seen
in Section IV, the traffic throughput of our protocol is much
better than that of existing schemes.
A. Detection of Dynamic Intersections
By definition, conflict points exist at a dynamic intersection.
A self-driving vehicle can detect the presence of a dynamic
intersection using its map database showing other vehicle
paths that intersect with its own when making turns, and/or
by detecting the presence of anomalies like lane closures or
work zones. Our protocol will then be initiated by the vehicle
looking to intervene and use the dynamic intersection.
B. Message Set
In our protocol, each vehicle uses 6 types of safety messages
to interact with other vehicles within its communication range.
DI REQUEST: A DI REQUEST message indicates that the
vehicle has a dynamic intersection along its path. The message
contains 5 parameters: Location of the dynamic intersection,
Estimated Arrival Time, Estimated Exit Time, and Current
Lane.
DI APPROVE: A DI APPROVE message is used by a
vehicle to respond to the DI REQUEST message in order to
acknowledge the requested maneuver of the vehicle on the
other lane. The message contains Location of the dynamic
intersection, Estimated Arrival Time, and Estimated Exit Time.
DI INTERRUPT: A DI INTERRUPT message is used to
potentially stop the vehicles coming from the other lane. The
message is used to accomplish Cyber Traffic Light functions.
The message contains Location of the dynamic intersection,
Cycle Length, and Allocated Green Periods.
DI YIELD: A DI YIELD message is used to respond
to the DI INTERRUPT message, and it indicates that the
transmitter vehicle allows the receiver to enter the dynamic
intersection. The YIELD message contains Location of the
1The name Cyber Traffic Light implies that this traffic light is NOT
physical, but only exists as a software abstraction at the dynamic intersection.
dynamic intersection, Cycle Length, and Allocated Green
Periods.
DI DECLINE: A DI DECLINE message is used for the re-
sponse to the DI REQUEST message and the DI INTERRUPT
in order to decline the requested maneuver of the vehicle on
the neighboring lane.
DI CROSS: A DI CROSS message indicates that the
transmitter vehicle is inside the dynamic intersection. This
message contains Location of the dynamic intersection and
Estimated Exit Time.
The DI CROSS message is not absolutely required because
the perception system may be sufficient to assure the safety
of the dynamic intersection, but it is used in order to enhance
the reliability and safety of our protocol.
C. Vehicle State Transitions
We now present our cooperative dynamic intersection pro-
tocol for vehicles to navigate themselves safely through a
dynamic intersection. To maintain high traffic throughput
while satisfying safety requirements and keeping the worst-
case waiting time at a dynamic intersection short, vehicles
use the state transition diagram shown in Figure 3.
As shown in Figure 3, there are 8 vehicle states. Vehicles on
lane LLcan enter the 3 states with a dark blue border. Also,
vehicles on lane LHcan enter the other 3 states. Any vehicles
not near the dynamic intersection are in the Not Around DI
state. When the vehicle approaches the dynamic intersection,
its state transitions to the Approaching DI state. Also, when
the vehicle receives the DI REQUEST packet, its state is
changed from the Not Around DI state to the Queried state.
The behaviors of the vehicles within each state are presented
next.
The vehicle state transitions of our dynamic intersection
protocol are captured in Algorithm 1 and Algorithm 2. Here,
Algorithm 1 presents the protocol for the vehicle on lane LL,
and Algorithm 2 presents the protocol for the vehicle on lane
LH.
Now, we present the other notation that we use in this paper.
•τt
H: Allocated green period calculated at time tfor the
lane LH.
•τt
L: Allocated green period calculated at time tfor the
lane LL.
•τmax: Maximum value for allocated green periods, τt
H
and τt
L.
Figure 3: State Transition used by Vehicles at Dynamic
Intersections.
Algorithm 1: Protocol for vehicle on lane LL
1while DI is in the near future path do
2if Not halting then
3Φstate =Approaching DI;
4else
5Measure wait time;
6if wait time > τt
H&Φstate =
CyberTL-calculating then
7Φstate =Turn-starting;
8else if wait time > δ &Φstate = Approaching DI
then
9Φstate =CyberTL-calculating;
10 else
11 Φstate =Approaching DI;
12 Keep halting before the Dynamic Intersection;
13 end
14 end
15 end
•τmin: Minimum value for allocated green periods, τt
Hand
τt
L.
•δ: Time threshold for state transition from Approaching
DI state to CyberTL-calculating state.
•LDI : Total lost time.
•CDI : Cycle length.
•DDI : Length for the Dynamic Intersection.
•lv: Vehicle length.
•vDI : Permitted velocity around the Dynamic Intersection.
•s0: Base saturation flow rate.
•fa: Area type adjustment factor.
•NH: Number of vehicles coming from the lane LH.
•NL: Number of vehicles coming from the lane LL.
•vH: Average velocity of moving vehicles on the lane LH.
•vL: Average velocity of moving vehicles on the lane LL.
•DH: Distance to the farthest vehicle on the lane LH.
•DL: Distance to the farthest vehicle on the lane LL.
•ψH: Estimated traffic volume for the lane LH.
•ψL: Estimated traffic volume for the lane LL.
•ST (τt
H): Switching time for the phase allocated τt
H.
•ST (τt
L): Switching time for the phase allocated τt
L.
Here, the values of τmin and δare set to satisfy τmin > δ
to meet safety requirements. That is, we always have τt
H> δ
and τt
L> δ.
1) Not Around DI state: When the vehicle is not near a
dynamic intersection and has no request from other vehicles,
the vehicle is in the Not Around DI state. When the vehicle
approaches a dynamic intersection, the vehicle state Φstate
transitions to the Approaching DI state (See Section III-C2).
Also, when the vehicle receives a DI REQUEST packet, the
state is changed to the Queried state (See Section III-C5).
2) Approaching DI state: Under the Approaching DI state,
the vehicle keeps sending2the DI REQUEST packet. When
the vehicle receives the DI APPROVE packet from other
vehicles, the vehicle starts to enter the dynamic intersection
2This means that the message will be sent out regularly at the 10 Hz rate
of DSRC Advanced Safety Messages.
Algorithm 2: Protocol for vehicle on lane LH
1if DI REQUEST is received then
2Φstate =Queried;
3else if DI INTERRUPT is received then
4if cannot stop safely then
5DI DECLINE is sent;
6else
7Halt before the DI;
8Φstate =CyberTL-counting;
9end
10 else
11 Φstate =Not Around DI;
12 end
and transitions to the Crossing state (See Section III-C8). If a
vehicle has been waiting for more than the time period δ, its
state transitions to the CyberTL-calculating state (See III-C3).
3) CyberTL-calculating state: A vehicle in the CyberTL-
calculating state calculates the green period for each direc-
tion. Once the vehicle completes the calculation, the state is
changed to the Turn-starting state (See Section III-C4). The
calculation method is presented in Section III-D.
4) Turn-starting state: In the Turn-starting state, the vehi-
cle measures its wait time before the dynamic intersection
until the time τt
Hpasses. After the time τt
Hpasses, the
vehicle keeps sending the DI INTERRUPT packet. When the
vehicle can confirm that no vehicles are within the dynamic
intersection, the vehicle enters the intersection and transitions
to the Crossing state.
5) Queried state: When a vehicle in the Not Around DI
state receives the DI REQUEST packet, the vehicle state
Φstate transitions to the Queried state. In the Queried state,
the vehicle estimates the potential of conflict with the sender.
The vehicle calculates the Arrival Time and the Exit Time of
the requested dynamic intersection. Then, the vehicle replies
with either the DI DECLINE packet or the DI APPROVE
packet, and the state returns back to the Not Around DI state.
6) CyberTL-counting state: When the vehicle receives a
DI INTERRUPT packet, its state is changed to CyberTL-
counting state. In the CyberTL-counting state, the vehicle has
to stop before the dynamic intersection for the time period
τt
L. During the time period τt
L, the vehicle keeps sending the
DI APPROVE packet. After the time τt
Lpasses, the vehicle
state transitions to the Turn-switching state. Also, if all of the
oncoming vehicles traverse the dynamic intersection before the
time period τt
Lpasses, the vehicle starts to enter the dynamic
intersection.
7) Turn-switching state: Under the Turn-switching state,
the vehicle keeps sending the DI INTERRUPT packet to
request the vehicles on the conflicting lane to stop before the
dynamic intersection. When the dynamic intersection becomes
clear, the vehicle enters the dynamic intersection and its state
is changed to the Crossing state.
8) Crossing state: When the vehicle traverses the dynamic
intersection, the vehicle is in the Crossing state and keeps
sending the DI CROSS message. Once the vehicle exits the
dynamic intersection, it transitions to the Not Around DI state.
Next, we present how our Request-response-based Nego-
tiation proceeds among multiple self-driving vehicles in our
protocol. The negotiation sequence is shown in Figure 4 and
Figure 5. First, in Figure 4, a vehicle arrives at the dynamic
intersection from one direction, LHor LL. As shown in
Figure 4-(a), when the vehicle comes from lane LL, the
vehicle transmits the DI REQUEST packet to the surrounding
vehicles. The vehicle also uses its perception system to confirm
safety traversed within the dynamic intersection. On the other
hand, as shown in Figure 4-(b), when the vehicle comes
from lane LH, the vehicle does not transmit the Advanced
Safety Message, because its lane has precedence. Overall, all
negotiations are initiated from the vehicles coming from lane
LL. Unlike human-driven vehicles under current traffic rules,
self-driving vehicles under our protocol do not have to stop
and/or slow down before the dynamic intersection.
Suppose that, as shown in Figure 5, vehicles come to the
dynamic intersection from two different directions, LHand
LL, negotiation based on the request-response is initiated by
the vehicle on lane LL. The vehicle from lane LLsends
the DI REQUEST message (at t=t0). Then, the vehicle
on lane LHcalculates its Arrival Time and Exit Time. The
two vehicles determine that they have the potential to collide.
The vehicle on lane LHsends the DI DECLINE message
in order to decline the request (at t=t1). In this case, the
vehicle from lane LLstops before the dynamic intersection
and keeps sending the DI REQUEST packet (at t=t2).
Once the vehicle on lane LHcrosses the dynamic intersection,
the vehicle on lane LLconfirms safe clearance and enters the
dynamic intersection if there is no oncoming traffic (at t=t3).
A key challenge of this request-response-based protocol is
that the vehicle on lane LLmight wait a long time before the
dynamic intersection when there is a high volume of traffic
coming from the other direction. If vehicles keep arriving from
lane LH, the waiting vehicle on lane LLwould keep receiving
the DI DECLINE packet and keep waiting. To address the
issue explicitly, we use the notion of a Cyber Traffic Light
(CyberTL). The details of the CyberTL are presented next.
A simulation video of our request-response-based protocol
can be seen at:
https://www.youtube.com/watch?v=xkfyAkXzAXg
D. Cyber Traffic Light
In this section, we present the Cyber Traffic Light (Cy-
berTL) that is used to keep the worst-case waiting time short
while maintaining high traffic throughput. The CyberTL works
as a self-optimizing traffic light by estimating the surrounding
traffic volume and by coordinating with oncoming vehicles.
The CyberTL is a V2V communication-based traffic manager,
(a) Only ωLcomes (b) Only ωHcomes
Figure 4: Safety Confirmation for Dynamic Intersection.
(a) t=t0(b) t=t1
(c) t=t2(d) t=t3
Figure 5: Negotiation based on the Request-response.
(a) t=t0−δ
(b) t=t0−δ+τt0
H
(c) t=t0−δ+τt0
H+ST (τt0
H)
(d) t=t1
(e) t=t1+ ∆
Figure 6: DI management with CyberTL.
and we use the OBU of each self-driving vehicle to support
this distributed software abstraction.
By using the CyberTL, the vehicles at a dynamic intersec-
tion coming from two different directions take turns. That is,
once nvehicles have passed the dynamic intersection from
one direction, mvehicles from the other direction can enter
and traverse the dynamic intersection (n, m ∈N).
The vehicles that together constitute the CyberTL are
grouped into 3 categories: (i) CyberTL Calculator; (ii) Cy-
berTL Counter; and (iii) CyberTL Follower. First, the first
waiting vehicle on lane LLbecomes the CyberTL Calculator
when its vehicle state Φstate is changed to the CyberTL-
calculating state. The CyberTL Calculator estimates the arrival
traffic volumes, and calculates the green periods, τt
Hand τt
L,
along the two directions for one cycle. Here, trepresents
the time when the green periods are estimated. Once the
waiting time of the CyberTL Calculator exceeds τt
H, the
vehicle transmits the DI INTERRUPT packet and starts the
maneuver. When the vehicle coming from lane LHreceives
the DI INTERRUPT and replies with the DI YIELD packet,
its state transitions to the CyberTL-counting state. Any vehicle
in the CyberTL-counting state has to stop and has to measure
the waiting time acting as a CyberTL Counter. The CyberTL
Counter vehicle keeps sending the DI YIELD message during
τt
L, and once the waiting time exceeds τt
L, the vehicle starts
to send the DI INTERRUPT packet. The other surrounding
vehicles, CyberTL Followers, just follow the decisions made
by the CyberTL Calculator and CyberTL Counter. Overall,
the green period allocation is calculated by the CyberTL
Calculator, and the time measurements are carried out by the
CyberTL Calculator and CyberTL Counter.
The dynamic intersection management scheme with the
CyberTL is illustrated in Figure 6. Also, an example timeline
of the dynamic intersection use is presented in Figure 7. In this
example, the CyberTL Calculator is assigned for the vehicle
ωL1at time t0and vehicle ωL5at time t1. Also, the CyberTL
Counter is assigned for the vehicle ωH5. At t=t0, when
the duration δis passed after its arrival, the vehicle ωL1
calculates the green periods τt0
Hand τt0
L. During the period τt0
H,
vehicles on lane LHcan enter the dynamic intersection. The
Switching Time duration is the required time period to reverse
the direction of traffic flow, and is equal to the time taken from
the beginning of the DI INTERRUPT message transmission
to the dynamic intersection clearing. This duration cannot be
estimated beforehand, because it depends on the abilities and
physical sizes of the traversing vehicles.
The green periods τt
Hand τt
Lare calculated based on the
real-time traffic volumes and on the metrics that are widely
used in existing traffic systems [1]. In the protocol, we first
estimate the real-time traffic volumes from the surroundings
using V2V communications, and we calculate the approximate
total lost time from the dynamic intersection length and vehicle
size.
First, connected self-driving vehicles estimate the arrival
traffic volumes, ψHand ψL, based on the information available
from the received BSM, as shown in Eqs. (3) and (4). Here,
NHand NLrepresent the number of vehicles on lane LHand
on lane LLwithin the communication range, respectively. vH
and vLare the average velocity of the vehicles on each lane
within the communication range. DHand DLrepresent the
distance from the transmitter to the farthest vehicle on lane
LHand on lane LL, respectively. Since the transmitter might
not know the accurate boundaries of the communication range,
we use these two distances for estimation purposes.
ψH=NH·vH
DH
(3)
Figure 7: Allocated Green Period for CyberTL.
ψL=NL·vL
DL
(4)
Then, to calculate the appropriate cycle length C, the
CyberTL Calculator vehicle estimates the approximate total
lost time LDI as shown in Eq. (5). Here, the total lost time is
the sum of the Switching Time in a cycle. Since an accurate
Switching Time cannot be estimated accurately beforehand, we
use the size of the CyberTL Calculator vehicle lv, the length
of the dynamic intersection DDI , and the permitted velocity
vDI .
LDI =2×(lv+DDI )
vDI
(5)
Based on Eqs. (3), (4), and (5), the appropriate cycle length
Cis estimated as shown in Eq. (6).
CDI =LDI
1−(ψH+ψL
0.9·s0·fa)(6)
Here, we use the two variables, the Base Saturation Rate
s0and Area Adjustment Factor fa. According to the National
Research Council [1], s0is set as 1,900, and fais set to 1.0
as default and set to 0.9if the area is in a central business
districts.
Finally, the green periods, τt
Hand τt
L, are calculated by
using the cycle length and total lost time, as shown in Eqs. (7)
and (8). The allocated green periods reflect the arrival traffic
volumes from the two directions.
τt
H=ψH
ψH+ψL
·(CDI −L)(7)
τt
L=ψL
ψH+ψL
·(CDI −L)(8)
The allocated green periods τt
Hand τt
Lare checked against
reasonable minimum and maximum values, τmin and τmax, in
order to avoid vehicles having to wait for a long time before
the dynamic intersection. Also, the cycle length must not
exceed a maximum allowable value set by the local jurisdiction
(such as 150 sec) and it must be long enough to serve at least
one vehicle.
A simulation video of our Cyber Traffic Light can be seen
at:
https://www.youtube.com/watch?v=xlFZgofh2iw
E. Anomalies and Packet Loss
Our dynamic intersection protocol assures safety by using
V2V communications and the perception systems coopera-
tively. It inherently accommodates the loss of transmitted
packets, since safety messages are transmitted at 10 Hz and
occasional packet losses are compensated by subsequent suc-
cessful transmission. In addition, our request-response-based
negotiation enhances safety by its design.
In fact, the loss of the DI REQUEST or the DI IN-
TERRUPT messages makes the waiting time longer for the
transmitter, but it will not cause a vehicle collision or a
deadlock because the sender vehicle keeps waiting before the
dynamic intersection until it receives the appropriate response.
In addition, when the DI APPROVE or the DI YIELD
message is lost, the receiver might lose the opportunity to
enter the dynamic intersection, but it again cannot cause a
vehicle collision. However, the waiting time of the receiver
vehicle might be longer. Moreover, the packet loss of the DI
DECLINE message poses no safety risk because the receiver
vehicle has to keep waiting until it receives the appropriate
response when there are other vehicles around the dynamic
intersection.
Finally, in addition to these packets, our protocol uses the
DI CROSS message to inform that the transmitter vehicle
is inside the intersection area. The vehicles using our traffic
protocol cannot enter the intersection if they receive a DI
CROSS message, meaning the message is used to enhance
the safety and reliability around the intersection.
IV. EVALUATION
In this section, we present the implementation and evalua-
tion of our cooperative dynamic intersection protocol in our
hybrid simulator-emulator named AutoSim. We evaluate the
traffic protocol in terms of traffic throughput and compare
against two baseline traffic control methods: (i) Temporary
Traffic Signal Method and (ii) Stop or Yield Control Method.
AutoSim is an extended version of Groovenet [15] with
3-D graphics and other capabilities. AutoSim consists of vari-
ous core models, including mobility, communication, control,
localization, and pose estimation for each simulated vehicle.
In AutoSim, each simulated vehicle can transmit and receive
a Basic Safety Message (BSM) which is one of the main
requirements of vehicular communications. The message for-
mat used follows the definition of the SAE J2735 standard,
which is the DSRC Message Set Dictionary. In addition,
AutoSim is a modular hybrid simulator-emulator for vehicular
communications that enables interactions between virtual and
real vehicles equipped with DSRC radios.
A. Metric
In order to evaluate the utility of our protocol, we define the
Trip Time as the time taken by a vehicle to go from a known
start-point before the dynamic intersection to a known end-
point after the intersection. We calculate the trip time for each
simulated car and compare that against the trip time taken by
the car assuming that it stays at a constant speed and does
not stop at the dynamic intersection. The difference between
these two trip times is considered to be the Trip Delay due
to the dynamic intersection, and the Average Trip Delay and
Figure 8: Map for One-way Alternative Use.
Figure 9: Average waiting time for symmetric traffic volume.
Figure 10: Worst waiting time for symmetric traffic volume.
the Worst Trip Delay across all vehicles in each simulation are
compared in our evaluation.
B. Scenarios
Since dynamic intersections in the real world take a variety
of forms and sizes, we restrict the experimental settings to two
specific kinds of dynamic intersections: One-way Alternative
Use (Temporary DI) and Turning Left through Traffic (Un-
mapped DI). We assume that all vehicles calculate the required
trajectories to traverse the dynamic intersection by themselves.
In addition, the traffic generation follows the Poisson random
distribution. Each simulation runs for 30 minutes, and we
evaluate the time delay of the vehicles that are generated
during the last 20 minutes to bypass initial observation and
measure steady-state behavior.
Our simulations evaluate our protocol by comparing against
two baseline protocols: Temporary Traffic Light Method and
Stop or Yield Control Method.
Temporary Traffic Light Method Traffic signals may be
temporarily used to control vehicular traffic movements in
Temporary DI, such as the areas requiring alternating one-
way traffic operations. In the simulations, the time length for
Figure 11: Trip Delay for Asymmetric Traffic Volume.
a green light is set as 30 seconds, and the time length for a
yellow light is set as 5 seconds.
Stop or Yield Control Method All vehicles follow the
priority given to the lane. Specifically, vehicles on a higher-
priority lane are not required to stop or slow down at the
dynamic intersection. Correspondingly, vehicles on a lower-
priority lane have to stop before the dynamic intersection for
guaranteeing safety. In the Temporary DI, a stop or yield sign
might be temporarily installed on roads where one side of the
road way is closed. The side that is closed should stop for and
yield to oncoming traffic on the side that is open.
We assume that there is no obstacle completely blocking
perception around the dynamic intersections.
The variables for the simulations are set by accounting for
the DSRC standards [9], [13] and typical vehicle abilities, as
shown in Table I. For example, the communication range and
frequency for V2V communications are set as 500 m and 10
Hz, respectively.
C. Traffic Throughput Evaluation
We run simulations for evaluating the traffic throughput of
our cooperative dynamic intersection protocol and compare
against the baseline protocols. To measure the traffic through-
put, we focus on two specific types of dynamic intersections.
1) One-way Alternative Use (Temporary DI): We evaluate
the traffic throughput for One-way Alternative Use (Temporary
DI). The length of the dynamic intersection (DI length) is
changed from 15 m to 60 m. In Figure 8, we present the map
of the dynamic intersection. We first assume that the traffic
volumes from two directions are symmetric, and then evaluate
the throughput when the traffic volumes are asymmetric.
For the symmetric traffic volume, we evaluate the average
trip delay and worst trip delay when the range of the traffic
volume for each lane is changed from 100 to 600 vehicles
per hour. In Figure 9, we present the average trip delay of
all the vehicles, and Figure 10 presents the worst trip delay of
all the vehicles. Our cooperative dynamic intersection protocol
has the shortest average trip delay and the shortest worst trip
Table I: Environmental settings for the experiments.
Communication range 500 (m)
Communication frequency 10 (Hz)
Max speed of vehicles 30 (km/h)
Vehicle length l4.5 (m)
Vehicle width w1.7 (m)
Figure 12: Worst waiting time for Turning Left through Traffic.
delay in all cases. The Temporary Traffic Light Method has
a constant and long average delay because all of the arriving
vehicles have to wait for the green period and the delay is
not related to the intersection size. When the traffic volume
is high, the Stop or Yield Control Method has the longest
average trip delay and worst trip delay, because the vehicles
on the lower-priority lane can enter the dynamic intersection
only when there is no oncoming traffic.
Second, as shown in Figure 11, we evaluate the average
trip delay and worst trip delay when the traffic volume from
the two directions is asymmetric. Here, we evaluate the trip
delay when the {traffic volume from LH,traffic volume from
LL}={200 [V/H], 400[V/H]},{400 [V/H], 200[V/H]}. In all
cases, our dynamic intersection protocol still has the shortest
average trip delay and the shortest worst trip delay. Since our
CyberTL dynamically allocates the green period depending on
the short-term arrival traffic volume, the average trip delays
of {200 [V/H], 400[V/H]}and of {400 [V/H], 200[V/H]}are
almost the same. On the other hand, under the Stop or Yield
Control Method, when we increase the traffic volume on the
lane LH, the worst trip delay becomes much larger.
2) Turning Left through Traffic (Unmapped DI): In addi-
tion, we evaluate the worst trip delay when vehicles turn left
through oncoming traffic. Since this dynamic intersection is
an Unmapped DI, not due to construction, we compare our
protocol only to the Stop or Yield Control Method. The traffic
volume of the oncoming lane LHis varied from 100 to 1,200
vehicles per hour. We assume that the traffic volume on lane
LLis 50 or 100 vehicles per hour, and there are no vehicles
going straight on lane LL.
In Figure 12, we present the worst trip delay of all the vehi-
cles around the dynamic intersection. Our dynamic intersection
protocol again has the shortest worst trip delay in all cases.
Under the Stop or Yield Control Method, all vehicles on lane
LLhave to slow down and halt before the dynamic intersection
to guarantee safety. Therefore, even when the traffic volume on
lane LHis relatively low, the Stop or Yield Control Method
has longer worst trip delay. Moreover, when the oncoming
traffic volume is high, under the Stop or Yield Control Method,
some vehicles on lane LLhave to wait a very long time before
the dynamic intersection. On the other hand, when the traffic
volume is high, our protocol uses the Cyber Traffic Light to
allow the waiting vehicles to use the dynamic intersection.
V. REL ATED WO RK
Since the DARPA Urban Challenge in 2007 [18], a com-
petition for autonomous driving vehicles, cooperation and
collaboration among self-driving vehicles has been widely
studied for intelligent intersection management, platooning,
merging, and lane-changing maneuvers.
First, Dresner and Stone proposed a multi-agent approach
for an intersection control, named Autonomous Intersection
Management [10]. In this system, all vehicles call ahead to
a reservation manager agent at the intersection to reserve
conflict-free trajectories. Since the protocol relies on Vehicle-
to-Infrastructure communications and on the manager agent,
the infrastructure might be a single point of failure for the
intersection management system. The authors extended the
system to heterogeneous traffic environments [5], in which
vehicles are assumed to be partially automated.
Secondly, as a priority-based intersection protocol, Azimi
et al. [7], [8] proposed several spatio-temporal intersection
protocols (STIP). In these protocols, vehicles are assigned
priorities based on their arrival times at the intersection,
with vehicles reaching the intersection earlier assigned higher
priorities than vehicles coming later. A vehicle with trajectory
conflicts will come to a complete stop before entering the
intersection, and waits until all vehicles having higher priority
than this vehicle have crossed the conflicting areas. Under
STIP, all vehicles exchange the expected arrival time and
required trajectory with Vehicle-to-Vehicle (V2V) communi-
cations before reaching the intersection. These priority-based
intersection protocols were extended to merge points, where
two lanes with different priorities meet [4]. In this work,
vehicles safely traverse merge points by using both vehicular
communications and their own perception systems. The merge
protocol also supports zipper merge where the vehicles from
different lanes merge at the merge points by taking turns when
the traffic density is high. In addition, the protocol enables
cooperation and collaboration between self-driving vehicles
and human-driven vehicles. Gregoire et al. [11] proposed
priority-based scheduling techniques for multi-robot planning
and cooperation, and Zhang et al. [20] extended the work
for self-driving vehicles. In this work, a priority scheduling
algorithm is designed based on the modeling of vehicular
spatio-temporal relations, for coordinating multiple vehicles
at intersections.
Thirdly, a novel class of protocols called synchronous inter-
section protocols, BRIP [6] and CSIP [3], were presented. Un-
der these synchronous intersection protocols, all self-driving
vehicles cross the intersection without stopping by following a
strict but well-defined spatio-temporal pattern tailored to each
intersection. BRIP and CSIP allow vehicles to efficiently and
continuously enter and cross the intersection in a synchronized
manner.
In addition, lane-changing maneuvers have been considered
as a longitudinal motion planning problem [16], [17]. In this
work, automated vehicles select an appropriate inter-vehicle
traffic gap and time instance to perform the maneuver by
estimating whether there is a longitudinal trajectory allowing
the vehicles to complete the maneuver. No cooperation is
assumed, expected, or accomplished.
The above approaches have addressed individual driving
contexts, such as simple multi-way intersections, roundabouts,
merging points, and lane-changing maneuver, and they did not
study more complicated situations. On the other hand, our
protocol can be used for many complex driving contexts we
encounter in practice.
VI. CONCLUSION
In this paper, we defined and classified Dynamic Inter-
sections where the paths of vehicles intersect dynamically
and perhaps temporarily. Often, these intersections are not
present in a map database. These dynamic intersections are
common in or near parking lots, multi-level parking garages,
work zones, accidents, potholes, and driveways. Such dynamic
intersections might appear almost anytime and anywhere along
public roads. We presented a cooperative dynamic intersection
protocol for self-driving vehicles to safely traverse dynamic
intersections by using V2V communications and perception
systems.
Under our protocol, connected and automated vehicles are
not required to stop before dynamic intersections when there
is no vehicle coming from other directions. Our protocol
increases the traffic throughput significantly at dynamic in-
tersections, compared to two baseline protocols. The protocol
has the shortest average trip delay and the shortest worst trip
delay in all the cases we considered. Under our protocol, self-
driving vehicles calculate the possible collision points, specify
the dynamic intersections, and manage the spatial area by
using the vehicle-state transition and a Cyber Traffic Light.
The Cyber Traffic Light enhances the traffic throughput around
the dynamic intersections when traffic volumes are high.
We note several limitations of our work. First, our coopera-
tive dynamic intersection protocol must be extended to support
Moving DI and Variegated DI where vehicle behaviors are
more complex, with interactions and effects between multiple
dynamic intersections. Secondly, we have currently assumed
that perception works perfectly, and there is no obstacle
blocking a perception around the dynamic intersection. In
future work, we will design protocols to support limits on
perception. Thirdly, we only consider the protocol for fully-
automated vehicles. Since a long transition period will likely
be needed to replace human-driven vehicles with fully self-
driving vehicles, we will need to design protocols for the
heterogeneous environment where self-driving vehicles and
human-driven vehicles co-exist. Finally, our cooperative dy-
namic intersection protocol assumed that vehicles always have
good GPS, and there are no component failures. In future
work, we will study the feasible protocols to enhance safety
when GPS and/or components fail.
REFERENCES
[1] Highway capacity manual. Transportation Research Board, Washington,
DC, 2000.
[2] Manual on uniform traffic control devices 2009 edition (MUTCD).
United States Department of Transportation, Federal Highway Admin-
istration, 2009.
[3] S. Aoki and R. R. Rajkumar. A configurable synchronous intersection
protocol for self-driving vehicles. In Embedded and Real-Time Comput-
ing Systems and Applications (RTCSA), 2017 IEEE 23rd International
Conference on, pages 1–11. IEEE, 2017.
[4] S. Aoki and R. R. Rajkumar. A merging protocol for self-driving
vehicles. In Cyber-Physical Systems (ICCPS), 2017 IEEE/ACM Third
International Conference on, pages 219–228. ACM, 2017.
[5] T.-C. Au, S. Zhang, and P. Stone. Autonomous intersection management
for semi-autonomous vehicles. In D. Teodorovi’c, editor, Handbook of
Transportation, pages 88–104. Routledge, 2016.
[6] R. Azimi, G. Bhatia, R. Rajkumar, and P. Mudalige. Ballroom in-
tersection protocol: Synchronous autonomous driving at intersections.
In Embedded and Real-Time Computing Systems and Applications
(RTCSA), 2015 IEEE 21st International Conference on, pages 167–175.
IEEE, 2015.
[7] R. Azimi, G. Bhatia, R. R. Rajkumar, and P. Mudalige. Reliable
intersection protocols using vehicular networks. In Cyber-Physical
Systems (ICCPS), 2013 ACM/IEEE International Conference on, pages
1–10. IEEE, 2013.
[8] R. Azimi, G. Bhatia, R. R. Rajkumar, and P. Mudalige. Stip: Spatio-
temporal intersection protocols for autonomous vehicles. In Cyber-
Physical Systems (ICCPS), 2014 ACM/IEEE International Conference
on, pages 1–12. IEEE, 2014.
[9] L. Cheng, B. E. Henty, D. D. Stancil, F. Bai, and P. Mudalige. Mobile
vehicle-to-vehicle narrow-band channel measurement and characteriza-
tion of the 5.9 ghz dedicated short range communication (dsrc) frequency
band. IEEE Journal on Selected Areas in Communications, 25(8), 2007.
[10] K. Dresner and P. Stone. A multiagent approach to autonomous
intersection management. Journal of Artificial Intelligence Research,
31:591–656, March 2008.
[11] J. Gregoire, S. Bonnabel, and A. De La Fortelle. Priority-based
intersection management with kinodynamic constraints. In Control
Conference (ECC), 2014 European, pages 2902–2907. IEEE, 2014.
[12] W. Gruel and F. Piller. A new vision for personal transportation, 2016.
[13] J. B. Kenney. Dedicated short-range communications (dsrc) standards
in the united states. Proceedings of the IEEE, 99(7):1162–1182, 2011.
[14] R. B. Langley. Rtk gps. GPS World, 9(9):70–76, 1998.
[15] R. Mangharam, D. Weller, R. Rajkumar, P. Mudalige, and F. Bai.
Groovenet: A hybrid simulator for vehicle-to-vehicle networks. In
Mobile and Ubiquitous Systems: Networking & Services, 2006 Third
Annual International Conference on, pages 1–8. IEEE, 2006.
[16] J. Nilsson, M. Br¨
annstr¨
om, E. Coelingh, and J. Fredriksson. Lane change
maneuvers for automated vehicles. IEEE Transactions on Intelligent
Transportation Systems, 18(5):1087–1096, 2017.
[17] J. Nilsson, J. Silvlin, M. Brannstrom, E. Coelingh, and J. Fredriksson. If,
when, and how to perform lane change maneuvers on highways. IEEE
Intelligent Transportation Systems Magazine, 8(4):68–78, 2016.
[18] C. Urmson, J. Anhalt, D. Bagnell, C. Baker, R. Bittner, M. Clark,
J. Dolan, D. Duggins, T. Galatali, C. Geyer, et al. Autonomous driving
in urban environments: Boss and the urban challenge. Journal of Field
Robotics, 25(8):425–466, 2008.
[19] J. Wei, J. M. Snider, J. Kim, J. M. Dolan, R. Rajkumar, and B. Litkouhi.
Towards a viable autonomous driving research platform. In Intelligent
Vehicles Symposium (IV), 2013 IEEE, pages 763–770. IEEE, 2013.
[20] K. Zhang, D. Zhang, A. de La Fortelle, X. Wu, and J. Gregoire. State-
driven priority scheduling mechanisms for driverless vehicles approach-
ing intersections. Intelligent Transportation Systems, IEEE Transactions
on, 16(5):2487–2500, 2015.