A Two-Stage Optimization Method For Schedule and Trajectory of CAVs at An Isolated Autonomous Intersection

A Two-Stage Optimization Method for Schedule and Trajectory of CAVs at an Isolated
Autonomous Intersection
Zhihong Yao1,2,3, Haoran Jiang1,2*, Yangsheng Jiang1,2,3*, and Bin Ran4
1. School of Transportation and Logistics, Southwest Jiaotong University, Chengdu, Sichuan 610031, China;
2. National Engineering Laboratory of Integrated Transportation Big Data Application Technology,
Southwest Jiaotong University, Chengdu, Sichuan 611756, China;
3. National United Engineering Laboratory of Integrated and Intelligent Transportation, Southwest Jiaotong
University, Chengdu, Sichuan 611756, China;
4. Department of Civil and Environmental Engineering, University of Wisconsin-Madison, Madison, WI
53706, USA.
Abstract
Autonomous intersection management has become a state-of-the-art control strategy customized for
connected and autonomous vehicles. Combining the advantages of tile-based and conflict point-based
approaches, this paper proposes a two-stage optimization method based on a developed intersection
modeling approach. The first stage is a timing schedule optimization model, assigning vehicle arrival
times at an intersection. Based on the output of the first stage, the second stage is a trajectory optimization
model, which gives the eco-driving strategies. Moreover, a rolling optimization with a variable cycle
length is adopted to run the method continuously. Simulation results show that the proposed method
outperforms the genetic algorithm-based method in terms of computation time, and can reduce vehicle
delay and fuel consumption by 89.48% and 46.84%, respectively, under different traffic demands
compared to the first-come-first-serve method. Furthermore, the performance of the proposed method
under asymmetric traffic demand is discussed. Sensitivity analyses suggest that (1) a long cycle length
benefits the proposed method within certain limits and (2) a proper deceleration within the intersection
can balance traffic delay with fuel consumption. In addition, an additional model with a heuristic rule is
compared with the original timing schedule optimization model. It is found that reducing binaries in the
first stage can make a tradeoff between the quality of the solution and efficiency, which can be used in
conjunction with long cycles.
Keywords: Connected and autonomous vehicles, optimal scheduling, MILP, autonomous intersection
control, trajectory planning
*Correspondence to: Haoran Jiang (jianghaoran.joe@foxmail.com) and Yangsheng Jiang (jiangyangsheng@swjtu.edu.cn),

School of Transportation and Logistics, Southwest Jiaotong University, Chengdu, Sichuan 610031, China.
1
I. INTRODUCTION
A. Background
With the booming development of the 5G, big data, and other emerging technologies, connected
and autonomous vehicles (CAVs) will gradually replace traditional ones [1]. Compared with traditional
vehicles, CAVs can communicate and execute precise control schemes in a connected automated
environment [2]. Because of this, non-stopping through intersections can be achieved, significantly
reducing vehicle delay while maintaining the continuity of traffic flow.
The role of traffic signals at traditional intersections is to sequentially assign right-of-way (ROW)
to vehicles in different directions to achieve spatial and temporal separation of conflicting traffic. Many
scholars optimize signal timings and precisely control the CAV trajectory to ensure that vehicles arrive
at the intersection at green phases [3]–[7]. However, these methods do not allow conflicting traffic to
present simultaneously at the intersection. Thus, the intersection space is not fully utilized. Relying on
vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) technologies, the intersection manager (IM),
which assigns ROW to individual vehicles, will take the place of traffic signals [8]. As a result, vehicles
can intersperse through intersections individually. This IM-led control mode is called autonomous
intersection management (AIM), a state-of-the-art strategy customized for CAVs.
B. Literature Review
AIM assigns times for vehicles to arrive at the intersection to avoid conflicts. Dresner and Stone
[9] proposed the earliest AIM, which is a reservation-based method. They divided the intersection into a
grid of tiles with n × n granularity, where each tile can be reserved by one vehicle per time step. The
vehicle sends a request for an arrival time and reserves tiles on its desired path. The IM simulates the
request and accepts it when the reserved tiles are not occupied by others. Based on the reservation idea,
studies emerged [10]–[12]. Liu et al. [10] proposed a hybrid strategy of cooperative resource reservation.
A trajectory planning-based mechanism was developed and the results showed that the mechanism has
a shorter evacuation time and more significant throughput than fixed lights and adaptive lights. Fajardo
et al. [11] and Li et al. [12] proved that the reservation system outperforms signal control in traffic delays
through simulations. Further, Yu et al. [13] transformed the reservation-based model into an
optimization-based model. Xu et al. [14] focused on the optimal passing order of vehicles in AIM and
turned it into a tree search problem. To overcome the fail-follow problem in AIM, Ma and Li [15]
proposed a time-independent trajectory optimization approach. Despite the changes of the reservation
process, these studies can be categorized as tile-based approaches based on their spatial dispersion [16].
In the tile-based approach, the granularity varies and leads to a trade-off between efficiency and
computation complexity [9]. Since a tile can only be occupied by one vehicle at a time, the intersection
space can be better utilized by decreasing the tile size, which improves the intersection's efficiency but
2
introduces higher complexities for algorithm design [17], [18].
Another category of intersection modeling approaches, independent of granularity, replaces the
tiles as conflict points [19]–[21]. Conflict points are defined as the intersections between different turning
paths. Due to its spatial continuity, the mathematical model constructed by the conflict point-based
approach is simpler and faster to solve [19]. The separation ways at the conflict points are diverse, but
none of them are good at considering the vehicle's shape. For instance, Zohdy and Rakha [20] set a
constant time for vehicles in different approaches as an interval to pass the conflict point, which,
however, does not mathematically consider vehicle length and width. Levin and Rey [19] evaluated the
vehicle length and set different time intervals for different turnings but still ignored the width.
To better consider the vehicle's shape, He et al. [22] extended a conflict point to a region along the
paths, which can be seen as a prototype of the combination of the two kinds of intersection modeling
approaches, but does not consider diversion points and the diversity of different conflicting angles. Lu
et al. [23] allowed vehicles to be present in the conflict region simultaneously by car-following, which
weakens the consideration of vehicle width. In conclusion, none of the studies have combined the
advantages of the two kinds of intersection modeling approaches well.
As for control policies, due to the nature of reservation, the most-used method is rule-based,
represented by first-come-first-serve (FCFS), where the reservation of vehicles is processed in the
sequence of arrivals [24]. It is demonstrated that FCFS-based methods outperform traffic signal control
at symmetrical intersections [11], [12], [22], [24], but generate enormous delays at asymmetric
intersections, oversaturation, and queue driving situations because of the fairness of vehicles [13], [15],
[17], [25].
The optimization-based method emerged to bridge the limitations of rule-based methods and
also changed the way of the reservation to the assignment. Optimization-based methods treat all vehicles
simultaneously to minimize traffic delay and fuel consumption [26]. Compared to the FCFS-based
method, it takes into account the optimization of the vehicle sequence. For example, Choi et al. [17] used
the genetic algorithm (GA) to filter the optimal vehicle passing order. A similar scheme can also be found
in [27]. Lee and Park [28] constructed a nonlinear model to minimize the total length of overlapped
trajectories. Zohdy and Rakha [20] optimized the vehicle trajectory to minimize the traffic delay with
nonlinear constraints. The above studies reduced vehicle delay and fuel consumption to some extent, but
all are computationally challenging. To improve the solution efficiency, the AIM problem is constructed
as a mixed-integer linear programming (MILP). Many researchers build the MILP with tile-based
approaches [13], [15], [29]–[31], with or without considering vehicle turnings, while Levin and Rey [19]
build the model with a conflict point-based approach, which is more amenable to mathematical
programming formulations.
After assigning the arrival time to vehicles, the IM makes travel strategies for them through a
3
trajectory planning method. The trajectory planning method ensures that the vehicle can reach the
intersection on time. Some studies developed driving rules [29], [32], or adopted a car-following model
[20], to plan the speed profile and avoid collisions on the road. With the eco-driving concept, trajectory
planning also aims to reduce traffic emissions [33]. Therefore, many studies have directly aimed at fuel
consumption or energy to develop the optimization model. However, most of them are optimal control
problems [34]–[39], which are difficult to solve and performed individually for each vehicle. Yao et al.
[30] used the infinitesimal method to deal with the optimal control problem and turned it into a nonlinear
program. To our best knowledge, rare studies transformed the optimal control problem into a linear
program.
C. Contribution of the Paper

To address the research gaps, this paper designs an intersection modeling approach that
combines the tile-based and conflict point-based approaches. Realizing the advantages that the tile-based
approach reflects the vehicle's shape through tiles, and the conflict point-based approach separates
conflicting vehicles by considering only a single point instead of multiple tiles, we expand conflict points
on an infinitely granular grid to a conflict region, resulting in the conflict region-based approach. As a
result, we consider vehicle length and width but do not need to add a large number of tiles to the model's
constraints.
Then, we propose a two-stage method to optimize timing schedules and trajectories for CAVs at
a signal-free intersection with the conflict region-based approach. The first stage, constructed as a MILP,
optimizes vehicle schedules. The second stage is linear programming aimed to reduce fuel consumption.
Moreover, the model is oriented to free-turning lanes based on the concept that CAVs can reach their
destinations without any on-road lane changes to achieve high traffic efficiency and safety [22].
In addition, we add a heuristic rule to the origin MILP to cut off some branches of the search tree
to improve the solution efficiency. Although it transforms into a suboptimal model, it can balance a trade-
off between optimality and computation time, especially for large-scale problems.
Furthermore, a rolling optimization strategy is adopted, and the cycle length is variable, which
suits different traffic demands. Numerical simulations verify the validity of the proposed method, and
the sensitivity analysis is discussed.
The contributions of this study are as follows. (1) A conflict region-based approach that combines
the advantages of the tile-based and the conflict point-based approaches is designed to model the
intersection. (2) A two-stage linear method is proposed based on the conflict region-based approach, and
a rolling optimization strategy with variable cycle length is designed. (3) The solving of the timing
schedule optimization is accelerated by adding a heuristic rule to reduce constraints, and a trade-off
between optimality and computation time is balanced. (4) The proposed two-stage method is compared
4
with both the FCFS-based method and an optimization-based method based on the genetic algorithm. In
addition, the discussions of the optimization-based method under asymmetric traffic demands are
displayed, and the pros and cons of the MILP under different degrees of asymmetry are concluded.
D. Organization of the Paper

The remainder of this paper is structured as follows. Section II describes the designed
autonomous intersection system. Section III designs an intersection modeling approach. Section IV
proposes a two-stage optimization method. Section V conducts numerical experiments. Section VI
concludes this paper and suggests future work.
II. DESIGN OF AN AUTONOMOUS INTERSECTION
A. Intersection Layout
Fig. 1 illustrates a four-legged autonomous intersection with a manager in the center and two
lanes on each side, where the lane is indicated by ℎ ∈ 𝐻 = {1,2, … ,8}. The area where vehicle paths may
interact is defined as the intersection area, which can be seen as a square with a side length 𝐷. The width
of each lane is 𝑤. The turning path is simplified to a quarter arc, whose center is the corner of the
intersection area. The entry lane comprises two segments: an optimization area (indicated in orange) and
a control area (marked in blue) with a length 𝐿𝑂 and 𝐿𝐶 , respectively. The vehicles communicate in these
two areas and are optimized and controlled in turn. Roadside facilities are equipped at the entry of the
optimization area, aiming to detect traffic demand and guide approaching vehicles.
Fig. 1. Illustration of an autonomous intersection.
For modeling and analysis purposes, the following assumptions needed to be made.
● The IM can obtain all vehicles' real-time information through V2I devices, and all CAVs can also
share real-time status information through V2V devices.
5
● This study does not consider communication delays [10], [40].
● The vehicles entering and exiting the intersection from the same entry and exit lanes share the
same turning path [21], [22].
● To accurately predict the time when vehicles enter the control area, it is assumed that vehicles
move at a uniform speed in the optimization area [41].
● All vehicles travel at the same speed in the intersection area to guarantee stability [35], [37].
B. The Process of Optimizing and Controlling the Vehicles

In the proposed intersection system, three subjects, namely Intersection Manager, Vehicle, and
Roadside Facility are cooperating with each other, as shown in Fig. 2. The vehicles follow the guidance of
roadside facilities before entering the optimization area, and start communicating with the IM when
entering the optimization area and finish the whole process after leaving the intersection. The overall
process is designed as follows.
Step 1: The vehicles enter the intersection system: The roadside facilities demonstrate the suggested
speed for vehicles to enter the optimization area. The suggested speed is calculated by the IM according
to the traffic state in the intersection system, which is not discussed in this paper. A proper slowdown
needs to be executed before entering the optimization area at oversaturated intersections. If the suggested
speed is 0, vehicles stop outside the optimization area and wait for an entry; otherwise, they enter the
optimization area at the suggested speed.
Step 2: The IM collects vehicle status information: The vehicles communicate with the IM in the
optimization area, and the IM collects their status information within a certain period, including the
initial position, speed, turning, and vehicle attributes. Vehicle attributes mainly include the maximum
acceleration, deceleration, and vehicle length. Based on the initial position and speed, the IM predicts the
time of vehicles entering the control area.
Step 3: The IM assigns the arrival time for each vehicle: Based on the status information and entry
time, the IM optimizes the schedule and calculates an arrival time (the time to leave the control area and
enter the intersection) for each vehicle in the optimization area.
Step 4: The IM design a speed profile for each vehicle: After determining the arrival time, the IM assigns
each vehicle a driving strategy based on the entering and leaving time of the control area.
Step 5: The vehicles execute the control strategy: After accepting the driving strategy, the vehicles
enter the control area and execute the strategy, aiming to reach the intersection boundary with the arrival
time through acceleration and deceleration processes. The vehicles then follow the fixed path in the
intersection and leave it from the target exit.
The framework of the two-stage method and the rolling optimization are introduced later in
Section IV.
6
Fig. 2. Cooperation of the intersection manager, vehicles, and roadside facilities.
III. DESIGN OF THE INTERSECTION MODELING APPROACH
A. Comparison of the Tile-Based and Conflict Point-Based Approaches

There are two main categories of intersection modeling approaches. One is the tile-based
approach, and the other is the conflict point-based approach, as shown in Fig. 3.
7
Fig. 3. Different intersection modeling approaches: (a) The tile-based approach; (b) The conflict point-
based approach.
1) The tile-based approach: As shown in Fig. 3 (a), the tile-based approach divides the intersection
into n × n tiles, where n is defined as the granularity. Modeling based on this approach constrains each
tile to ensure that no tile is occupied by more than one vehicle simultaneously to avoid crashes. With this
feature, the shape of the vehicle is fully considered regardless of its movement. However, the degree of
utilization of the intersection space is affected by the granularity, for example, the yellow tiles in Fig. 3
(a) indicate the saved unoccupied tiles when the granularity is marked from 4 to 8. Although increasing
the granularity improves intersection utilization, it leads to computational challenges.
2) The conflict point-based approach: As shown in Fig. 3 (b), the conflict point-based approach fixes
the turning movement path and defines the intersection of conflicting paths as conflict points.
Separations are also required in this approach but aim at each conflict point. In other words, this
approach constrains the time interval between vehicles passing the conflict point, which is more
amenable to mathematical programming formulations. However, it is hard to consider the vehicle's
geometry when setting the time interval because this approach reduces the path to a line while the vehicle
is two-dimensional.
B. Combination of the Tile-Based and Conflict Point-Based Approaches: The Conflict Region-Based Approach
Starting from two basic points: (1) the vehicle geometry, including length and width, is
considered; and (2) the approach is independent of granularity and easy to model and solve, this paper
combines the tile-based and conflict point-based approaches to reflect the advantages of both.
Suppose we construct the turning movement at an infinitely granular intersection space, as
shown in Fig. 4. It can be seen as extending the lane into the intersection as a path, which gives the path
width and turns it into a channel. In this way, it retains the characteristics of the tile-based approach that
can reflect the vehicle's geometric shape. On this basis, following the definition in the conflict point-based
approach, we define the overlap of conflicting channels as the conflict region. As shown in Fig. 5, the
shape of the conflict region depends on conflict angles. For safety, only one vehicle can exist in the conflict
region at any time, which is the same as the tile-based approach. However, in the tile-based approach,
8
each tile in the conflict region needs a constraint. Inspired by the conflict point-based approach, only the
key points entering and leaving the conflict region need to be considered by treating the conflict region
as a point. Therefore, we calculate when vehicles arrive and leave the conflict region to separate the
conflict in time and space.
Fig. 4. Gradually increase the granularity.
Fig. 5. Illustration of the conflict region: (a) a cross conflict and the entering key point; (b) a cross
conflict and the leaving key points; (c) a diverge conflict; (d) a merge conflict.
We define the crossover points of the channel's centerline and the boundaries of the conflict
region as key points, and use 𝑘𝑝𝐼 and 𝑘𝑝𝑂 to indicate the entering key point and leaving key point,
respectively. The arc length on the channel's centerline represents the distance between the intersection's
9
entrance and the key points. Here, we use 𝑢𝑖,𝑗 and 𝑢𝑖,𝑗 to denote the distance along the channel of vehicle
𝑖, from the entering point to the conflict region with vehicle 𝑗. Specifically,
𝑘𝑝𝐼
𝑢𝑖,𝑗 = 𝑑𝑖𝑠(𝑝𝑖𝑒𝑛𝑡𝑒𝑟 , 𝑝𝑖 ), (1)
𝑘𝑝𝑂
𝑢𝑖,𝑗 = 𝑑𝑖𝑠(𝑝𝑖𝑒𝑛𝑡𝑒𝑟 , 𝑝𝑖 ), (2)
where 𝑑𝑖𝑠(𝐴, 𝐵) is the distance between 𝐴 and 𝐵 on the channel. 𝑝𝑖∗ is the coordinates of vehicle 𝑖. The
different superscript ∗ denotes different positions, where 𝑒𝑛𝑡𝑒𝑟 means the entrance of the intersection
area, and 𝑘𝑝𝐼 and 𝑘𝑝𝑂 are the key points on the vehicle's channel.
Note that we are not the only ones to assign widths to paths, but we clarify the differences and
connections between this approach and the other two widely used approaches. He et al. [22] ignore the
diverging conflict region and assume identical crossing and merging conflict regions. However, the
diverging conflict region is as important as the merging conflict region (as shown in Fig. 5 (c) and (d)),
and the area of conflict regions varies a lot, especially for two straight-ahead conflicts and two left-turn
conflicts from the outer lane. In the model in [23], two vehicles involved in a diverging or merging conflict
region stick to a car-following rule:
𝑥𝑗𝑒 (𝑡) − 𝑥𝑖𝑒 (𝑡) ≥ 𝑙𝑗 + 𝑣𝑖 (𝑡)𝛥𝑡, (3)
where 𝑥𝑖𝑒 (𝑡) , 𝑣𝑖 (𝑡) are the distance between vehicle 𝑖 ’s current position from the entrance of the
intersection area along its channel and vehicle 𝑖’s current speed at time 𝑡, respectively. 𝑙𝑗 is the length of
vehicle 𝑗, and 𝛥𝑡 is the minimum time headway between adjacent vehicles. However, Eq. (3) is not
accurate when two vehicles are not engaged in one lane. As shown in Fig. 6, although Eq. (3) is satisfied
as 𝑥𝑗𝑒 (𝑡) − (𝑙𝑗 + 𝑣𝑖 (𝑡)𝛥𝑡) > 𝑥𝑖𝑒 (𝑡), vehicle 𝑖 invades the safety gap between two vehicles, which may lead
to potential collisions. In contrast to [22], [23], although similar in presentation, the nature of our
modeling requires us to consider different angles of conflict, including diverging and merging, as well
as not allowing vehicles to remain following in the conflict region. Therefore, the proposed approach,
named as the conflict region-based approach, can more accurately describe reality.
10
Fig. 6. The invaded area in a diverging conflict region.
IV. TWO-STAGE OPTIMIZATION METHOD

The proposed two-stage method includes a timing schedule optimization as the first stage and
trajectory optimizations as the second stage, as shown in Fig. 7. The timing schedule optimization is
modeled as a MILP to minimize traffic delay. The output is the time for each vehicle to arrive at the
intersection. Based on this result, trajectory optimizations are modeled as linear programs to reduce fuel
consumption. The output of the second stage is the instantaneous acceleration, which gives the speed
profile used to guide the vehicles. Note that the trajectory optimization is constructed separately for each
lane (the total lane number is |𝐻|), and the models are run in parallel.
Fig. 7. Framework of the two-stage method.
The parameters and decision variables used in this paper are summarized in TABLE I. Note that
some parameters in the first-stage model also show up in the second-stage model.
TABLE I
LIST OF NOTATIONS
The intersection layout
Parameters Descriptions
𝐷 The width of the intersection
𝑤 The width of each lane
The first-stage model
11
𝐼 The set of vehicles in an optimization
𝑡𝑖− The time when vehicle 𝑖, 𝑖 ∈ 𝐼 enters the control area
𝑡𝑖∗ The ideal travel time of vehicle 𝑖, 𝑖 ∈ 𝐼
𝛥𝑡 The minimum time headway
𝑣𝑖− The speed when vehicle 𝑖, 𝑖 ∈ 𝐼 enters the control area
max
𝑣road The speed limit of the entry lane
𝑣int The travel speed within the intersection area
𝑎max The max acceleration of vehicles
𝑏max The max deceleration of vehicles
𝐿𝐶 The length of the control area
𝑙𝑖 The length of vehicle 𝑖, 𝑖 ∈ 𝐼
𝑢𝑖,𝑗 The distance along the channel of vehicle 𝑖, 𝑖 ∈ 𝐼, from the entering point to the leaving
key point with vehicle 𝑗, 𝑗 ∈ 𝐼
𝑢𝑖,𝑗 The distance along the channel of vehicle 𝑖, 𝑖 ∈ 𝐼, from the entering point to the entering
key point with vehicle 𝑗, 𝑗 ∈ 𝐼
𝜃𝑖 The lane vehicle 𝑖, 𝑖 ∈ 𝐼 comes from
𝜎𝑖 The turning of vehicle 𝑖, 𝑖 ∈ 𝐼
𝑀 A large positive number
𝜛𝑗 A threshold of the time difference between vehicle 𝑖, 𝑖 ∈ 𝐼 and vehicle 𝑗 ∈ [1, 𝑖) entering
the control area
Decision variables Descriptions
+
𝑡𝑖 The time when vehicle 𝑖, 𝑖 ∈ 𝐼 leaves the control area (the time arrives at the intersection)
𝜆𝑖𝑗 A binary variable, 𝜆𝑖𝑗 = 1 when vehicle 𝑗, 𝑗 ∈ [1, 𝑖) leaves the conflict region before
vehicle 𝑖, 𝑖 ∈ 𝐼 enters
The second-stage model
𝐻 The set of entry lanes
𝐼ℎ The set of vehicles on lane ℎ ∈ 𝐻 in an optimization
𝑘 The index of a time step
𝛿 A sufficiently small number
𝑧𝑖 [𝑘] The auxiliary variable for linearization
𝑣𝑖 [𝑘] The speed of vehicle 𝑖, 𝑖 ∈ 𝐼 at time step 𝑘
𝑥𝑖 [𝑘] The position of vehicle 𝑖, 𝑖 ∈ 𝐼 at time step 𝑘
Decision variables Descriptions
𝑎𝑖 [𝑘] The acceleration (or deceleration) of vehicle 𝑖, 𝑖 ∈ 𝐼 at time step 𝑘
The rolling optimization strategy
𝑇 The total optimization time
𝛥𝑇 The cycle length
𝑄 The total traffic demand of the intersection
A. Timing Schedule Optimization

Timing schedule optimization separates vehicle conflicts within an intersection and makes
vehicles pass in alternating patterns. Before building the model, we first construct a virtual platoon to
facilitate indexing. We combine and project vehicles from different lanes on a virtual lane in the order of
12
the time when they are expected to enter the control area to form a virtual platoon [42], as shown in Fig.
8.
The vehicles are numbered with 𝑖, 𝑖 ∈ 𝐼 = {1,2, ⋯ , |𝐼|}, where 𝐼 denotes the set of vehicles in an
optimization. The purpose of constructing the indexing sequence is to simplify modelings. We need to
determine the constraints between any two vehicles based on their relationship to avoid crashes. Based
on the virtual platoon, the constraints between vehicle 𝑖 and 𝑗 are constructed in the manner of 𝑖 ∈
[1, 𝐼], 𝑗 ∈ [1, 𝑖), ensuring no missed constraints.
Fig. 8. Illustration of the virtual platoon.
The following content proposes a MILP model to optimize the timing schedule of vehicles passing
through the signal-free intersection.
1) Objective function: The model aims to minimize vehicle delays, which is the difference between
the actual and ideal travel times.
𝐼
min 𝑍 = min ∑(𝑡𝑖+ − 𝑡𝑖− − 𝑡𝑖∗ ), (4)

𝑖=1
where 𝑡𝑖− and 𝑡𝑖+ are the time when vehicle 𝑖 enters and leaves the control area, respectively. Thus, 𝑡𝑖+ −
𝑡𝑖− is the actual travel time of a vehicle passing through the control area. 𝑡𝑖∗ is the ideal travel time, which
is the minimum travel time for a vehicle to pass through the control area without interference.
13
2) Minimum travel time constraint: To minimize vehicle delay, the vehicle needs to pass through
the control area as quickly as possible. The minimum travel time passing through the control area
depends on the speed, acceleration, deceleration, and area length. When the entering and leaving speeds
are restricted, the maximum travel distance 𝑥max can be divided into two situations, as shown in Fig. 9.
(1) The vehicle accelerates to a speed lower than the speed limit of the control area and then decelerates;
(2) the vehicle speeds up to the speed limit and moves at a constant speed for a while, before decelerating
to the leaving speed. To ensure the vehicle can pass through the control area, the maximum travel
distance has to be larger than the length of the control area, i.e., 𝑥max ≥ 𝐿𝐶 .
In the first situation,
max 2 max 2
(𝑣road ) − (𝑣𝑖− )2 (𝑣int )2 − (𝑣road )
+ ≥ 𝐿𝐶 , (5)
2𝑎max 2𝑏max
max
where 𝑣road is the speed limit of the entry lane, 𝑣int is the travel speed within the intersection area. 𝑎max
and 𝑏max are the max acceleration and deceleration, respectively. 𝑣𝑖− is the initial speed entering the
max max
control area, which is between a lower bound 0 and an upper bound 𝑣road (i.e., 0 ≤ 𝑣𝑖− ≤ 𝑣road ).
As shown in Fig. 9 (a), the area surrounded by the red dashed line and the axes is larger than 𝐿𝐶 .
In this situation,
(𝑣𝑖max )2 − (𝑣𝑖− )2 (𝑣int )2 − (𝑣𝑖′ )2

𝑥ma𝑥 = + , (6)
2𝑎max 2𝑏max
𝑥max ≥ 𝐿𝐶 , (7)
𝑣𝑖′ − 𝑣𝑖− 𝑣int − 𝑣𝑖′

𝑡𝑖′ = + , (8)
𝑎max 𝑏max
𝑡𝑖+ = 𝑡𝑖− + 𝑡𝑖′ . (9)
Fig. 9. Illustration of travel distance.
Joint Eqs. (6)-(9), we have
14
2
𝑏 (𝑣 − ) − 𝑎max (𝑣int )2 + 2𝐿𝐶 𝑏max 𝑎max
𝑎max 𝑣int − 𝑏max 𝑣𝑖− + (𝑏max − 𝑎max )√ max 𝑖
𝑏max − 𝑎max
𝑡𝑖+ ≥ 𝑡𝑖− + ,
𝑎max 𝑏max (10)
max 2 max 2
∀ + ≥ 𝐿𝐶 .
2𝑎max 2𝑏max
In the second situation,

max 2 max 2
+ < 𝐿𝐶 . (11)
2𝑎max 2𝑏max
As shown in Fig. 9 (b), the area surrounded by the red dashed line and the axes is smaller than
𝐿𝐶 . In this situation,
max max
max ′
(𝑣road − 𝑣𝑖− )2 (𝑣road − 𝑣int )2
𝑥max = 𝑣road 𝑡𝑖 − + , (12)
2𝑎max 2𝑏max
𝑥max ≥ 𝐿𝐶 , (13)
𝑡𝑖+ = 𝑡𝑖− + 𝑡𝑖′ . (14)
Then, joint Eqs. (12)-(14), we have

max max
𝐿𝐶 (𝑣road − 𝑣𝑖− )2 (𝑣road − 𝑣int )2
𝑡𝑖+ ≥ 𝑡𝑖− + max + max − max ,
𝑣road 2𝑎max 𝑣road 2𝑏max 𝑣road
(15)
max 2 max 2
∀ + < 𝐿𝐶 .
2𝑎max 2𝑏max
3) Arrival time constraint: When the vehicles come from the same lane, the headway should be
more significant than the minimum safe time interval 𝛥𝑡 to avoid rear-end collisions,
𝑙𝑗
𝑡𝑖+ − 𝑡𝑗+ ≥ + 𝛥𝑡. (16)
𝑣int
In Eq. (16), 𝑖 and 𝑗 are same-lane vehicles, and vehicle 𝑗 is the preceding one. 𝑙𝑗 represents the
length of vehicle 𝑗.
Moreover, when the same lane vehicles have different turnings, their channels will create a
conflict region. Since the conflict region can only be occupied by one vehicle simultaneously (and there
will be no overtaking on the same lane), the preceding vehicle has to leave the conflict region before the
rear vehicle enters.
15
𝑢𝑗,𝑖 − 𝑢𝑖,𝑗 + 𝑙𝑗
𝑡𝑖+ − 𝑡𝑗+ ≥ ( ). (17)
𝑣int
𝑘𝑝𝐼 𝑘𝑝𝐼
In Eq. (17), 𝑢𝑖,𝑗 = 𝑑𝑖𝑠(𝑝𝑖𝑒𝑛𝑡𝑒𝑟 , 𝑝𝑖 ), where 𝑝𝑖𝑒𝑛𝑡𝑒𝑟 = 𝑝𝑖 when vehicle 𝑖 and 𝑗 come from the same
lane. Thus, 𝑢𝑖,𝑗 = 0, and this equation can be simplified as
𝑢𝑗,𝑖 + 𝑙𝑗
𝑡𝑖+ − 𝑡𝑗+ ≥ . (18)
𝑣int
When the vehicles come from different lanes, their channels may overlap, creating a conflict
region. As mentioned before, the vehicles with the same conflict region must be separated in time. In
order to ensure safety, when the rear vehicle reaches the conflict region, the preceding vehicle must have
left. Here, we introduce a 0-1 variable 𝜆𝑖𝑗 and a large positive number 𝑀 to help construct the sequential
problem.
𝑢𝑗,𝑖 − 𝑢𝑖,𝑗 − 𝑙𝑖
𝑡𝑖+ − 𝑡𝑗+ ≤ ( ) (1 − 𝜆𝑖𝑗 ) + 𝜆𝑖𝑗 𝑀, (19)
𝑣int
𝑢𝑗,𝑖 − 𝑢𝑖,𝑗 + 𝑙𝑗
𝑡𝑖+ − 𝑡𝑗+ ≥ ( ) 𝜆𝑖𝑗 + (𝜆𝑖𝑗 − 1)𝑀. (20)
𝑣int
In Eqs. (19) and (20), 𝜆𝑖𝑗 = 1 indicates that vehicle 𝑗 enters and leaves the conflict region first;
otherwise, vehicle 𝑖 enters and leaves the conflict region first. In particular, for vehicles coming from
different lanes but not constituting a conflict region, it is defined that 𝑢𝑗,𝑖 − 𝑢𝑖,𝑗 = 𝑀, 𝑢𝑗,𝑖 − 𝑢𝑖,𝑗 = −𝑀.
4) The timing schedule optimization model: We propose two MILP models with different constraints.
Each model has minimum travel time and arrival time constraints but with different application ranges.
We define 𝜃𝑖 as the lane vehicle 𝑖 comes from, and 𝜎𝑖 as its turning. The two models are expressed as
follows.
a) M1: The original optimization model.
M1
Objective function:
Eq. (4),
Subject to:
Minimum travel time:
Eqs. (10) and (15), ∀ 𝑖 ∈ [1, 𝐼],
Car-following:
Eq. (16), ∀ 𝑖 ∈ [1, 𝐼], 𝑗 = 𝑗 ′ , 𝜃𝑖 = 𝜃𝑗 ′ ,
Diverge collision avoidance:
Eq. (18), ∀ 𝑖 ∈ [1, 𝐼], 𝑗 = 𝑗 ′ , 𝜃𝑖 = 𝜃𝑗 ′ , 𝜎𝑖 ≠ 𝜎𝑗 ′ ,
Cross and merge collision avoidance:
Eqs. (19) and (20) ∀ 𝑖 ∈ [1, 𝐼], 𝑗 ∈ [1, 𝑖), 𝜃𝑖 ≠ 𝜃𝑗 .
16
In M1, Eqs. (10) and (15) are the minimum travel time constraints, that are applied to all vehicles.
The arrival time constraints for same-lane vehicles are only applied to adjacent vehicles. Thus, we use 𝑗 ′
to represent the preceding vehicle of vehicle 𝑖 on the real lane, i.e.,
∀ 𝑗 ′ ∈ [1, 𝑖), 𝜃𝑗′ = 𝜃𝑖 , ∄ 𝑗 ′′ ∈ (𝑗 ′ , 𝑖), 𝑠. 𝑡. 𝜃𝑗′′ = 𝜃𝑖 . (21)
If the adjacent vehicles have the same turning, Eq. (16) is applied. However, if they have different
turnings, both Eq. (16) and Eq. (18) are applied. For vehicles coming from different lanes, Eqs. (19) and
(20) are applied for each vehicle pair.
b) M2: The sub-optimal model with fewer binaries.
M2
Objective function:
Eq. (4),
Subject to:
Minimum travel time:
Eqs. (10) and (15), ∀ 𝑖 ∈ [1, 𝐼],
Car-following:
Eq. (16), ∀ 𝑖 ∈ [1, 𝐼], 𝑗 = 𝑗 ′ , 𝜃𝑖 = 𝜃𝑗 ′ ,
Diverge collision avoidance:
Eq. (18), ∀ 𝑖 ∈ [1, 𝐼], 𝑗 = 𝑗 ′ , 𝜃𝑖 = 𝜃𝑗 ′ , 𝜎𝑖 ≠ 𝜎𝑗 ′ ,
Cross and merge collision avoidance:
Eqs. (19) and (20) ∀ 𝑖 ∈ [1, 𝐼], 𝑗 ∈ [1, 𝑖), 𝜃𝑖 ≠ 𝜃𝑗 , 𝑡𝑖− − 𝑡𝑗− ≤ 𝜛𝑗
Eq. (17) ∀ 𝑖 ∈ [1, 𝐼], 𝑗 ∈ [1, 𝑖), 𝜃𝑖 ≠ 𝜃𝑗 , 𝑡𝑖− − 𝑡𝑗− > 𝜛𝑗 .
M2 modifies the application range of Eqs. (19) and (20) based on M1. In M2, the sequential
problem between vehicles from different lanes is only for the vehicle pair whose time difference of
entering the control area is less than 𝜛𝑗 . In other words, when the time difference between two vehicles
entering the control area is greater than 𝜛𝑗 , the vehicle that enters first passes the conflict region first by
default. This heuristic rule reduces many constraints that contain binaries and is expected to improve the
solving speed to a large extent. However, the solution is often not optimal due to the reduced form of
vehicle order combinations.
B. Trajectory Optimization
The goal of trajectory optimization is to reduce fuel consumption, which can be calculated by Eq.
(22) [43].
𝛽2 𝑚𝑎(𝑡)2 𝑣(𝑡)
𝛼 + 𝛽1 𝑅𝑇 𝑣(𝑡) + [ ] , 𝑅𝑇 > 0
𝑓(𝑣, 𝑎; 𝑡) = { 1000 𝑎(𝑡)>0
, (22)
𝛼 , 𝑅𝑇 ≤ 0
where 𝑅𝑇 is the total tractive force,
17
𝑚𝑎(𝑡)
𝑅𝑇 = 𝑑1 + 𝑑2 𝑣(𝑡) + 𝑑3 𝑣(𝑡)2 + + 9.81 × 10−5 𝑀𝐺. (23)
1000
According to [44], the values of the relevant parameters in Eqs. (22) and (23) are 𝛼 = 0.666mL/s,
𝛽1 = 0.0717 , 𝛽2 = 0.0344mL/(kJ ⋅ m/s 2 ) , 𝑑1 = 0.269kN , 𝑑2 = 0.0171kN/(m/s) , 𝑑3 = 0.000672kN/(m/
s)2 , 𝑚 = 1600kg, and 𝐺 = 0.
1) Objective function: Since Eq. (22) is highly nonlinear, the absolute value of vehicle acceleration
(or deceleration) is used as a simplified function [4], which is a common practice in optimal control theory
[45]. In addition, to avoid the formation of an optimal control problem, which is hard to solve, we assume
a sufficiently small number named 𝛿, and let 𝑡𝑖− ≤ 𝑘𝛿 ≤ 𝑡𝑖+ , where 𝑘𝑖− 𝛿 = 𝑡𝑖− and 𝑘𝑖+ 𝛿 = 𝑡𝑖+ . The objective
function can be written as:
𝐼ℎ 𝑘𝑖+
min 𝐽 = ∑ ∑ |𝑎𝑖 [𝑘]|𝛿 . (24)

𝑖=1 𝑘=𝑘𝑖−
where 𝑎𝑖 [𝑘] is the acceleration (or deceleration) of vehicle 𝑖 at time step 𝑘. 𝐼ℎ denotes the total number of
|𝐻|
vehicles on lane ℎ, ℎ ∈ 𝐻 in an optimization, ∑ℎ=1 𝐼ℎ = 𝐼.
The absolute form in Eq. (24) can be linearized as Eqs. (25)-(27) according to [46].
𝐼ℎ 𝑘𝑖+
min 𝐽 = ∑ ∑ 𝑧𝑖 [𝑘]𝛿 , (25)

𝑖=1 𝑘=𝑘𝑖−
𝑧𝑖 [𝑘] ≥ 𝑎𝑖 [𝑘], (26)
𝑧𝑖 [𝑘] ≥ −𝑎𝑖 [𝑘]. (27)
2) Kinematic constraint: The vehicle needs to satisfy the following dynamic equations in driving.
1
𝑥𝑖 [𝑘] = 𝑥𝑖 [𝑘 − 1] + 𝑣𝑖 [𝑘 − 1]𝛿 + 𝑎𝑖 [𝑘 − 1]𝛿 2 ,
2 (28)
∀𝑘 ∈ [𝑘𝑖− + 1, 𝑘𝑖+ ],
𝑣𝑖 [𝑘] = 𝑣𝑖 [𝑘 − 1] + 𝑎𝑖 [𝑘 − 1]𝛿, ∀𝑘 ∈ [𝑘𝑖− + 1, 𝑘𝑖+ ], (29)
where 𝑥𝑖 [𝑘] and 𝑣𝑖 [𝑘] are the position and speed of vehicle 𝑖 at time step 𝑘, respectively.
Meanwhile, the speed and acceleration (deceleration) of the vehicles have a certain range in the
process of driving, which is formulated as follows:
max
0 ≤ 𝑣𝑖 [𝑘] ≤ 𝑣road , ∀𝑘 ∈ [𝑘𝑖− , 𝑘𝑖+ ], (30)
18
𝑏max ≤ 𝑎𝑖 [𝑘] ≤ 𝑎max , ∀𝑘 ∈ [𝑘𝑖− , 𝑘𝑖+ ]. (31)
3) State constraint: The vehicle's initial and termination state must satisfy the following constraints:
𝑥𝑖 [𝑘𝑖− ] = 0, 𝑣𝑖 [𝑘𝑖− ] = 𝑣𝑖− , 𝑎𝑖 [𝑘𝑖− ] = 0, (32)
𝑥𝑖 [𝑘𝑖+ ] = 𝐿𝐶 , 𝑣𝑖 [𝑘𝑖+ ] = 𝑣int , 𝑎𝑖 [𝑘𝑖+ ] = 0. (33)
Eq. (32) indicates that vehicle 𝑖 enter the control area at speed 𝑣𝑖− . Eq. (33) indicates that vehicle 𝑖
enter the intersection area at speed 𝑣int and maintains a constant speed.
4) Safety spacing constraint: To ensure the safety of adjacent vehicles in the same lane, there must
be a minimum safety spacing requirement.
𝑥𝑖−1 [𝑘] − 𝑥𝑖 [𝑘] ≥ 𝑙𝑖−1 + 𝑣𝑖 [𝑘]𝛥𝑡,

(34)
∀𝑘 ∈ [𝑘𝑖− , 𝑘𝑖−1
+ ],
𝑖 ∈ [2, 𝐼ℎ ],
where 𝑥𝑖−1 [𝑘] is the position of the preceding vehicle to vehicle 𝑖 at time step 𝑘. It is better to use 𝑣𝑖 [𝑘]𝛥𝑡
as the spacing than a fixed distance because it varies with speed.
5) The trajectory optimization model: It is formed as a linear program as below. Eq. (25) is the
objective function after linearization. Eq. (26) and (27) are auxiliary constraints helping to linearize the
objective. Eqs. (28)-(31) update a vehicle’s position and speed. Eqs. (32) and (33) hold the initial and
terminal states of a vehicle. Eq. (34) makes sure no rear-end collision occurs.
M3
Objective function:
Eq. (25),
Subject to:
Auxiliary constraints:
Eqs. (26) and (27).
Kinematic conditions:
Eqs. (28)-(31).
State conditions:
Eqs. (32) and (33).
Car-following:
Eq. (34).
C. Rolling Optimization Strategy

The computation time of the optimization model increases exponentially with the number of
vehicles [36], and it is challenging to optimize even one hour's worth of vehicles at a time. Therefore,
rolling optimization is adopted to split a large-period problem into multiple consecutive short-period
problems [47]. The vehicles are optimized with a specific time interval 𝛥𝑇, as shown in Fig. 10. The black,
red and green vehicles denote the vehicles which are not optimized, are being optimized, and have been
19
optimized, respectively. At time 𝑡0 , there are four vehicles in the optimization area being optimized.
After a few seconds, they finished optimization and kept moving (assuming that the vehicles move at a
constant speed). The rear vehicles enter the area successively but are not being optimized before 𝑡0 + 𝛥𝑇.
At time 𝑡0 + 𝛥𝑇, the IM optimizes the vehicles in the optimization area again.
The cycle length 𝛥𝑇 is changed according to the traffic demand because the number of vehicles
that can be optimized simultaneously has an upper bound; otherwise, it will make the model difficult to
solve. Combining the conclusions of [19] that a MILP can be solved in real-time for up to 30 vehicles and
our experiments, the optimized cycle length is set to Eq. (35).
5𝑠 𝑄 > 7200
𝛥𝑇 = { , (35)
10𝑠 𝑄 ≤ 7200
where 𝑄 is the total traffic demand of the intersection. That is to say, when the traffic demand is over
7200 veh/h, the cycle length needs to be shortened to improve the efficiency of the model solution. Note
that the traffic demand of each entry lane can be collected by roadside facilities and transmitted to the
IM.
To ensure that no vehicles are missed between optimizations, the length of the optimization area
has the following requirements.
road (max{𝛥𝑇} 𝑂𝑃 ),
𝐿𝑂 = 𝑣max ⋅ + 𝑡max (36)
OP
where 𝐿𝑂 is the length of the optimization area, and 𝑡max is the maximum tolerable time for the IM to
optimize the vehicles in a cycle. Eq. (36) ensures that the vehicle will not cross the optimization area
within a cycle. Therefore, every vehicle will be optimized at some time. In this study, we ignored the
𝑂𝑃
optimization duration, so 𝑡max = 0.
Note that adjacent optimizations are not independent. In the timing schedule optimization, the
current vehicles need to be constrained with the optimized vehicles that have not left the intersection. In
the trajectory optimization, the last vehicle in the previous optimization and the current first vehicle
should satisfy the safety spacing constraint. Therefore, constraints (16), (18)-(20), and constraint (34) are
adopted between adjacent optimizations. In particular, since the re-optimization of the remaining
vehicles will reduce the efficiency of the model, only the vehicles newly entering the optimization area
in the current cycle are optimized.
20
Fig. 10. Illustration of rolling optimization strategy.
To sum up, suppose the total optimization time is 𝑇, and the rolling optimization strategy is as
follows, as shown in Fig. 2.
Step 1. Initializing parameter 𝑡 = 0;
Step 2. Determine the cycle length 𝛥𝑇 according to traffic demand and form the current optimization
time interval [𝑡, 𝑡 + 𝛥𝑇];
Step 3. Predict the time of arrival at the control area for the vehicles that are in the optimization area
within the current optimization time interval;
Step 4. Optimize the timing schedule of the vehicles in the optimization area within the current
optimization time interval;
Step 5. Based on the timing schedule results, optimize the trajectory of vehicles on different lanes
independently and parallelly.
Step 6. Let 𝑡 = 𝑡 + 𝛥𝑇. If 𝑡 > 𝑇, the rolling optimization ends; otherwise, go to Step 2.
V. NUMERICAL EXPERIMENTS
A. Simulation Setup
To verify the effectiveness of the proposed method, the four-legged, two-lane autonomous
intersection in Fig. 1 is taken as an example to design simulation experiments. For simplicity, it is
assumed that all CAVs are identical and have the same physical properties and characteristics. The main
simulation parameters are shown in TABLE II. It is assumed that the arrival of vehicles obeys a Poisson
distribution [36]. By utilizing a point-queue model [48], we assume that before vehicles enter the
optimization area, they travel at a free-flow speed [36]. Based on the Dedicated Short-Range
Communication (DSRC) technique [49], [50], each segment of the entry lane does not exceed 200 m. The
minimum time headway between adjacent CAVs is set as 1s [30], [36].
21
TABLE II
SIMULATION PARAMETERS
Parameters Value Unit
𝑇 600 s
𝛥𝑇 Calculated by Eq. (35)
𝐷 18 m
𝜔 3 m
𝐿𝐶 200 m
max
𝑣road 15 m/s
𝑣int 15 m/s
𝛥𝑡 1 s
𝑎max 3 m/s2
𝑏max -6 m/s2
𝑣max 15 m/s
𝑣min 0 m/s
𝑙 4.5 m
Note that the experiment is a numerical simulation. Therefore, pre-warm time is not included in
the duration of 𝑇. Besides, we assume that all CAVs can be controlled in the simulation experiment,
regardless of the vehicle communication failure and other abnormal factors. All simulation experiments
are implemented in Python, and the timing schedule optimization model and the trajectory optimization
model are solved by Gurobi [51]. The computer processor is an Intel(R) Xeon(R) CPU E3-1225 v6,
3.30GHz, with 16.0 GB RAM. To consider the randomness effect, each scenario is simulated ten times
and averaged.
The FCFS-based method [24] and the GA-based method [17] are applied as benchmarks. In the
FCFS-based method, the vehicle arrival time is planned according to the order of entering the control
area, i.e., when 𝑡𝑖+ is determined, subsequent vehicles plan their arrival times based on it, to achieve a
conflict-free crossing. In the GA-based method, the vehicle sequence is determined by the genetic
algorithm, and the arrival time is then calculated according to the sequence. For more details about the
GA-based method, readers can refer to [17]. When testing the effectiveness of the proposed method and
making sensitivity analysis, M1 is used as the timing schedule optimization model for the proposed
method. Note that all three methods (the proposed method and two benchmarks) only differ in the
timing schedule optimization. As for trajectory optimization, all methods use M3 to optimize the
trajectory based on the arrival time results, to minimize fuel consumption.
B. Test the Effectiveness of the Proposed Method

1) Traffic demand cases: Five cases are designed to verify the control performance of the proposed
method under different traffic demands, as shown in TABLE III. In each case, the ratio of left-turn,
straight-ahead, and right-turn vehicles is 3:5:2.
22
TABLE III
DIFFERENT TRAFFIC DEMANDS
Flow per Lane (veh/h)
Case Total Flow (veh/h)
South-North East-West
1 600 600 4800
2 600 900 6000
3 900 900 7200
4 900 1200 8400
5 1200 1200 9600
2) Results and discussion: TABLE IV demonstrates the results of timing schedule optimizations
among the three methods. The delay values in the table are the average delay of one vehicle. The
reduction value (reduction percentage) is the difference between the results of the proposed method and
the benchmark. A positive reduction means the proposed method earns a lower average delay. The
columns named ‘Time’ record the average computation time (CPU time) of each method. Note that the
CPU time is the computation time for solving the entire traffic flow (i.e., in the proposed method and the
GA-based method, it is the sum of solving all cycles, while in the FCFS-based method, it is the sum of
solving all vehicles).
TABLE IV
RESULTS OF TIMING SCHEDULE OPTIMIZATION
The Proposed
The GA-based Method The FCFS-based Method
Method
Cas Reducti Reductio Reducti Reductio
Avg.
e Time Avg. on n Time Avg. on n Tim
Delay
(s) Delay (s) Value Percentag (s) Delay (s) Value Percenta e (s)
(s)
(s) e (s) ge
1 0.27 1.63 0.27 0.01 2.53% 396.25 0.46 0.19 42.22% 1.38
2 0.38 6.10 0.40 0.02 5.45% 517.61 0.76 0.38 50.19% 1.99
3 0.51 29.03 0.56 0.05 9.26% 708.55 1.28 0.77 60.52% 2.86
4 0.87 3.30 0.92 0.05 5.91% 638.01 2.98 2.11 70.83% 3.74
1060.9
5 1.45 11.14 1.59 0.14 8.74% 13.78 12.33 89.48% 5.26
9
It can be seen that higher demand causes more significant delays, and the average vehicle delays,
controlled by the proposed method are always the lowest among the three methods. However, the
improvement of the proposed method relative to the GA-based method is not significant in terms of
delays. When compared to the FCFS-based method, the delay reduction percentage shows an increasing
trend up to 89.48%. We argue that when the traffic demand is low, the continuity between vehicles is
poor, which means the combination of vehicle passing orders is homogeneous. When the traffic demand
23
is high, the change in the sequence of two adjacent vehicles will significantly impact the traffic delay.
Therefore, the methods which optimize the vehicle passing order (both the proposed method and the
GA-based method) show more significant improvement at higher demands.
Using only average delays for comparison is not comprehensive because high delays of
individual vehicles may affect the average. Therefore, as shown in Fig. 11, the delay data are used to
draw a boxplot. It can be seen that the high average delays from the FCFS-based model are not caused
by individuals but rather by the overall characteristics of the platoon. Not only is the maximum delay
higher, but the median delay is also more significant. In addition, the vehicle delays obtained by the
proposed method and the GA-based method are more concentrated, indicating that their performance is
more stable.
Fig. 12, as a supplement to Fig. 11, illustrates the traffic delay of different vehicles. It indicates
that not all vehicles have lower delays under the proposed method, which reveals the nature of the
optimization model: Sacrifice the interests of some vehicles for the efficiency of the whole platoon.
Conversely, the FCFS-based method is greedy. Each vehicle aims to minimize its delay while not
colliding with preceding ones, leading to low delays for some vehicles but a high system delay.
In addition to the delay result, CPU time is also an important metric for evaluating methods.
From TABLE IV, it can be seen that the CPU time of the proposed method is much less than that of the
GA-based method. Although the proposed method makes no significant improvement to the GA-based
method on traffic delay, it outperforms the GA-based method in terms of solving time. Note that some
unnormal values exist in the proposed method and the GA-based method (e.g., the CPU time of the
proposed method under case 3 is 29.03s). This is due to the large CPU time of individual cycles, which
leads to a large average CPU time. As shown in Fig. 13, the black dots show the outliers beyond the caps.
These outliers indicate those cycles that contribute to increasing the average CPU time. After filtering the
outliers (by replacing the outliers beyond the caps as the caps), the average CPU time for solving the
entire traffic flow is shown in TABLE V. The filtered CPU times are more likely to represent the efficiency
of the methods in general scenarios, however, the unfiltered values are comprehensive because they
contain the situations where the method is hard to converge. An interesting phenomenon appears that
the CPU time in TABLE V does not show an increasing trend with the traffic demand. For example, the
CPU time of case 4 is shorter than that of case 3. It is because the cycle length used in cases 1-3 is 10 s and
the cycle length used in cases 4-5 is 5 s based on Eq. (35). Therefore, although traffic demand increases,
decreasing the cycle length reduces the number of decision and binary variables in the MILP (reduces
the combinations of vehicle sequence), which improves the solving efficiency. Consequently, the CPU
time shows an increasing trend with the traffic demand under the same cycle length in both the proposed
method and the GA-based method. As for the FCFS-based method, which does not have the concept of
‘cycles’, its computation time increases approximately linearly with the number of vehicles.
24
Fig. 11. Comparison of vehicle delay: (a) Case 1; (b) Case 2; (c) Case 3; (d) Case 4; (e) Case 5.
Fig. 12. Different single-vehicle delays.
Another interesting phenomenon is that the vehicle delay and CPU time under the proposed
method vary against asymmetric traffic demands. We assume that traffic demands in opposite directions
are the same and fix the total traffic demand at 7200 veh/h. Therefore, when the North-South traffic
demand ranges from 100 veh/h/lane to 1700 veh/h/lane, the East-West traffic demand decreases from
1700 veh/h/lane to 100 veh/h/lane. Fig. 14 (a) shows that severe asymmetric traffic demand yields high
traffic delay but low CPU time under the proposed method (a similar trend appears in the GA-based
method), and the reasons will be discussed later. Fig. 14 (b) testifies that the FCFS-based method
generates enormous delays at asymmetric intersections. Note that this conclusion is not the same as [25].
25
In [25], the lane configuration is asymmetrical; however, we demonstrate that the same results occur
despite symmetric lane configuration.
Fig. 13. Comparison of CPU Time: (a) Case 1; (b) Case 2; (c) Case 3; (d) Case 4; (e) Case 5.
TABLE V
COMPARISON OF COMPUTATION TIMES
The Proposed Method The GA-based Method
Case
Time (Origin) Time (Filtered) Time (Origin) Time (Filtered)
1 1.63 0.62 396.25 378.81
2 6.10 1.93 517.61 487.67
3 29.03 6.26 708.55 659.34
4 3.30 2.48 638.01 621.23
5 11.14 6.67 1060.99 1036.09
Let's discuss the above phenomenon. The area of overlap channels between parallel directions
(e.g., North-South) is larger than that of vertical directions (e.g., North-West) in the conflict region-based
approach. Generally, the delay between parallel directions should be more significant. Further, the
delayed vehicle will affect its following vehicles and may form a traffic wave. Hence, asymmetric traffic
demands lead to significant delays.
Suppose an intersection with only two approaches and only straight-ahead vehicles are
26
considered (the same as [29], [30]). Assume 𝓂 and 𝓃 vehicles are on two approaches, respectively. As a
result, 𝓂 × 𝓃 binaries are needed to decide the passing order according to Eqs (19)(20). We denote 𝑠 ≔
𝓂 × 𝓃 and 𝑞 ≔ 𝓂 + 𝓃. Thus, 𝑠 = −𝓂2 + 𝓂𝑞 and 𝑠̇ = −2𝓂 + 𝑞. Since 𝑞 is fixed (=1800), 𝑠 is a quadratic
function, and its maximum value is obtained at 𝓂 = 0.5𝑞 . Therefore, without loss of generality,
intersections with symmetric traffic demands yield a high CPU time of the proposed method.
Conversely, the more severe asymmetric traffic demand is, the lower the CPU time is.
Fig. 14. Variation of average vehicle delay and total computation time against Nort-South traffic
demand: (a) The proposed method; (b) The FCFS-based method.
TABLE VI
RESULTS OF TRAJECTORY OPTIMIZATION
The Proposed
The GA-based Method The FCFS-based Method
Method
Case
Avg. FC Reduction Reduction Reduction Reduction
Avg. FC (mL) Avg. FC (mL)
(mL) Value (s) Percentage Value (s) Percentage
27
1 19.86 19.88 0.03 0.13% 20.62 0.77 3.72%
2 20.29 20.37 0.08 0.39% 21.75 1.46 6.73%
3 20.78 20.96 0.18 0.86% 23.59 2.81 11.92%
4 22.14 22.33 0.19 0.85% 28.69 6.55 22.83%
5 24.14 24.60 0.46 1.87% 45.40 21.27 46.84%
TABLE VI shows the comparison of fuel consumption. The abbreviation ‘Avg. FC’ means the
average fuel consumption of one vehicle. The CPU times (the average of one cycle) for each case under
three methods are all around 0.01 s. Note that the objective in the trajectory optimization is Eq. (25), but
the fuel consumption is calculated by Eq. (22). In all methods, trajectory optimization is performed,
except that the optimizations are based on different timing schedule results. TABLE VI shows that the
proposed method can reduce vehicle fuel consumption by optimizing the timing schedule. Although the
reduction is not evident at low traffic demand, it increases up to 46.84% as traffic demand grows
compared to the FCFS-based method. Fuel consumption is related to acceleration and deceleration
processes. The more frequently the vehicle accelerates and decelerates, the higher the fuel consumption.
When the traffic demand is low, vehicle delays are minor, and vehicles pass through the control area at
a near free-flow speed, so the fuel consumption does not change significantly under the control of all
methods. When the traffic demand is high, the proposed method significantly reduces the vehicle delay
and shortens the process of vehicle acceleration and deceleration, thus reducing fuel consumption.
Fig. 15 illustrates the trajectory of vehicles on the Eastbound Lane 4 of one of the experiments in
case 5. Traffic waves appear in the FCFS-based method, and vehicles are likely to queue. In addition, the
number of trajectories in the figure shows that fewer vehicles enter the entry lane in the FCFS-based
method due to significant delays from preceding vehicles compared to the other two methods.
(a)
28
(b)
(c)
Fig. 15. Space-time trajectory of one lane: (a) The proposed method; (b) The GA-based method (c) The
FCFS-based method.
If the stop wave spreads to the entire traffic flow, vehicles may slow down to a complete stop,
causing severe traffic jams. As a result, delays for later-entry vehicles can be significant. Therefore, the
trend of single-vehicle delay can determine whether the intersection reaches saturation. In Fig. 16, a bar
chart is used to show the delays of different vehicles, and the black dashed line shows the trend of
average delay per 20 vehicles. Intersections that are unsaturated show a fluctuating trend of vehicle
delays, as shown in Fig. 16 (a). In contrast, saturated intersections have an increasing trend of single-
vehicle delay with vehicle entry sequence due to vehicle queues, as shown in Fig. 16 (b).
In addition to the visual representation of the image, we can use the correlation between vehicle
ID and delay to represent the degree of intersection saturation. The correlation can be calculated as Eq.
(37).
∑(𝑥 − 𝑥̅ )(𝑦 − 𝑦̅)

𝐶𝑜𝑟𝑟𝑒𝑙(𝑋, 𝑌) = , (37)
√∑(𝑥 − 𝑥̅ )2 ∑(𝑦 − 𝑦̅)2
where 𝑥 and 𝑦 are the data of properties 𝑋 and 𝑌. 𝑥̅ and 𝑦̅ are the average data. Since the vehicles at the
rear of the virtual platoon must have entered the control area later, vehicle delays will increase with the
order of vehicles in the virtual platoon at oversaturated intersections. Thus, the correlation between
29
vehicle delay and vehicle ID will be significant.
(a)
(b)
Fig. 16. The delay trend: (a) An unsaturated intersection; (b) An oversaturated intersection.
Fig. 17. Correlation of vehicle numbers and delays under different traffic demands.
Fig. 17 shows the correlation between vehicle ID and delays at different traffic demands. Under
30
the control of the FCFS-based method, the intersection starts to approach saturation when traffic demand
reaches 1100 veh/h/lane. However, the intersections are low saturation until traffic demand reaches 1300
veh/h/lane under the control of the optimization-based method (the proposed method and the GA-based
method). The intersection reaches an oversaturated state earlier under the control of the FCFS-based
method, and the maximum intersection capacity is 300 veh/h/lane less than that of the optimization-
based method. In addition, the intersection switches between the unsaturated and oversaturated states
very abruptly in the FCFS-based method, while it is in the critical saturation state during a period of
demand interval with the optimization-based method. Therefore, the optimization-based method can
increase the capacity of intersections.
C. Sensitivity Analysis
Since the proposed method and the GA-based method are both categorized as optimization-
based methods, and we’ve proved that they both earn quality solutions (in terms of delay and fuel
consumption results), this section only compares the proposed method and the FCFS-based method.
1) Cycle length: The setting of optimization cycle length is a critical issue. The number of
optimizations varies according to cycle length for the same optimization time. Cases 1, 3, and 5 in TABLE
III are examples of the low, medium, and high traffic demands. The cycle length is set to 1-10 s and 1-5 s
for cases 1, 3, and case 5, respectively, both with 1 s as the interval. The rest of the simulation parameters
are shown in TABLE II.
31
Fig. 18. Sensitivity analysis of rolling optimization cycle length: (a) Low traffic demand; (b) Medium
traffic demand; (c) High traffic demand.
Because the timing schedule model in the FCFS-based method is independent of the optimization
cycle, the average delay controlled by the FCFS-based method will not vary with cycle length. In terms
of fuel consumption, the trajectory is optimized according to the timing schedule results in a certain cycle.
Regardless of the cycle division, the intra-cycle constraints and inter-cycle constraints in the trajectory
optimization model are always for two adjacent vehicles. Therefore, for one vehicle, its constraint
relationship is irrelevant to the cycle division. In other words, its optimization trajectory is cycle-
independent unless the established trajectory of the previous optimization cycle impacts the current
vehicle. In the above cases, this effect is minimal, so the manifested result is that the fuel consumption
result is cycle-independent with constant delay. Thus, only the delay and fuel consumption of the
proposed method is demonstrated.
Fig. 18 shows that, under the same traffic demand conditions, the proposed method performs
increasingly well in both delay and fuel consumption as the cycle length increases. For traffic delay, the
main idea of the proposed method is to consider all vehicles as a system and find a system optimal delay
result. A single optimization cycle can be regarded as a subsystem whose solution is locally optimal. The
longer the optimization cycle, the fewer partitioned subsystems and the higher the system integrity. The
32
globally optimal solution is obtained when the length of the optimization cycle is equal to the total
simulation time. Conversely, when the length of the optimization cycle is small enough that only one
vehicle exists in a single cycle, the proposed method is equivalent to the FCFS-based method. Therefore,
a longer optimization cycle can stimulate better performance in the proposed method. However, it is
essential to note that this trend does not vary monotonically and causes fluctuations because the arrival
times that have been optimized are given conditions that affect the vehicles being optimized. This
drawback is similar to that of the FCFS-based method, where individual cycles are also greedy in rolling
optimization. For fuel consumption, the trend is consistent with the trend of delay, which again proves
that the proposed method reduces fuel consumption by optimizing the vehicle timing schedule.
As shown in Fig. 18 (a) and (b), for the low and medium traffic demand, when the cycle length is
10 s, the increasing trend of the delay reduction has been regionally smooth. It means a longer cycle
length only increases the computational consumption and does not improve the method's performance.
Therefore, a 10 s cycle length is appropriate for these two traffic demands. As shown in Fig. 18 (c), for
the high traffic demand, the delay still tends to decrease when the cycle length reaches 5 s. Therefore,
increasing the cycle length can reduce vehicle delays to a greater extent, but the complexity of the model
increases exponentially with the cycle length under high traffic demand, and the decision to increase the
cycle length depends on the computational performance of the IM. Furthermore, another factor that
affects the cycle length is the length of the entry road. According to Eq. (36), the length of the optimization
area varies with cycle length, and a longer cycle cannot be chosen if the road conditions do not allow it.
2) Travel speed within intersections: In a fully CAV environment, all vehicles are committed to
maintaining the same speed to reduce confusion caused by speed differences. As a result, vehicles are
often expected to have the same travel speed at intersections and entry lanes. However, it is often
assumed that vehicles need to slow down during turning to gain more stability [52]. Therefore, the effects
of the speed difference caused by vehicle deceleration at the intersection on the overall traffic flow
became one of the main points of this paper.
Fig. 19 shows the trend of vehicle delay and fuel consumption with the degree of intersection
deceleration for different traffic demands. The horizontal axis is the travel speed within the intersection,
and the vehicle is designed to enter the control area at 15 m/s. The left vertical axis is the average delay
or fuel consumption for one vehicle, and the right vertical axis is the delay or fuel consumption reduction.
The remaining parameters in the experiment are shown in TABLE II.
33
Fig. 19. Sensitivity analysis of travel speed within intersections: (a) Low traffic demand; (b) Medium
traffic demand; (c) High traffic demand.
34
Fig. 20. Average instantaneous fuel consumption: (a) Low traffic demand; (b) Medium traffic demand;
(c) High traffic demand.
As shown in Fig. 19, deceleration at intersections causes extra vehicle delays and may lead to
oversaturation. In addition, the proposed method shows a decreasing trend in delay reduction
improvement over the FCFS-based method as the speed increase up to 15 m/s for both low and medium
traffic demand, but the reverse is true for high traffic demand. It indicates that: (1) deceleration at
intersections has a more significant impact on the FCFS-based method and that (2) the FCFS-based
method is strained to cope with high traffic demands even if there is no deceleration at intersections,
while the proposed method worked well under all three demands.
The proposed method not only reduces traffic delay, but also reduces fuel consumption. Fuel
consumption is influenced by many factors, such as travel time, travel speed, and acceleration or
deceleration processes. The travel time mainly acts on the cumulative fuel consumption, while travel
speed and acceleration impact instantaneous fuel consumption. Therefore, we first investigate the
instantaneous fuel consumption and then the cumulative fuel consumption. We defined the average
instantaneous fuel consumption as the ratio of the average fuel consumption to the average travel time,
as shown in Fig. 20. It can be seen that decelerations at the intersection reduce the average instantaneous
fuel consumption.
Fig. 21. Speed profiles: (a) Travel Speed within Intersections: 10 m/s; (b) Travel Speed within
Intersections: 14 m/s; (c) Travel Speed within Intersections: 15 m/s.
It is easy to explain. According to Eqs. (22)(23), frequent speed changes can lead to high fuel
35
consumption. When the travel speed within the intersection is low, most vehicles only need to slow down
to meet the speed requirements, as shown in Fig. 21 (a). However, when the speed is high, some vehicles
with large delays need to slow down first and then accelerate, as shown in Fig. 21 (b). When the travel
speed within the intersection is equal to the speed of entering the control area, as shown in Fig. 21 (c), all
vehicles have to slow down first and then accelerate, increasing the acceleration and deceleration process.
Besides, higher cruising speed also leads to higher fuel consumption. When the travel time is long, the
vehicle speed profile can be seen as a three-segment [53], with the longer the travel time, the lower speed
the vehicle will cruise at. That’s why the instantaneous fuel consumption of the FCFS-based method is
low when speed within intersections is approaching 10 m/s. As a result, high-speed driving and frequent
acceleration or deceleration will increase instantaneous fuel consumption.
As mentioned before, travel time affects cumulative fuel consumption. Therefore, a small average
instantaneous fuel consumption does not mean less cumulative fuel consumption, depending on travel
time, i.e., vehicle delay. When vehicle delay is small (e.g., less than 2 s), travel speed and acceleration are
the dominant factors in the increase of fuel consumption. Therefore, Fig. 19 shows that for low and
medium traffic demand under the proposed method, vehicle fuel consumption increases with an increase
in travel speed within the intersection. When vehicle delay is significant, it becomes the dominant factor.
Long travel times, caused by considerable delays, result in significant cumulative fuel consumption.
Therefore, under the combined effect of these factors, the trend of fuel consumption with travel speed
within the intersection changes from increasing (results of the proposed method under low traffic
demand) to decrease-then-increase (results of the FCFS-based method under medium traffic demand),
or even to a decreasing trend (results of the FCFS-based method under high traffic demand).
D. Test for Accelerated Solution Methods.

As analyzed before, a long cycle leads to a better solution but causes challenges in computation.
However, if we can accelerate the solving process, longer cycles can be used. To this end, this paper
proposes an additional timing schedule model in Section IV. We argue that M2 can sacrifice the quality
of the solution in exchange for efficiency. In this section, we verify the efficiency of the two models. In
M2, the threshold 𝜛𝑗 is defined as the time vehicle 𝑗 takes to pass through the intersection area and attach
a buffer time.
𝑟𝑗
𝜛𝑗 = 𝜛𝑗 (𝜖) = +𝜖 (38)
𝑣int
where 𝑟𝑗 is the route length of vehicle 𝑗 (the channel length), and 𝜖 is the buffer time. The buffer times are
set to 0 s and 1 s, corresponding to the model representations M2(0) and M2(1), respectively.
We eliminate the concept of cycles and consider only one optimization. In this optimization, the
total number of vehicles at the intersection is increased from 5 to 80 in intervals of 5 vehicles. Traffic
36
demand is set at 900 veh/h/lane to limit vehicle arrivals. Thus, a vehicle number of 80 corresponds to a
cycle time of 40 s. Ten experiments are conducted and the results are shown in Fig. 22.
Fig. 22 (a) and (b) show the average sum-up CPU time and average single-vehicle delay of
multiple experiments, respectively. By setting up a threshold to reduce some constraints, M2 is solved
faster than M1. And it can be seen from Fig. 22 (b) that the reduced constraint does not have much effect
on the model optimality. The gap between the optimal results and the result of M2(1) is less than 3.5%.
While the quality of the solution in M2(0) decreases to some extent, it is still better than that of the GA-
based method. Fig. 22 (c) and (d) display the fluctuation range of CPU time and delay under multiple
experiments, respectively. The CPU time of M1 exhibits uncertainty and wide fluctuations under large-
scale problems because M1 is NP-hard. After introducing a threshold, the CPU time fluctuation range of
the proposed model is compressed, indicating that M2 performs more consistently. In Fig. 22 (d), the
delay results obtained by the optimization-based methods are all more concentrated than those obtained
by the FCFS-based methods, which is consistent with the findings in Fig. 11.
Overall, the average results from subfigures (a) and (b) and the median results from subfigures
(c) and (d) both yield that reducing constraints with binaries can speed up the model solution, and a
suitable threshold can make a tradeoff between the delay result and the CPU time.
Fig. 22. Comparison of different timing schedule models: (a) Computation time; (b) Average delay.
37
VI. CONCLUSION AND FUTURE WORK
In the CAV environment, AIM is a state-of-the-art control strategy. Many scholars have studied
AIM and proposed different control methods to plan a collision-free trajectory for vehicles to cross an
intersection without stopping. However, all previous studies were lacking in some respects.
This paper designs an intersection modeling approach, combining the advantages of the tile-
based and conflict point-based approaches, to consider the vehicle's shape well while maintaining the
simplicity of the model.
Based on the designed intersection modeling approach, a two-stage method is proposed with the
idea of the virtual platoon, considering vehicle delay and fuel consumption. The first stage is a vehicle
timing schedule optimization model formed as a MILP to minimize traffic delay. Based on the original
model, an additional model with different constraints is designed with the aim of speeding up the
solution. The additional model sacrifices the quality of the solution for efficiency. The vehicle arrival
timing sequence is obtained by solving the MILP, and then a trajectory optimization model is constructed
based on it as the second stage. It is formed as a linear program. The output is the instantaneous
acceleration of each vehicle, from which the driving strategy is obtained. The two-stage model runs
continuously using a rolling optimization strategy, and the cycle length is set as an adjustable variable
to accommodate different levels of traffic demand.
Numerical simulations are performed to demonstrate the proposed method. A four-legged
intersection with two lanes on each side is taken as an example. To achieve high traffic efficiency, vehicles
are allowed to turn freely on any entry lane. The FCFS-based method and the GA-based method are used
as benchmarks. Results show that the proposed method ensures the safety and performs the best. It
outperforms the GA-based method in terms of computation time, while reducing delays by up to 89.48%
and fuel consumption by up to 46.84%, under different traffic demands compared to the FCFS-based
method. The proposed method can deal with the problem from a system perspective, while the FCFS-
based method is greedy. Meanwhile, the proposed method has a stable ability in controlling delays.
However, the computational burden of the proposed method is greater under symmetric traffic demand
due to the nature of a MILP. In trajectory planning, the proposed method can reduce fuel consumption
by optimizing the vehicle timing schedule. In addition, the proposed method can increase the capacity
of intersections compared to the FCFS-based method. The sensitivity analysis of the cycle length reveals
that the proposed method performs better at longer cycle lengths, but the improvement is no longer
significant when the cycle length reaches a certain level. Moreover, the sensitivity analysis of the travel
speed within intersections indicates that maintaining the same speed within the intersection as entering
the control area reduces delays, but it may not be the optimal strategy for eco-driving. As for speeding
up the model solution by adding a heuristic rule, setting a suitable threshold to reduce binaries can speed
up the solution to a great extent. Although the quality of the solution is reduced, it is still better than the
38
two benchmarks.
This paper presents a complete methodology for isolated signal-free intersection management
from modeling approach to method design. In future work, we are eager to propose a joint optimal
control strategy for a road network with multiple intersections. Moreover, this paper only changes the
constraints to accelerate the solution, and the solution algorithm will be improved later.
ACKNOWLEDGMENTS
The paper received research funding support from the National Natural Science Foundation of
China (52002339), the Sichuan Science and Technology Program (2021YJ0535, 2022YFG0152) and the
Guangxi Science and Technology Program (2021AA01007AA).
REFERENCES
[1] P. Bansal and K. M. Kockelman, “Forecasting Americans’ long-term adoption of connected and
autonomous vehicle technologies,” Transportation Research Part A: Policy and Practice, vol. 95, pp. 49–
63, Jan. 2017, doi: 10.1016/j.tra.2016.10.013.
[2] G. Guo and L. Wang, “Control Over Medium-Constrained Vehicular Networks With Fading
Channels and Random Access Protocol: A Networked Systems Approach,” IEEE Trans. Veh. Technol.,
vol. 64, no. 8, pp. 3347–3358, Aug. 2015, doi: 10.1109/TVT.2014.2360438.
[3] Z. Li, L. Elefteriadou, and S. Ranka, “Signal control optimization for automated vehicles at isolated
signalized intersections,” Transportation Research Part C: Emerging Technologies, vol. 49, pp. 1–18, Dec.
2014, doi: 10.1016/j.trc.2014.10.001.
[4] C. Yu, Y. Feng, H. X. Liu, W. Ma, and X. Yang, “Integrated optimization of traffic signals and vehicle
trajectories at isolated urban intersections,” Transportation Research Part B: Methodological, vol. 112, pp.
89–112, Jun. 2018, doi: 10.1016/j.trb.2018.04.007.
[5] Y. Feng, C. Yu, and H. X. Liu, “Spatiotemporal intersection control in a connected and automated
vehicle environment,” Transportation Research Part C: Emerging Technologies, vol. 89, pp. 364–383, Apr.
2018, doi: 10.1016/j.trc.2018.02.001.
[6] R. Niroumand, “Joint optimization of vehicle-group trajectory and signal timing_ Introducing the
white phase for mixed-autonomy traffic stream,” p. 17, 2020.
[7] G. Qian, M. Guo, L. Zhang, Y. Wang, S. Hu, and D. Wang, “Traffic scheduling and control in fully
connected and automated networks,” Transportation Research Part C: Emerging Technologies, vol. 126,
p. 103011, May 2021, doi: 10.1016/j.trc.2021.103011.
39
[8] C. Wuthishuwong, A. Traechtler, and T. Bruns, “Safe trajectory planning for autonomous
intersection management by using vehicle to infrastructure communication,” J Wireless Com Network,
vol. 2015, no. 1, p. 33, Dec. 2015, doi: 10.1186/s13638-015-0243-3.
[9] K. Dresner and P. Stone, “Multiagent traffic management: a reservation-based intersection control
mechanism,” New York, p. 8, 2004.
[10] B. Liu, Q. Shi, Z. Song, and A. El Kamel, “Trajectory planning for autonomous intersection
management of connected vehicles,” Simulation Modelling Practice and Theory, vol. 90, pp. 16–30, 2019,
doi: 10.1016/j.simpat.2018.10.002.
[11] D. Fajardo, T.-C. Au, S. T. Waller, P. Stone, and D. Yang, “Automated Intersection
Control:Performance of Future Innovation Versus Current Traffic Signal Control,” Transportation
Research Record, vol. 2259, no. 1. pp. 223–232, 2011.
[12] Z. Li, M. V. Chitturi, D. Zheng, A. R. Bill, and D. A. Noyce, “Modeling Reservation-Based
Autonomous Intersection Control in VISSIM,” Transportation Research Record, vol. 2381, no. 1. pp. 81–
90, 2013.
[13] C. Yu, W. Sun, H. X. Liu, and X. Yang, “Managing connected and automated vehicles at isolated
intersections: From reservation- to optimization-based methods,” Transportation Research Part B:
Methodological, vol. 122, pp. 416–435, Apr. 2019, doi: 10.1016/j.trb.2019.03.002.
[14] H. Xu, Y. Zhang, C. G. Cassandras, L. Li, and S. Feng, “A bi-level cooperative driving strategy
allowing lane changes,” Transportation Research Part C: Emerging Technologies, vol. 120, p. 102773, Nov.
2020, doi: 10.1016/j.trc.2020.102773.
[15] M. Ma and Z. Li, “A time-independent trajectory optimization approach for connected and
autonomous vehicles under reservation-based intersection control,” Transportation Research
Interdisciplinary Perspectives, vol. 9, p. 100312, 2021, doi: 10.1016/j.trip.2021.100312.
[16] M. W. Levin and S. D. Boyles, “Intersection Auctions and Reservation-Based Control in Dynamic
Traffic Assignment,” Transportation Research Record, vol. 2497, no. 1, pp. 35–44, 2015, doi:
10.3141/2497-04.
[17] M. Choi, A. Rubenecia, and H. H. Choi, “Reservation-based traffic management for autonomous
intersection crossing,” International Journal of Distributed Sensor Networks, vol. 15, no. 12, p.
155014771989595, Dec. 2019, doi: 10.1177/1550147719895956.
[18] L. Chen and C. Englund, “Cooperative Intersection Management: A Survey,” IEEE Trans. Intell.
Transport. Syst., vol. 17, no. 2, pp. 570–586, Feb. 2016, doi: 10.1109/TITS.2015.2471812.
[19]M. W. Levin and D. Rey, “Conflict-point formulation of intersection control for autonomous vehicles,”
Transportation Research Part C: Emerging Technologies, vol. 85, pp. 528–547, Dec. 2017, doi:
10.1016/j.trc.2017.09.025.
40
[20] I. H. Zohdy and H. A. Rakha, “Intersection Management via Vehicle Connectivity: The Intersection
Cooperative Adaptive Cruise Control System Concept,” Journal of Intelligent Transportation Systems,
vol. 20, no. 1, pp. 17–32, Jan. 2016, doi: 10.1080/15472450.2014.889918.
[21] H. Ahn and D. Del Vecchio, “Safety Verification and Control for Collision Avoidance at Road
Intersections,” IEEE Trans. Automat. Contr., vol. 63, no. 3, pp. 630–642, Mar. 2018, doi:
10.1109/TAC.2017.2729661.
[22] Z. He, L. Zheng, L. Lu, and W. Guan, “Erasing Lane Changes From Roads: A Design of Future Road
Intersections,” IEEE Trans. Intell. Veh., vol. 3, no. 2, pp. 173–184, 2018, doi: 10.1109/TIV.2018.2804164.
[23] G. Lu, Z. Shen, X. Liu, M. Nie, and Z. Xiong, “Are autonomous vehicles better off without signals at
intersections? A comparative computational study,” SSRN Journal, 2021, doi: 10.2139/ssrn.3812649.
[24] K. Dresner and P. Stone, “A Multiagent Approach to Autonomous Intersection Management,” jair,
vol. 31, pp. 591–656, Mar. 2008, doi: 10.1613/jair.2502.
[25] M. W. Levin, S. D. Boyles, and R. Patel, “Paradoxes of reservation-based intersection controls in
traffic networks,” Transportation Research Part A: Policy and Practice, vol. 90, pp. 14–25, Aug. 2016, doi:
10.1016/j.tra.2016.05.013.
[26] Z. Deng, Y. Shi, Q. Han, L. Lv, and W. Shen, “A Conflict Duration Graph-Based Coordination Method
for Connected and Automated Vehicles at Signal-Free Intersections,” Applied Sciences, vol. 10, no. 18,
p. 6223, Sep. 2020, doi: 10.3390/app10186223.
[27]L. Xu et al., “Cooperative Merging Strategy for Connected Vehicles at Highway On-Ramps,” J. Transp.
Eng., p. 11.
[28] J. Lee and B. Park, “Development and Evaluation of a Cooperative Vehicle Intersection Control
Algorithm Under the Connected Vehicles Environment,” IEEE Trans. Intell. Transport. Syst., vol. 13,
no. 1, pp. 81–90, Mar. 2012, doi: 10.1109/TITS.2011.2178836.
[29]E. R. Müller, R. C. Carlson, and W. K. Junior, “Intersection control for automated vehicles with MILP,”
IFAC-PapersOnLine, vol. 49, no. 3, pp. 37–42, 2016, doi: 10.1016/j.ifacol.2016.07.007.
[30] Z. Yao, H. Jiang, Y. Cheng, Y. Jiang, and B. Ran, “Integrated Schedule and Trajectory Optimization
for Connected Automated Vehicles in a Conflict Zone,” IEEE Trans. Intell. Transport. Syst., pp. 1–11,
2020, doi: 10.1109/TITS.2020.3027731.
[31] W. Wu, Y. Liu, W. Liu, F. Zhang, V. Dixit, and S. T. Waller, “Autonomous Intersection Management
for Connected and Automated Vehicles: A Lane-Based Method,” IEEE Trans. Intell. Transport. Syst.,
pp. 1–16, 2022, doi: 10.1109/TITS.2021.3136910.
[32] S. A. Fayazi and A. Vahidi, “Mixed-Integer Linear Programming for Optimal Scheduling of
Autonomous Vehicle Intersection Crossing,” IEEE Trans. Intell. Veh., vol. 3, no. 3, pp. 287–299, Sep.
2018, doi: 10.1109/TIV.2018.2843163.
41
[33] M. Barth and K. Boriboonsomsin, “Energy and emissions impacts of a freeway-based dynamic eco-
driving system,” Transportation Research Part D: Transport and Environment, vol. 14, no. 6, pp. 400–410,
Aug. 2009, doi: 10.1016/j.trd.2009.01.004.
[34] A. M. I. Mahbub, A. A. Malikopoulos, and L. Zhao, “Decentralized optimal coordination of
connected and automated vehicles for multiple traffic scenarios,” Automatica, vol. 117, p. 108958, Jul.
2020, doi: 10.1016/j.automatica.2020.108958.
[35] H. Xu, Y. Zhang, L. Li, and W. Li, “Cooperative Driving at Unsignalized Intersections Using Tree
Search,” IEEE Trans. Intell. Transport. Syst., vol. 21, no. 11, pp. 4563–4571, Nov. 2020, doi:
10.1109/TITS.2019.2940641.
[36] H. Xu, C. G. Cassandras, L. Li, and Y. Zhang, “Comparison of Cooperative Driving Strategies for
CAVs at Signal-Free Intersections,” IEEE Trans. Intell. Transport. Syst., pp. 1–14, 2021, doi:
10.1109/TITS.2021.3071456.
[37] B. Chalaki and A. A. Malikopoulos, “Optimal Control of Connected and Automated Vehicles at
Multiple Adjacent Intersections,” IEEE Trans. Contr. Syst. Technol., pp. 1–13, 2021, doi:
10.1109/TCST.2021.3082306.
[38] Y. Zhang and C. G. Cassandras, “A Decentralized Optimal Control Framework for Connected
Automated Vehicles at Urban Intersections with Dynamic Resequencing,” in 2018 IEEE Conference on
Decision and Control (CDC), Miami Beach, FL, Dec. 2018, pp. 217–222. doi: 10.1109/CDC.2018.8618871.
[39] A. A. Malikopoulos, C. G. Cassandras, and Y. J. Zhang, “A decentralized energy-optimal control
framework for connected automated vehicles at signal-free intersections,” Automatica, vol. 93, pp.
244–256, Jul. 2018, doi: 10.1016/j.automatica.2018.03.056.
[40] Y. Lu, X. Xu, C. Ding, and G. Lu, “A Speed Control Method at Successive Signalized Intersections
Under Connected Vehicles Environment,” IEEE Intell. Transport. Syst. Mag., vol. 11, no. 3, pp. 117–
128, 2019, doi: 10.1109/MITS.2019.2919638.
[41] A. Zhou, S. Peeta, M. Yang, and J. Wang, “Cooperative signal-free intersection control using virtual
platooning and traffic flow regulation,” Transportation Research Part C: Emerging Technologies, vol. 138,
p. 103610, 2022, doi: https://doi.org/10.1016/j.trc.2022.103610.
[42] B. Xu et al., “Distributed conflict-free cooperation for multiple connected vehicles at unsignalized
intersections,” Transportation Research Part C: Emerging Technologies, vol. 93, pp. 322–334, Aug. 2018,
doi: 10.1016/j.trc.2018.06.004.
[43] R. Akcelik, “Efficiency and drag in the power-based model of fuel consumption,” Transportation
Research Part B: Methodological, vol. 23, no. 5, pp. 376–385, Oct. 1989, doi: 10.1016/0191-2615(89)90014-
3.
42
[44] P. Typaldos, I. Papamichail, and M. Papageorgiou, “Minimization of Fuel Consumption for Vehicle
Trajectories,” IEEE Trans. Intell. Transport. Syst., vol. 21, no. 4, pp. 1716–1727, Apr. 2020, doi:
10.1109/TITS.2020.2972770.
[45] A. Naidu , DS and R. Kolmanovsky , I, “Optimal Control Systems,” Applied Mechanics Reviews, vol.
57, no. 1, pp. B3–B4, Feb. 2004, doi: 10.1115/1.1641776.
[46] D. Bertsimas and J. N. Tsitsiklis, Introduction to linear optimization, vol. 6. Athena Scientific Belmont,
MA, 1997.
[47] G. Lu, Y. Nie, X. Liu, and D. Li, “Trajectory-based traffic management inside an autonomous vehicle
zone,” Transportation Research Part B: Methodological, vol. 120, pp. 76–98, Feb. 2019, doi:
10.1016/j.trb.2018.12.012.
[48] X. (Jeff) Ban, J.-S. Pang, H. X. Liu, and R. Ma, “Continuous-time point-queue models in dynamic
network loading,” Transportation Research Part B: Methodological, vol. 46, no. 3, pp. 360–380, Mar. 2012,
doi: 10.1016/j.trb.2011.11.004.
[49]Jing Zhu and S. Roy, “Mac for dedicated short range communications in intelligent transport system,”
IEEE Commun. Mag., vol. 41, no. 12, pp. 60–67, Dec. 2003, doi: 10.1109/MCOM.2003.1252800.
[50] J. B. Kenney, “Dedicated Short-Range Communications (DSRC) Standards in the United States,” Proc.
IEEE, vol. 99, no. 7, pp. 1162–1182, Jul. 2011, doi: 10.1109/JPROC.2011.2132790.
[51] L. Gurobi Optimization, “Gurobi Optimizer Reference Manual.” 2021. [Online]. Available:
http://www.gurobi.com
[52] M. O. Sayin, C.-W. Lin, S. Shiraishi, J. Shen, and T. Basar, “Information-Driven Autonomous
Intersection Control via Incentive Compatible Mechanisms,” IEEE Trans. Intell. Transport. Syst., vol.
20, no. 3, pp. 912–924, Mar. 2019, doi: 10.1109/TITS.2018.2838049.
[53] W. Ma, J. Li, and C. Yu, “Shared-phase-dedicated-lane based intersection control with mixed traffic
of human-driven vehicles and connected and automated vehicles,” Transportation Research Part C:
Emerging Technologies, vol. 135, p. 103509, Feb. 2022, doi: 10.1016/j.trc.2021.103509.
43

A Two-Stage Optimization Method For Schedule and Trajectory of CAVs at An Isolated Autonomous Intersection

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Two-Stage Optimization Method For Schedule and Trajectory of CAVs at An Isolated Autonomous Intersection

Uploaded by

Copyright:

Available Formats

A Two-Stage Optimization Method for Schedule and Trajectory of CAVs at an Isolated

Zhihong Yao1,2,3, Haoran Jiang1,2*, Yangsheng Jiang1,2,3*, and Bin Ran4

*Correspondence to: Haoran Jiang (jianghaoran.joe@foxmail.com) and Yangsheng Jiang (jiangyangsheng@swjtu.edu.cn),

C. Contribution of the Paper

D. Organization of the Paper

II. DESIGN OF AN AUTONOMOUS INTERSECTION

Fig. 1. Illustration of an autonomous intersection.

B. The Process of Optimizing and Controlling the Vehicles

III. DESIGN OF THE INTERSECTION MODELING APPROACH

A. Comparison of the Tile-Based and Conflict Point-Based Approaches

Fig. 4. Gradually increase the granularity.

𝑥𝑗𝑒 (𝑡) − 𝑥𝑖𝑒 (𝑡) ≥ 𝑙𝑗 + 𝑣𝑖 (𝑡)𝛥𝑡, (3)

IV. TWO-STAGE OPTIMIZATION METHOD

Fig. 7. Framework of the two-stage method.

A. Timing Schedule Optimization

Fig. 8. Illustration of the virtual platoon.

min 𝑍 = min ∑(𝑡𝑖+ − 𝑡𝑖− − 𝑡𝑖∗ ), (4)

(𝑣𝑖max )2 − (𝑣𝑖− )2 (𝑣int )2 − (𝑣𝑖′ )2

𝑣𝑖′ − 𝑣𝑖− 𝑣int − 𝑣𝑖′

𝑡𝑖+ = 𝑡𝑖− + 𝑡𝑖′ . (9)

Fig. 9. Illustration of travel distance.

Joint Eqs. (6)-(9), we have

In the second situation,

𝑡𝑖+ = 𝑡𝑖− + 𝑡𝑖′ . (14)

Then, joint Eqs. (12)-(14), we have

∀ 𝑗 ′ ∈ [1, 𝑖), 𝜃𝑗′ = 𝜃𝑖 , ∄ 𝑗 ′′ ∈ (𝑗 ′ , 𝑖), 𝑠. 𝑡. 𝜃𝑗′′ = 𝜃𝑖 . (21)

where 𝑅𝑇 is the total tractive force,

min 𝐽 = ∑ ∑ |𝑎𝑖 [𝑘]|𝛿 . (24)

min 𝐽 = ∑ ∑ 𝑧𝑖 [𝑘]𝛿 , (25)

𝑧𝑖 [𝑘] ≥ 𝑎𝑖 [𝑘], (26)

𝑧𝑖 [𝑘] ≥ −𝑎𝑖 [𝑘]. (27)

𝑣𝑖 [𝑘] = 𝑣𝑖 [𝑘 − 1] + 𝑎𝑖 [𝑘 − 1]𝛿, ∀𝑘 ∈ [𝑘𝑖− + 1, 𝑘𝑖+ ], (29)

𝑥𝑖 [𝑘𝑖− ] = 0, 𝑣𝑖 [𝑘𝑖− ] = 𝑣𝑖− , 𝑎𝑖 [𝑘𝑖− ] = 0, (32)

𝑥𝑖 [𝑘𝑖+ ] = 𝐿𝐶 , 𝑣𝑖 [𝑘𝑖+ ] = 𝑣int , 𝑎𝑖 [𝑘𝑖+ ] = 0. (33)

𝑥𝑖−1 [𝑘] − 𝑥𝑖 [𝑘] ≥ 𝑙𝑖−1 + 𝑣𝑖 [𝑘]𝛥𝑡,

C. Rolling Optimization Strategy

B. Test the Effectiveness of the Proposed Method

Fig. 12. Different single-vehicle delays.

∑(𝑥 − 𝑥̅ )(𝑦 − 𝑦̅)

D. Test for Accelerated Solution Methods.

You might also like

Zhihong Yao1,2,3, Haoran Jiang1,2, Yangsheng Jiang1,2,3, and Bin Ran4