DOI:10.1109/MITS.2020.3014074 (IN PRESS) 1

Autonomous and Semi-Autonomous IntersectionManagement: A Survey

Zijia Zhong, Member, IEEE, Mark Nejad, Member, IEEE, and Earl E. Lee, Member, IEEE

Abstract—Intersection is a major source of traffic delaysand accidents within modern transportation systems. Comparedto signalized intersection management, autonomous intersectionmanagement (AIM) coordinates the intersection crossing at anindividual vehicle level, which provides additional flexibility.AIM can potentially eliminate stopping in intersection crossingdue to traffic lights while maintaining a safe separation amongconflicting movements. In this paper, the state-of-the-art AIM re-search among various disciplines (e.g., traffic engineering, controlengineering) is surveyed from the perspective of three hierarchi-cal layers: corridor coordination layer, intersection managementlayer, and vehicle control layer. The key aspects of AIM designsare discussed in details, including conflict detection schemes,priority rules, control centralization, computation complexity,etc. The potential improvements for AIM evaluation with theemphasis of realistic scenarios are provided. This survey servesas a comprehensive review of AIM design and provides promisingdirections for future research.

Index Terms—Connected and automated vehicle, Autonomousintersection management, Vehicle control, Trajectory planning,Priority policy, Measure of effectiveness

I. INTRODUCTION

Intersection is a major source of traffic delays and accidents.According to the National Motor Vehicle Causation Surveyconducted between 2005 and 2007, 36% of the surveyedcrashes (i.e., 2,188,969) were intersection related in the UnitedStates. Among them, inadequate surveillance (44.1%) andfalse assumption of other’s actions (8.4%) were the mostfrequent culprits [1]. In 2010, crashes at intersection in theUnited States amounted to US$ 120 billion in economic costsand US$ 371 billion in societal cost [2].

Connected and Automated Vehicles (CAVs) are expected toassume a revolutionary role in mitigating traffic accidents andcongestion. CAVs encompasses connected vehicles (CVs) andautomated vehicles (AVs). The former relies on the two-waywireless communication, which enables real-time informationsharing and cooperation among agents within a transportationsystem, whereas the latter eliminates human driver errorsthat may potentially cause crashes, traffic flow oscillations,and shock-waves. Thus far, the CAV-based improvements forintersection traffic management can be categorized into twogroups: i) the incorporation of real-time, high-resolution CAVtraffic data in signalized intersections management (SIM) for

Z. Zhong is with the Center for Integrated Mobility Science at the NationalRenewable Energy Laboratory. The work was performed when he was with theDepartment of Civil and Environmental Engineering, University of Delaware,e-mail: zijia.zhong@ieee.org.

M. Nejad and E. E. Lee are with the Department of Civil and EnvironmentalEngineering, University of Delaware, Newark, DE, 19716 USA, e-mail:{nejad, elee}@udel.edu

enhancing the signal phase and timing (SPaT) plans and ii) thedevelopment of signal-free autonomous intersection manage-ment (AIM) that is made possible by vehicle automation andconnectivity.

A. Scope of the Survey

Insofar, there are several notable survey papers in theliterature focusing on intersection management. Guo et al. [3]conducted a detailed review of the integration of the CAVdata into signalized intersection management. Guo et al.’ssurvey emphasized the integration of CAV technology into theexisting signal control framework, for instance, vehicle platoonfor signalizing. With the focus on the mixed-traffic conditionand CAV-enabled augmentations, they chose to exclude signal-free intersection control, which is the focus of this paper.Rios-Torres and Malikopoulos [4] reviewed the intersectionmanagement and on-ramp merging from the centralized anddecentralized control perspectives. Chen and Englund [5]reviewed studies on cooperative intersection management andhighlighted the major AIM research. Li et al. [6] surveyedthe traffic control system with the focus of contrasting pref-erence, such as global planning-based versus local and self-organization-based control.

A systematic review that provides the overall landscape ofAIM, including design philosophies, evaluation approaches,and cross-discipline perspectives still lacks, in spite of thesignificant potentials of AIM. Additionally, AIM studies havebeen steadily emerging in the recently years owing to the rapiddevelopment of CAV technology. There are numerous papersregarding using CAV to enhanced intersection performance.The scope of this review, however, is confined to the studiesthat deal with coordinating conflicting vehicle movements.Hence, intersection studies without dealing with conflictingintersection movements are excluded, such as eco-approachand departure of intersection.

B. Contribution and Organization of the Paper

The primary contribution of the paper is the survey ofthe state-of-the-art research on AIM from a multi-disciplinaryperspective. More specifically, this survey focuses on thefollowing areas:

• the transition from SIM to AIM• intersection conflict planning• vehicle control for AIM• AIM evaluationThe remainder of the paper is organized as follows: Sec-

tion II introduces the hierarchical intersection management

arX

iv:2

006.

1313

3v2

[cs

.MA

] 6

Aug

202

0

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 2

framework that encompasses both AIM and SIM. Section IIIreviews the key design aspects of AIMs. The studies on thevehicle control layer are reviewed in Section IV, followed bythe evaluation scenarios for AIM in Section V. Discussion ofthe future research trends are presented in Section VI, followedby Conclusion in Section VII

II. INTERSECTION MANAGEMENT

Ensuring safety by separating the time-space conflictsamong approaching vehicles is of the utmost importance forintersection management. This is the case for both SIM andAIM. A conflict point is the intersection of two vehicle trajec-tories, where a collision could potentially occur. A standardintersection (with 4 approaches and 12 movements) has 16conflict points as illustrated in Fig. 1(a). Roundabout is a typeof non-standard intersection that are with a different set ofconflict points, as shown in Fig. 1(b).

(a) standard intersection

(b) roundabout

Fig. 1: Intersection conflict points (source: [7])

A. Hierarchical Layers

As illustrated in Fig. 2, three hierarchical layers can bedistilled from existing intersection control practices in trafficengineering, which are i) corridor coordination layer, ii) tra-jectory planning layer, and iii) vehicle control layer, The hi-erarchical framework is also suitable for evaluating AIM. Theconnections among the layers are enabled by communicationnetworks. In SIM, magnetic loop detectors collect prevailingtraffic conditions for the signal controllers that host the inter-section management protocol (e.g., SPaT plans). The humandrivers are notified with crossing permission by traffic lights.The AIM replaces the aforementioned procedure with V2Xcommunication. Furthermore, the human control of vehiclesis anticipated to be replaced by advanced driving assistancesystems (ADAS) and eventually by automated driving systems(ADS) in AIM.

Fig. 2: Intersection control layer

Corridor coordination layer deals with the coordinationof multiple intersections at a corridor level. Such coordinationis common for major arterials under SIM. Maximization of thegreen band [8] is commonly used to ensure the progressionof the major through movements across multiple intersections.Fixed-time SPaT plans among intersection are the most com-mon approach to achieve coordination.

Intersection management (trajectory planning) layerassigns the crossing sequence for vehicles in AIM or vehiclegroups in the case of SIM for an intersection. For SIM,the signal phases that are with conflicting movements arecycling based on a predefined phase sequence. For AIM, theintersection manager is responsible for allocating the limitedtime-space resources of an intersection. This aspect of AIMis often referred as trajectory planning since AIM separatesconflict movements at the level of individual vehicles.

Vehicle Control layer focuses on the motion control foran individual vehicle both longitudinally and laterally. Con-ventionally, vehicles are driven by a human who is primarilyresponsible for the movement of a vehicle, sometimes with theaid of vehicle sub-systems such as power steering and assistedbraking. Under the CAV environment, an automated drivingsystem is envisioned to complement and ultimately replace thedriving inputs from a human.

B. Roadmap for AIM

The transition to full CAV penetration could take decades.Therefore, the three layers (illustrated in Table I) could beassumed by various entities and levels of automation, as the

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 3

development of CAV technologies progresses. The Society ofAutomotive Engineers International (SAE) defined six levelsof vehicle automation from 0 (zero) being solely-driven bya human to 5 (five) in which the automation is in effect inany conditions [9]. Thus far, with only a few exceptions, theAIM is generally not compatible with HVs. However, semi-AIMs has been proposed for the mid-level automation (Level2 or Level 3) during the transition to full CAV penetration.The transition from SIM to AIM is accompanied by a gradualdecrease in human involvement in various dynamic drivingtasks.

TABLE I: Intersection Management Roadmap

Hierarchical LayerLevel ofAutomation Vehicle Control Intersection

ManagementCorridor

Coordination

SIM (Lv. 0) human driver SIM pre-timed orhardwired connection

SIM (Lv. 1) human driver+ ADAS

SIM +Vehicle Info.

pre-timed orhardwired connection

SIM (Lv.2) human driver+ ADAS

SIM +Vehicle Info.

pre-timed orhardwired connection

SIM (Lv.3) human driver+ ADAS

SIM orSemi-AIM

hardwired connection orwireless communication

Simi-AIM (Lv.4) human driver+ ADAS

SIM orSemi-AIM

hardwired connection orwireless communication

AIM (Lv.5) ADS AIM wireless communicationADAS: advanced driver-assistance systemsADS: automated driving systems

III. AUTONOMOUS INTERSECTION MANAGEMENT

An isolated AIM is comprised of two layers: the trajectoryplanning layer and the vehicle control layer. To differentiateintersection management with SIM, we use the term “tra-jectory planning” exclusively for AIM in this paper. Priorityassignment and a reservation system are the two key aspectsof trajectory planning, which are discussed in this section. Thevehicle control layer is discussed in Section IV.

A. Time-space Reservation

The vehicle trajectory assignment is akin to aircraft sepa-ration planning in the time-space dimension. Fig. 3 visualizesthe shared use of the finite intersection time-space resource ofan all-way stop control (AWSC) intersection with a standard4-leg layout. There are three vehicles crossing the intersection.Each line represents the trajectory of a individual vehicle.A crossing assignment is feasible as long as the trajectorylines (plus a safety buffer) do not intersect with each other.The separation in AIM is conducted at vehicle level withreservation-based systems, whereas the separation in SIM isensured at the vehicle group level with traffic signals.

There are four reservation systems for separating conflictsin AIM: i) intersection-based reservation [10], ii) tile-basedreservation [11], iii) conflict point-based reservation [12], andiv) vehicle-based reservation [13]. Intersection-based reserva-tion allows one and only one vehicle within an intersection. Intile-based reservation (shown in Fig. 4(b)), space is discretizedinto a grid of tiles. A reservation is rejected if two vehiclesoccupy the same tile at the same time. Tiles could alsobe grouped into bigger regions to reduce the computationcomplexity for the reservation [14]. The conflict point-based

Fig. 3: Vehicle-level conflict separation

reservation, shown Fig. 4(c), is able to take full advantage ofthe intersection space [12]. Li et al. [13] proposed a radicallydifferent vehicle-based reservation system that is able to guideall the CAVs within a standard intersection (including theuse of opposing travel lanes), provided collision is avoided.The vehicle-based reservation, as demonstrated in Fig. 4(d),is the least restrictive reservation system; however, it im-poses demanding computational expense to solve the nonlinearprogramming (NLP) problem due to the high-dimensionalcollision avoidance constraints.

The trade-off between computational tractability and uti-lization of intersection space was commonly tailored to theresearch need. Insofar, nearly all the studies for AIM dealt withstandard intersections, in which there are three movementsassociated with each of the four approaches. Table II showsthe reservation systems in the previous studies. The majority ofthe AIMs included twelve (full) movements [15–18]. However,it is not uncommon to make simplifications by reducingthe movements to only four [19–23], or even two [24–26]during the concept development of an AIM. To the best ofour knowledge, no computational deadline was set for thetrajectory planning among the reported studies, which meansthe AIMs were allowed to take as much time for solving theenter sequence.

B. Priority Policy

The priority policy dictates the allocation of the intersectionresources. It is independent of the reservation framework,whose main objective is to separate conflict movements. InSIM, the priority among vehicle groups is determined by op-erational needs (e.g., queue length or delay) and implementedvia the SPaT plan. Priority can be increased by extending thegreen signal timing for movement groups of interest.

Different from SIM, the priority assignment of AIM is onan individual vehicle level. The only vehicle-level priorityassignment on SIM is transit- or emergency- vehicle signalpreemption. However, it only makes exceptions for a few spe-cial vehicles. When it comes to AIM, the first-come-first-serve(FCFS) policy, which is based on fairness, has been adopted inthe majority of the AIM research. The system-optimal policyis the second most-used priority policy where the crossingsequence is determined based on system-level performance

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 4

TABLE II: AIM Reservation System

Literature Reservation System Movementintersection-based tile-base conflict point-based vehicle-based

Fajardo et al. [27] X 12Carlino et al. [28] X 12Bashiri and Fleming [21] X n/aBashiri et al. [29] X 12Jiang et al. [30] X n/aKamal et al. [12] X 12Muller et al. [25] X 2Levin and Rey [31] X 12Zohdy et al. [19] X 4Du et al. [32] X 2Ding et al. [15] X 12Bashiri et al. [33] X 12Li et al. [13] X unlimitedJin et al. [26] X 2Mirheli et al. [34] X 8Li and Zhang[35] X unlimitedWuthishuwong et al.[36] X 12Liu et al. [17] X 12Zohdy and Hakha [18] X 12Zhao et al. [37] X 12Hassan and Hakha [38] X 4Stone et al. [39] 12Fayazi et al.[22] X 4Lam and Katupitiya [24] X 2Creemers et al. [40] X 5

(a) intersection-based (b) tile-based

(c) conflict point-based (d) vehicle-based

Fig. 4: Types of reservation for intersection management

measures, such as overall delay, throughput, travel time, etc.Other priority policies have been reported, such as longest-queue-first policy[41], vehicle type-based policy [42], custompriority score-based policy [17], and auction-based policy [28].

Table III lists the major AIM study with regard to the cover-age of the three-layer intersection management structure. Fewstudies of AIM investigated the corridor-level coordination ofAIMs. An increasing amount of AIM studies that coupled

an explicit vehicle dynamics model was observed. Game-theoretic priority policy [19, 43] is a popular choice among theheuristic methods. Platoon-based performance metrics [21, 33]is another type of popular heuristic methods.

1) Fairness-based Priority Policy: The FCFS policy hasbeen widely adopted in existing studies. First-in-first-out(FIFO) is an alternative term that has been used. The FCFSsequence is determined by two criteria: i) the estimated ar-rival time to the infrastructure-to-vehicle (V2I) communicationboundary of the intersection (i.e., stop line), ii) the arrivalof the communication boundary. If the identical arrival timesof multiple vehicles are identified, the right-of-way could beassigned to the vehicle on this right (in right-hand drivingcountries). This rule has been in practices to resolve the con-flict arisen from AWSC intersection when two vehicles arriveat the stop sign at the same time, albeit more sophisticatedrule is certainly possible for AIM.

Algorithm 1 generalizes the conflict point-based FCFSreservation system. The algorithm is also applicable to non-standard intersection layouts, for instance the diverging dia-mond interchange (DDI), as long as the conflict points areidentified. N is the total number of vehicles that need to crossan intersection; tc2k is the time at conflict point c2 for vehiclek; δ is minimum distance; S is the distance to be traversedwithing the intersection; tinj is the time vehicle j enters theintersection; toutj is the time vehicle j exits the intersection;tinj,k is the enter time for vehicle j considering the conflict withvehicle k.

FCFS is able to achieve good performance under certaincircumstances. Fajardo et al. [27] compared the FCFS-basedAIM with SIM and found that the FCFS protocol signifi-cantly outperformed SIM in various testing scenarios. Thecomparison was among three FCFS-AIMs (with combinations

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 5

TABLE III: Layer Coverage of AIM Studies

Study Reservation Layer Prioritycorridor intersection vehicle FCFS system optimal heuristic

Jiang et al. [30] IB X XBichiou and Rakha[14] TB X X XLiu et al.[17] TB X XDing et al.[15] CP X XFajardo et al.[27] CP X XHassan and Hakha[38] IB XKamal et al.[12] (VICS) CP X X XStone et al.[39] TB+SIM XLevin and Rey [31] (CPIC-AIM) CP X X XLam and Katupitiya[24] IB X XLi et al.[13] VB X X XDu et al.[32] IB X X X XJin et al.[26] CP X XZhao et al.[37] CP X XBashiri and Fleming[21] IB X XBashiri et al.[33] IB X XMuller et al.[25] IB X XWuthishuwong et al.[36] CP X XMirheli et al.[34] CP X XCarlino et al.[28] IB X X XCreemers et al.[40] CP X X X

Algorithm 1 Conflict point-based FCFS-AIM scheduler

initialization . get intersection info. (e.g., conflictpoints)get the intersection entry sequencefor 1 < j < N do

for k < j doidentify conflict point cj,k between vehicle j and

k, cj,k ∈ {1, 2, · · · , Qj}obtain the entry time tink of vehicle ktinj,k = tQk + S

vintx+ δ . assuming Q is the

conflict pointend fortinj = max

{tinj,1, t

inj,2, · · · , tinj,k

}. get the most

conservative entry timesubsequently update and store tinj , t

1j , t

2j , · · · t

Qj

j , toutj

end for

of a static buffer, internal time buffer, and edge time buffer,respectively) and SIM (with single protected left-turn phase)under low, medium, and high volume scenarios.

There are several issues with the FCFS policy. First, theFCFS may impose an external cost for other vehicles withhigher priority due to its priority-agnostic nature. Imaginean extreme case where an emergency vehicle is at the backof the queue, waiting to clear the intersection, such vehicleis only granted permission to enter the intersection after allthe preceding vehicles in the FCFS queue. The handling ofemergency vehicle was reported in a handful of studies. Dreserand Stone [42] proposed the FIFS-EMERG AIM, an augmen-tation of existing FIFS-AIM that grants priority to the laneof the incoming emergency vehicle. The lane-level priority isimplemented to ensure that the non-emergency vehicles donot stop on the travel lane, which could potentially blockthe emergency vehicle. It was found that the FIFS-EMERGyielded lower average delay for emergency vehicles. Thehandling of emergency vehicle is relatively straight forward

for non-FIFS-AIM, such as by increasing emergency vehicle’spriority score [44] or giving it a virtually infinite budget in theauction-based AIM [28]

Second, a reservation is meaningful only if the requestingvehicle is able to execute it. In another words, a vehicle in aqueue may not request a reservation until it is able to enteran intersection. This suggests that an intersection approachwith more lanes is likely to obtain a greater share of theintersection capacity as more vehicles in the front of the lanescan request reservations at the same time. Furthermore, theFCFS policy assigns equal weight to all approaches, whichmeans a vehicle on a minor approach can break the progressionof vehicle platoons on the major approach. Third, FCFS doesnot strictly maintain the order of entry and the order ofreservation requests. For instance, let i, j, k be the indexedvehicles in the FCFS queue in ascending order. Vehicle iobtains a reservation, whereas vehicle j got rejected due toconflict with vehicle i, but vehicle k was accepted in theabsence of conflicts with vehicle i and j. In this case, wehave an entry sequence of [i, k, j], different from the FCFSorder [i, j, k].

Levin et al. [45] presented a theoretical example of the ex-ploitation of the FCFS policy and demonstrated the superiorityof SIM to FCFS-AIM. The simulation of an arterial network(with 5 signalized intersections and 21 links) revealed thatthe AIM was outperformed by a traffic signal in all demandlevels with the exception of under low demand scenarios.The FCFS-AIM was subsequently evaluated on a large-scaleurban network (i.e., downtown Austin, TX), where all SIMswere replaced with FCFS-based AIM. With the additionalassumption of user equilibrium route choice, interestingly, thesuperior performance of FCFS-based AIM was observed in theurban grid network: the overall travel time decrease by over50% in all scenarios. Levin et al. concluded that it was theavailability of parallel links in an urban network, combinedwith the user equilibrium route choice, that evenly distributedtraffic to avoid high delay intersections, despite the theoretical

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 6

disadvantages of the FCFS policy.A probabilistic model based on the turning ratio of a

standard intersection was put forward to theoretically computethe saturation flow rate of a conflict point-based AIM. Thecapacity was found to be 1,667 vehicle per hour (vph) assum-ing no turning movements, compared to the capacity rangeof 1700-1900 vph for SIM [46]. Zhang and Cassandras [47]compared the FCFS-AIM with decentralized optimal controlat vehicle level with SIM. They concluded that as trafficgrows, a higher CAV penetration is necessary to match theperformance of SIMs. 750 vph per lane was considered as thecritical flow rate. When the traffic demand is above 750 vphper lane, even assuming 100% CAV penetration, FCFS-AIMstill cannot outperform SIM in terms of energy saving. Undersaturated condition, nearly all vehicles have to slow downor even stop to create the necessary separation for enteringthe intersection. However, non-signalized coordination is moreeffective in reducing travel delays than signalization.

2) System-optimal Priority Policy: FCFS is not likely toproduce the system optimal solution, as it does not explicitlyoptimize the global intersection performance [18]. A trajectorymanagement layer for the entire intersection can be addedto relax the FCFS policy. Lee and Park [23] put forward acooperative vehicle intersection control (CVIC) system whichdid not require any signalization even under moderate inter-section demand (1,900 vph). In CVIC, each passing vehiclewith potential conflict with another vehicle was assigned withan individual trajectory to minimize the trajectory overlappingwithin an intersection. A general trajectory planning for mini-mizing the overlapping of vehicle trajectories is shown in (1).

minv(t)

J =

P∑φ=1

L∑l=1

N∑n=1

∫ p

q

√(1 + vn(t))2)dt (1a)

subject to: (1b)an,min ≤ an(t) ≤ an,max, (1c)

0 ≤ vn,min ≤ vn(t) ≤ vn,max, (1d)τ < xn(t)− xn+1(t) (1e)

where, P is the total number of the movement types; φ isthe movement type index; L is the total number of lanesof movement φ with lane index l; N is the total number ofvehicles; n is the vehicle index; vn(t) is the time-dependentspeed of vehicle n; p is the arrival time at the beginning ofintersection; q is the exiting time at the end of the intersection;vn,max is the maximum speed; vn,min is the minimum speed;an,max is the maximum acceleration; an,min is the minimumacceleration; xn(t) is the position of vehicle n; τ is theminimum headway on consecutive vehicle on the same lane.

A centralized cooperative intersection control was proposedby Ding et al. [15]. The control strategy was designed tominimize intersection delay, fuel consumption as well asemission, while avoiding a collision by separating the arrivaltime for conflict vehicle at each conflict point. Fayazi etal. [22] formulated the intersection management as a mixed-integer linear program (MILP) to maximize the number ofvehicles that clear the intersection within a given interval and,

at the same time, minimize the difference between desiredarrival time and assigned time. A signal-free intersectioncontrol logic (SICL) was proposed in [34] to maximize theintersection throughput. Dynamic programming was employedto find the near-optimal trajectory under safety constraints.Later the framework was improved with a cooperative vehicle-level structure to account for CAV preferences and scalability[48]. Ding et al. [15] proposed a multi-objective optimizationmodel for minimizing delay, emission, and discomfort level.Kamal et al. [12] proposed a vehicle-intersection coordinationscheme (VICS), which used a risk score as the objective. Therisk score quantitatively indicated the potential risk of collisionat a time step for a vehicle pair based on the overlapping areaof the two-dimensional Gaussian functions.

3) Heuristic Priority Policy: To account for service pri-orities, Liu et al. [17] proposed an intersection managementframework called TP-AIM. Here, a window searching algo-rithm (illustrated in Fig. 5) was proposed to find the entrywindow that yields a collision-free trajectory, while factoringin the service priority (e.g., emergency vehicles, heavy-dutytrucks, school buses) as well as the vehicle-based score. Thevehicle-based score was determined by the distance to theintersection, headings, etc. The assignment for vehicles withlower service priority had to first take into account the vehicleswith higher service priority.

Fig. 5: Window searching for a vehicle going straight (source:[17])

Game-theoretic approaches have been incorporated intoAIM. To put into a transportation perspective, all CAVscould potentially form a cooperative game along with theintersection controller via V2X communication. Elhenawyet al. [43] proposed a game-theory-based algorithm, whereCAVs communicate vehicle status (i.e., speed and location)to a centralized intersection manager. In the proof-of-conceptsimulation, two sets of vehicles (north-sound and east-west)were classified as two players, each of whom tried to minimizetheir delay at the intersection. Each player had three options: toaccelerate, to decelerate, and to maintain current speed. Uponobtaining vehicle information, the intersection manager solvesthe game matrix and obtains the Nash equilibrium. Then, theoptimal actions were distributed to each vehicle. Comparedto an AWSC intersection, the proposed scheme achieved 49%and 89% reduction in vehicle travel time and delay, respec-tively. A CACC-CG (Cooperative Adaptive Cruise Control -Cooperative Game) was proposed in [19]. The CACC-CG iscomprised of a manager agent and reactive agents at eachtime step. The manager agent selected one reactive agent for

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 7

movement optimization. The reactive agents, using symmetricinformation that is shared among players, choose amongacceleration, deceleration, or maintaining current speed. Allthe players choose the minimum utility value from the payofftable at each time step to achieve an equilibrium state.

The platoon-based reservation has steadily gained recogni-tion. In [21], the benefits of forming platoon among crossingvehicles were studied. The study proposed using platoonleaders to communicate on-behalf of followers to decreasecommunication complexity. The Platoon-based Delay Mini-mization cost function and the Platoon-based Variance Min-imization cost function were formulated for scheduling thecrossing sequence. In a subsequent study, Bashiri and Fleming[21] introduced a reservation policy that minimized delayor optimized schedules. To tackle the exponential nature ofthe permutation on the crossing sequence (i.e., (O(NN )), aheuristic method that ignores the non-conflicting trajectories,was proposed to reduce the computation complexity to O(N !).

C. Centralization

“Centralized”, “decentralized”, and “distributed” are theterms used for describing the organization of AIM systems.The information flow and the corresponding organization ofthe three types of planning is illustrated in Fig. 6.

Centralized AIM has a single-point contact among nodes(vehicles) for information sharing and decision making, asshown in Fig. 6(a). As such, single-point failure is the primaryconcern for centralized AIM. Early AIMs relied on centralizedcontrol strategies where the intersection manager guided theCAVs to safely traverse the intersection. Centralized intersec-tion management strategies are costly to implement, and theirscalability is open to question [38]. The current state of V2Xwireless communication may not technologically guaranteesuch performance with thousands of vehicles in the vicinityof an intersection.

Decentralized AIM contains several central hubs within thesystems. Note that the processing in the distributed system isshared across multiple nodes. For instance, for the platoon-based AIM [21, 33], the platoon leader acts as the decen-tralized hub to communicate with the intersection manager toobtain permission to enter the intersection. The intra-platooncommunication for platoon following is assumed to conductlocally among the platoon members as demonstrated in Fig.6(b).

Distributed AIM is an extreme case for decentralization.In a distributed system, as exhibited in Fig. 6(c), there isnot a single point where the decision is made, and eachnode makes a decision for its own behavior. The systembehavior is the result of the aggregated response for eachnode within the system. The distribution of the schedulingamong vehicles has the potential of becoming truly fault-tolerant. Hassan and Rakha [38] proposed a fully-distributedheuristic intersection control strategy that aims to minimize thecommunication (information exchanges) in each time step. Thevehicles approaching the intersection are categorized into fourgroups (“Out,”, “Last”, “Mid”, and “Head”), the group closestto the intersection (the “Head” group) assumes the role of

the schedulers, which are responsible for the passages of theintersection of all vehicles at different non-conflicting times.

D. Summary

The trajectory planning for ensuring safety is the corefunction of AIM. The vehicle-level crossing assignment allowsa great deal of flexibility, but at the same time increases thesearch space of entry sequence drastically. Trajectory planningis usually formulated as a nonlinear, non-convex problem [49]in order to fully satisfy the collision avoidance requirement.Mixed-integer programming (MILP) has been seen in manyformulations of AIM [22, 31, 44, 48]. The FCFS trajectoryplanning provides a simple and fair way to gain intersectionaccess. However, its efficiency is open to debate. Efforts havebeen made to factor in the priority of different vehicles.

IV. VEHICLE CONTROL FOR AIM

AIM research from the traffic engineering perspective typ-ically assumes the availability of the vehicle control andemphasizes on coordinating conflicting crossing movements.The roadside unit (RSU) take over the control of the vehicleand guide it to safely cross the intersection. Intersectioncontrol, coupled with vehicle control from control engineering,is a promising and necessary direction in AIM research. Withvehicle automation, the driving function is expected to bereplaced by the vehicle controller which have been activelystudied from control engineering perspective. As shown inFig. 2, the control of individual vehicle is classified as thelowest layer in the hierarchical framework for AIM. A second-order dynamics model is typically used for vehicle control inrelevant AIM studies, as expressed in (2), where pi(t) is theposition of vehicle i at time t and vi(t) is the speed of vehiclei at time t

pi(t) = vi(t), vi(t) = ai(t) (2)

Fig. 7: Intersection zones (source: [50])

A. Optimal Control

The underlying concept of optimal control is to find acontrol strategy that yields minimum cost for the associated

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 8

(a) centralized AIM (b) decentralized AIM (c) distributed AIM

Fig. 6: Communication structure of of AIM

process while satisfying the applicable control and state con-straints [51]. A basic optimal control framework for a CAVis expressed in 3, which aims at minimizing the performanceindex J (a.k.a., cost function). The configuration of the per-formance index J influences the optimal control of a vehicle.The common performance indexes include control input u2i (t),acceleration a2i (t), jerk a2i (t), derivative of jerk a2i (t), and

evacuation time (expressed as∫ tfit0idt).

minui(t)

J =

∫ tfi

t0i

U(x, t)dt (3a)

Subject to: (2), (3b)ui,min ≤ ui(t) ≤ ui,max, (3c)ai,min ≤ ai(t) ≤ ai,max, (3d)

0 ≤ vmin ≤ vi(t) ≤ vmax, ∀t ∈ [t0i , tfi ], (3e)

and given t0i , tmi , t

fi , pi(t

0i ) = 0, pi(t

fi ) = L, vi(t

0i ), ui(t

0i ),

where t0i is the entry time for vehicle i to the control zone(Fig. 7); tmi is the entry time to the intersection; tfi is thetime for clearing the intersection; ui(t) is the control inputfor vehicle i at time t; pi is the location in the control zonewith length L; all other variables are as previously defined.

Decentralized optimal vehicle coordination was proposedfor intersection management by Jiang et al. [30]. Underthis framework, each vehicle solves its own optimal controlproblem and then exchanges arrival and exiting time withneighboring vehicles. The entry sequence of the vehicles wasassumed to be available. With ten vehicles, the non-convexproblem can be reliably solved. However, the algorithm maynot scale well in high volume scenarios as more collisionavoidance constraints may become active, making the problemcomputationally intractable.

The co-design of optimal vehicle controls and crossingscheduling for intersection is complex with little availablemethods [52]. The majority of the studies used an upperintersection management layer to assigned a collision-freeentry sequence to the intersection. Zhang et al. [10] proposedan optimal intersection control framework with intersection-based FCFS policy. An intersection manager scheduled the

entry time of each CAV, which separated the conflicts amongdifferent movements within the intersection. With the assignedentry time, a vehicle then executes the optimal control strat-egy, which is derived by Hamiltonian analysis [53] with theassumption that none of the constraints was active within[t0i , t

mi ]. The consideration of left and right turning movements

was introduced in [54]. The AIM framework consideringturning movements with optimal control was presented inAlgorithm 2. Their system solves the control strategy for anindividual vehicle, but it did not implicitly coordinate thescheduling among different conflicting movements. The FCFSsequence may need to update depending on the movements ofconsecutive vehicles set in the FCFS queue.

Zhang et al. [50] later proved the existence of a non-empty set of initial conditions that keep the collision avoidanceconstraint inactive over the entire control zone. Hence the op-timal control strategy is theoretically attainable. To ensure thefeasible initial condition, they proposed a feasibility enforce-ment zone located upstream of the control zone in Fig.7. Theconstrained optimal control was addressed by Malikopouloset al. [55], where the constrained and unconstrained arcs werepieced together according to the activation of one or multipleconstraints. First, the time when the control (or state constraint,or both) become active was determined. Then, the constraintand non-constraint arcs needed to be pieced together. For twostate constraints and two control constraints, there are six casesfor constructing the constrained arcs.

Obtaining the analytical solution for a constrained optimalcontrol problem is computationally demanding and sometimesinfeasible, because of the many possible combinations of theactivation period of subsequent constraints [56]. Instead ofsolving the otpimal control analytically as in [55], Wang etal. [57] used a iterative process to adjust minimum safe spaceheadway of the following vehicle to ensure collision avoid-ance. Bichiou and Rakha [29] proposed an optimal intersectioncontrol system, which is designed to minimize travel timefor CAVs. The expected distance derived from the Rakha-Parsumarthy-Adjerid (RPA) car-following model was appliedas the collision avoidance constraint for each vehicle. Twoversions of the proposed framework-the optimal control timeand the optimal control effort-were tested on a roundabout,

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 9

Algorithm 2 Decentralized Optimal Control Framework forAIM (source: [54])

initialize intersection infofor vehicle j < N do . loop through FCFS queue

for k < j doassign vehicle k into one of the four pre-

determined conflict sets (i.e., Ei, Si, Li, Oi ) based onintersection movements• Ei with all the vehicles that can cause rear-end collision

at the end of the intersection• Si with vehicles on the same lane that can cause rear-

end collision at the beginning of the intersection• Li with vehicles with different origin-destination that

can cause collision within the intersection• Oi with vehicles with different origin-destination that

cannot cause collision within the intersectionretrieve the terminal time tfk and entry time tmk for

the closest preceding vehicle within each setend fordetermine the terminal time tfj from each of the

conflict setsadopt the most conservative (maximum) tfjsolve Xj(t) with boundary conditions Xj(t

oj), Xj(t

fj )

. Xj(t) defined as (3)end for

AWSC intersection, and a SIM. In spite of the significantreduction in CO2 emission, the proposed model has highcomputational expense for conducting nonlinear optimization:it takes up to five minutes to solve the optimization for a setof four vehicles. They concluded that the computational costof solving for the optimum solution makes it impractical forreal-time implementation.

B. Model Predictive Control

The model predictive control (MPC) has the advantages ofdealing with a constrained system. In MPC, the optimal controlproblem is solved in each time step over a finite time horizon,but only control for the current time step is implemented [32].Ntousakis et al. [56] integrated MPC control into a finite-horizon optimal control problem, where the possible real-time disturbance was compensated. In the VICS frameworkproposed in [12], a risk indicator of two conflict movementswas integrated into the MPC controller, along with speed errorand control input. Different from most of the studies, the actualtrajectory coordination among vehicles was implemented inthe MPC framework. But the nonlinear nature of the MPCframework and its complexity does not guarantee a globaloptimum solution. Additionally, a good guess of the initialcondition is necessary to ensure fast computation for MPC.

The optimal control framework, proposed in [29], hadintense computational demand. As a subsequent enhancement,Bichiou and Rakha [14] simplified the framework by intro-ducing MPC into the system and solved the problem numer-ically, instead of analytically. The improved algorithm canprovide real-time solutions through convex optimization once

the estimated time of arrival to the intersection is obtained.The trade-off in the slightly-reduced precision due to the con-vexification was justified as the authors concluded. Du et al.[32] formulated the corridor-level AIM where vehicle collisionwas avoided by enforcing a road segment-based referencespeed that is calculated by using a consensus algorithm ina decentralized manner. The control for each vehicle wasformulated as a tracking system, which aimed to minimizethe error between the vehicle speed and the reference speedof a particular road segment. The fast MPC method wasimplemented in [58] for the tracking system.

C. Other Vehicle Control

Lam and Katupitiya [24] adopted a proportional-derivative(PD) vehicle controller which was designed to maintain thegap of consecutive vehicles. The crossing sequence of vehicleswas determined by a wining contest that was based on theefficiency of the crossing plan submitted by each vehicle. Alane-free AIM was proposed by Li et al. [13]. Such frameworkis a re-imagination of intersection control by relaxing thepass-through paths of vehicles. This control problem aims tominimize the total clearance time and maximize the terminalpositions at each designating leg of the intersection. Due to theintensive computational requirement, the motion planning wasdecomposed into two stages. The first stage provides solutionsby solving the feasible trajectory in advance and the secondstage directs vehicles to form the standard formation (withequidistant row and columns) for executing the offline solutionobtained in stage 1.

V. EVALUATION OF AIM

Compared to other CAV applications, the evaluation of AIMis mostly conducted via computer simulation. In this section,we focus on the simulation scale, benchmark comparison,measures of effectiveness, and AIM performance.

A. Simulation Scale

Recall the coverage AIM layers shown in Table III. AIMapproached from the traffic engineering domain typicallyfocuses on the corridor coordination and trajectory planninglayers. Carlino et al. [28] evaluated the auction-based AIM infour major US cities with 30,000 drivers. This study was oftransportation planning nature, and no coordination betweenintersections was made. Other studies covered mesoscopic ormicroscopic level of traffic, and they can be further dividedinto two subgroups.

The first subgroup focused on system-optimal or near-optimal trajectory planning. No explicit vehicle dynamicsmodel was employed. The trajectory planning was typicallyformulated as a non-linear programming problem with variousobjectives, for example, minimum overall delay, minimumrisk, maximum throughput. Their evaluation scenarios dealtwith full intersection movement and with traffic demand thatis close to the real world with thousands of vehicles perhour. The second subgroup approached the isolated AIMfrom the control engineering standpoint. Its focus is realistic

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 10

vehicle dynamics model (with second- or higher-order), andthe separation of conflict movements used basic requirements(e.g., FCFS policy). Additionally, the simulation scenarios arewith much less traffic demand (as few as seven vehicles) andfewer movements (as few as two through movements).

Nonetheless, an increasing amount of research from thevehicle control domain extend their interest in replacing FCFSwith other forms of trajectory planning. Fig. 8 displays thesimulated vehicles and the number of turning movements foran intersection among the reviewed studies. The vehicle-based(VB) AIM [13, 35] is not shown, as it theoretically has aninfinite number of turning movements. Due to the natureof the intersection-based (IB) reservation (occupancy of anentire intersection), the simulated movements did not excess4 in the previous studies. The highest number of vehicles(2,000 vehicles) to be simulated for IB-AIM was conductedin [21]. The conflict point (CP)-based and tile-based (TB)reservation were often evaluated with full (12) movements andwith greater demands (upto 4,000 vehicles).

Fig. 8: Simulation scale

B. Benchmark Intersection ManagementThe composition of the comparing intersection management

in the reviewed studies is shown in Fig. 9. Fixed-time SIM(FT-SIM) is the mostly-used (46%) baseline for demonstratingthe performance of proposed AIMs, followed by FCFS-AIM(23%) and AWSC (8%).

Fig. 9: Comparing intersection managements

Insofar, the FT-SIMs used in the comparisons were un-optimized in accordance with traffic patterns, which may

potentially limit the performance of FT-SIM. As shown withdecades of practices in traffic engineering [59, 46, 60], anoptimized SPaT plan based on prevailing traffic patterns couldsignificantly improve intersection performance. The optimiza-tion relies on the balance between cycle time, phase time,queue, etc. Furthermore, a better comparison for SIM canbe found with actuated SIM [34] or adaptive SIM [17].The second most-frequent intersection management used forcomparison was FCFS-AIM, which accounted for 23% inthe previous studies. FCFS is the default priority rule ofAIM, hence FCFS-AIM was often adopted as benchmarkfor evaluating subsequent AIM variants: auction-based [28],control input-based (acceleration) [40], platoon-based [26]AIMs, etc. AWSC is the third most-used baseline, largely dueto its similar nature with AIM: 1) the priority of crossing isassigned at a vehicle level; 2) it is unsignalized; and 3) itoperates on a FCFS basis. However, AWSC is a relativelyinefficient scheme, as the required stop for each vehicle canpotentially increase the delay and queue. At median demand,SIM is recommended as per the Manual on Uniform TrafficControl Devices [61].

C. Measure of Effectiveness

The most common measures of effectiveness (MOEs) forassessing intersection performance are average delay, averagefuel consumption, average CO2 emission, travel time, evacu-ation time, intersection throughput, as shown in Table IV.

Among them, delay is the measure that relates to drivers’experience the most, as it represents the excessive amountof time in traversing an intersection. Delay can be furtherbroken down to stopped time delay, approach delay, travel timedelay, time-in-queue delay, and control delay [62]. Thoughanalytical delay prediction models (e.g., Webster’s, Akcelik,HCM2000) have been proposed along the years, simulationprovides an innovative and robust way of evaluating thedelays for intersections. Queue length provides an indication ofwhether a given intersection impedes the vehicle dischargingfrom an upstream intersection. Queue length is typically takeninto account for SIM coordination. For AIM, since vehiclecrossing is scheduled at an individual level, the queue lengthbecomes less effective in representing the overall intersectionperformance.

The number of stops is an important parameter when itcomes to emission model, since the regaining of speed forma stopped vehicle requires additional acceleration, thereforeburning more fuel. The use of fuel consumption model hasbecome more prevailing, such as the VT-Micro model [63].Other less common emissions have been adopted for studies aswell. Hydrocarbon (HC), carbon monoxide (CO), nitric oxide(NO), and fuel consumption were used by [64] for evaluatingthe proposed triangabout. The average speed and ratio ofaveraging moving time are used by [65] and [66], respectively.Safety surrogate assessment measure (SSAM) is often usedto gauge the safety performance for human-driven vehicles.However, its applicability to CAVs is still open to debate,because the performance of CAV is expected to exceeed thephysiological limitations of human drivers (e.g., much quicker

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 11

TABLE IV: Comparison of AIM Studies

Study Mobility Environmental Comfort ComputationalFramework OtherDelay Evac Time Throughput CO2 Fuel

Bichiou and Rakha [14] X X XLiu et al. [17] X X XDing et al.[15] X X X X XFajardo et al. [27] XHassan and Rakha [38] XKamal et al. [12] X XStone et al. [39] XLevin and Rey [31] XLam and Katupitiya [24] XLi et al. [13] XDu et al. [32] XJin et al. [26] X X XBashiri and Fleming [21] XBashiri et al. [33] X X XMuller et al. [25] XMirheli et al. [34] X XCarlino et al. [28] XCreemers et al. [40] XJiang et al. [30] X

response time). Evaluation time for a fixed amount of vehiclesand the minimal entry time of consecutive vehicles in thepriority queue were also used by [13] and [55], respectively.Other innovative comparing metrics had also been adapted.The required number of iterations in the proposed distributedAIM was used as a performance measure in [30]. The timerequired for stabilizing the intersection queue length to aminimal level was used as the MOE in [40].

D. AIM Performance

Fig. 10 offers an overview of the performance gain in the sixmost common MOEs. Each sample point represents a scenario-based comparison. As shown, some studies conducted multiplecomparisons, while others only did fewer. The left-hand-sidefigure shows the comparison of proposed AIMs and FT-SIMs;whereas the one on the right-hand side exhibits the comparisonbetween proposed AIMs and the FCFS-AIMs, the defaultconfiguration for AIM. Fig. 10(a) plots the MOEs that oneaims to decrease (e.g., delay, fuel consumption), and Fig. 10(b)show the MOE that one tries to increase, such as intersectionthroughput. For the papers without explicit numeration of theMOEs for all the scenarios (e.g., only figures was shown), weemployed a plot trace tool [67] to extract the numeric valuesin the original figures.

Two patterns can be observed in Fig. 10. First, the averagedelay has been adapted more frequently than any other MOE.Second, the percentage of gain or loss spans a wide range, forinstance, from 6% (increase) down to nearly -100% (decrease).The throughput of the proposed AIM in [17] was doubled(increased to 200%). Based on the reported results, the AIMperforms exceptionally well.

Besides the benefits attributed to AIM, the evaluation sce-nario (e.g., turning movements, traffic demand) likely playsa significant role in obtaining such high performance. Never-theless, there are several factors that could potentially skewperformance. The first factor is the traffic pattern, mainlythe level of saturation of the intersection. The saturationflow rate of a signalized intersection lies within the range

(a) Metric to Decrease

(b) Metric to Increase

Fig. 10: Measures of effectiveness

of 1,700-1,900 vph with variations [46]. Due to the safetyseparation among crossing vehicles, it is not hard to imagine ascenario where a vehicle has to slow down, even to a completestop at high traffic density to maintain safe separation. It isreasonable to believe that the performance of an AIM couldbe impacted under such circumstances. Second, the benchmarkintersection management scheme could play a significant rolein comparison. At median or high traffic demand, an optimizedFT-SIM can substantially outperform its non-optimized coun-terpart. Hence, the SIM should be calibrated or optimized toincrease the validity of the comparison. Additional realistictraffic scenarios for AIM are needed in order to draw astatistically sound conclusion for the applicability betweenAIM and conventional SIM.

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 12

VI. RESEARCH TRENDS AND TOPICS

Intersections are the common traffic bottlenecks in themodern transportation network. In this section, we discussfuture research directions for AIM.

1) Benchmarking: The majority of the studies served asproof-of-concept studies with two primary focuses: safe op-eration and potential benefits. Based on the reported results,the AIM performs exceptionally well (i.e., nearly 100% re-duction in average delay in certain cases) under low demandscenario without the consideration of lane change activity.Thus far, there is no consensus on a benchmark scenariofor reliably and consistently assessing the performance ofAIM. For SIM, the Highway Capacity Manual [46] providesa detailed methodology for the optimization of SPaT plansaccording to the local traffic pattern. For AIM, however, onlyfew proposed ones were subject to a wide variety of scenarios.Investigations regarding the carrying capacity of AIMs arevery much desired.

2) Semi-autonomous Intersection Management: The VolpeNational Transportation Systems Center [68] estimated thatit might take 25-30 years for CAVs to reach a 95% marketpenetration rate (MPR), even with federally mandatory instal-lation of DSRC devices on new light vehicles manufacturedin the United States. In anticipation of the transition to fullpenetration, semi-AIM which integrates non-equipped vehiclesis a logical step. So far, there are a few Semi-AIMs that havebeen proposed. The potential compromise needs to carefullybalance between compatibility and performance. The semi-AIM with signalization proposed by Dresner and Stone [69]suffered from significant degradation in performance evenwith only 5-10% presence of human operation. It was foundthat the CAVs were blocked by human-driven vehicles thatare controlled by the signalization. An enhanced version wasproposed in [39] to relax this limitation, where the controlof a vehicle can be provisionally transferred to the ADAS.Furthermore, the traffic signal also acted as a fallback strategywhen a reservation cannot be obtained. Under the semi-AIMframework, similar performance as AIM was achieved withno more than 40% MPR. However, additional research is stillrequired to quantify the possible trade-off that has to makewhen it comes to semi-AIM.

3) Priority Policy: AIM enables priority to be assignedat individual vehicle level, compared to SIM that operatesat a vehicle group movement level. Insofar, the majority ofthe proposed AIMs adopted FCFS policy, which is not likelyto lead to system optimum if solely implemented. Utilizingother network level information (e.g., routing) seems an goodcomplement to the FCFS policy for its further enhancement[45]. Besides FCFS, other non-fairness-based policies haveshown their potentials for AIM, which could be based onsystem optimality, travel mode, game theory, or custom prior-ity score system. Nevertheless, the priority policy remains anunderexplored area, and it is crucial in increasing applicabilityof AIM to a wide variety of traffic operation goals.

4) Computational Efficiency: Among the four typesof reservation systems for AIM, the vehicle-based, free-movement reservation holds the greatest amount of vehicles intheory. However, its high computational complexity hinders it

from being implemented in real-time. The practicality of thegame-theoretic reservation remains an open question accordingto the literature. The tile-based and the conflict point-basedreservations are the better candidates for real-world imple-mentation, though their computation is still nontrivial. Intensecomputation is still one of the major hurdles for AIM. Hence,the computational efficiency of the reservation system desiresmuch enhancement. The decomposition of intersection man-agement and vehicle control in several studies has exhibitedgreat potentials in improving computational efficiency.

5) Applicability with Intersection Layout: A generalizedreservation-based algorithm is needed as not all the intersec-tions in the real world follow the standard symmetrical 4-leglayout. Such variations may require necessary modificationsof the existing framework. Besides roundabout, there is anincreasing amount of alternative intersections that have beenimplemented in the real world [70]. For example, the divergingdiamond interchange (DDI) has an unconventional layoutwhere the through movements on the major directions intersecttwice (resulting in two conflict points), instead of once.Intersection-based reservation is rather counterproductive inthe case of DDI, as the interchange could span 300-meterlong. Therefore, the suitability of reservation systems to theintersection layout is worthy of investigation. In the DDIcase, the conflict point-based reservation systems are likelyto function better than the tile-based reservation system.

6) Cyber-security of CAV: More emphasises have been puton the cyber-security aspect of CAV. As a communicationplatform, CAV is susceptible to both passive and active formsof malicious attack. Passive attack has lower risk, for instance,the eavesdrop of information of a target vehicle [71]. Activeattack may include spoofing incorrect data [72], unauthorizedmessage modification, denying of service, and GPS jamming.A recent study conducted by Chen et al. [72] illustrated thatthe computational complexity should be factored in when itcomes to algorithm design for intersection management. Eventhough the I-SIG system [73] in the study is robust in theory,it is subject to exploitation in practice when only a simplifiedversion of the algorithm can be used due to computationallimitation of the RSU. Therefore, security research should alsoextend to the design of an AIM algorithm.

7) Decentralization: Centralized control is often subjectto single-point failures, making it a worthwhile target forattackers. The current information computing infrastructureexhibits the trend of decentralization, and we have seen anincreasing amount of decentralized vehicle control for AIMs.With the increased amount of computational power, CAV iscapable of computing and analyzed location traffic informationwith the onboard unit before sending actionable information toAIM. This trend on vehicle control coincides with the conceptof edge computing. Other latest technologies can evolve AIMto evolve into a decentralized, robust system further. Forinstance, blockchain technology enables the tamper resistancefor any transaction that is stored by each CAV [74, 75].

VII. CONCLUSION

Managing traffic safely and efficiently at intersections re-mains one of the most challenging problems for our trans-

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 13

portation system. The CAV technology extends the intersectionmanagement down to individual vehicle control, offering anew degree of flexibility to meet operational goals. This papersystematically reviews the state of the research of autonomousintersection management (AIM). The intersection managementis distilled into three hierarchical layers, which are corridorcoordination layer, intersection management layer, and vehiclecontrol layer. The underlying design concepts for AIM arediscussed in details. Additionally, the necessary connection toexisting signalized intersection management is also made topresent a full picture of the overall intersection management.This review shows that the reservation system with highcomputational efficiency and the extension to vehicle controllevel are two active, and yet underexplored, areas. Also, aconsensus on the evaluation scenario for AIMs is necessary toaccelerate the AIM research. Lastly, this paper highlights keyfuture research topics from an interdisciplinary standpoint.

APPENDIX

A. List of Abbreviations

TABLE A1: List of Abbreviations

Abbreviation DefinitionADAS advanced driver-assistance systemsADS automated driving systemsAIM autonomous intersection managementAV automated vehiclesAWSC all-way stop controlCV connected vehiclesCAV connected and automated vehiclesCACC-CG cooperative adaptive cruise control-cooperative

gameCVIC cooperative vehicle intersection controlCP conflict point-basedDDI diverging diamond interchangeDSRC dedicated short-range communicationFIFO first-in-first-outFCFS first-come-first-serveFT-SIM fix-time signalized intersection managementHCM Highway Capacity ManualHV human-driven vehicleIB intersection-basedI-SIG Intelligent Traffic Signal SystemMILP mixed-integer programmingMPC model predictive controlMPR market penetration rateMOE measure of effectivenessNLP non-linear programmingRSU roadside unitSICL signal-free intersection control logicSPaT signal phase and timingSIM signalized intersection managementSAE Society of Automotive Engineers InternationalTB tile-basedVB vehicle-basedVICS vehicle-intersection coordination schemeV2I vehicle-to-infrastructureV2X vehicle-to-everything

REFERENCES

[1] E.-H.Choi, “Crash factors in intersection-related crashes:An on-scene perspective,” National Highway TrafficSafety Administration, Tech. Rep. DOT HS 811 366,2010.

[2] L.Blincoe et al., “The economic and societal impact ofmotor vehicle crashes, 2010 (revised),” National High-way Transportation Safety Agency, Tech. Rep. DOT HS812 013, 2015.

[3] Q.Guo, L.Li, and X. J.Ban, “Urban traffic signal con-trol with connected and automated vehicles: A survey,”Transp. Res. Part C Emerg. Technol., 2019.

[4] J.Rios-Torres and A. A.Malikopoulos, “A survey on thecoordination of connected and automated vehicles atintersections and merging at highway on-ramps,” IEEETrans. Intell. Transp. Syst., vol. 18, no. 5, pp. 1066–1077,2017.

[5] L.Chen and C.Englund, “Cooperative intersection man-agement: A survey,” IEEE Trans. Intell. Transp. Syst.,vol. 17, no. 2, pp. 570–586, 2016.

[6] L.Li, D.Wen, and D.Yao, “A survey of traffic control withvehicular communications,” IEEE Trans. Intell. Transp.Syst., vol. 15, no. 1, pp. 425–432, 2014.

[7] F.Gross et al., “Safety effectiveness of converting sig-nalized intersections to roundabouts,” Accident Anal. &Prevention, vol. 50, pp. 234–241, 2013.

[8] N. H.Gartner et al., “A multi-band approach to arterialtraffic signal optimization,” Transp. Res. Part B: Method-ological, vol. 25, no. 1, pp. 55–74, 1991.

[9] Soceity of Automotive Engineers International, “Sur-face Vehicle Recommended Practice,” Tech. Rep.J3016 201806, 2018.

[10] Y. J.Zhang, A. A.Malikopoulos, and C. G.Cassandras,“Optimal control and coordination of connected andautomated vehicles at urban traffic intersections,” in 2016Amer. Control Conf. (ACC), 2016, pp. 6227–6232.

[11] K.Dresner and P.Stone, “A multiagent approach to au-tonomous intersection management,” J. Artif. Intell. Res.,vol. 31, pp. 591–656, 2008.

[12] M. A. S.Kamal et al., “A vehicle-intersection coordina-tion scheme for smooth flows of traffic without usingtraffic lights,” IEEE Trans. Intell. Transp. Syst., vol. 16,no. 3, pp. 1136–1147, 2015.

[13] B.Li et al., “Near-optimal online motion planning ofconnected and automated vehicles at a signal-free andlane-free intersection,” IEEE Intell. Vehicles Symp., vol.2018-June, no. IV, pp. 1432–1437, 2018.

[14] Y.Bichiou and H. A.Rakha, “Real-time optimalintersection control system for automated/cooperativevehicles,” Int. J. Transp. Sci. & Technol., pp. 1–12,2018. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S2046043018300315

[15] J.Ding et al., “Centralized cooperative intersection con-trol under automated vehicle environment,” no. iv, pp.972–977, 2017.

[16] E. R.Muller, B.Wahlberg, and R. C.Carlson, “Optimalmotion planning for automated vehicles with scheduledarrivals at intersections,” in 2018 European Control Conf.(ECC), 2018, pp. 1672–1678.

[17] B.Liu et al., “Trajectory planning for autonomous in-tersection management of connected vehicles,” Simul.Model. Pract. & Theory, vol. 90, pp. 16–30, 2019.

[18] I. H.Zohdy and H. A.Rakha, “Intersection management

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 14

via vehicle connectivity: the intersection cooperativeadaptive cruise control system concept,” J. Intell. Transp.Syst., vol. 20, no. 1, pp. 17–32, 2016.

[19] I. H.Zohdy and H.Rakha, “Game theory algorithm forintersection-based cooperative adaptive cruise control(CACC) systems,” in 15th IEEE Int. Conf. Intell. Transp.Syst. (ITSC). IEEE, 2012, pp. 1097–1102.

[20] A. I. M.Medina, N.van deWouw, and H.Nijmeijer, “Co-operative intersection control based on virtual platoon-ing,” IEEE Trans. Intell. Transp. Syst., vol. 19, no. 6, pp.1727–1740, 2018.

[21] M.Bashiri and C. H.Fleming, “A platoon-based intersec-tion management system for autonomous vehicles,” in2017 IEEE Intell. Veh. Symp. (IV), no. iv, 2017, pp. 667–672.

[22] S. A.Fayazi, A.Vahidi, and A.Luckow, “Optimal schedul-ing of autonomous vehicle arrivals at intelligent intersec-tions via MILP,” in 2017 Amer. Control Conf. (ACC),2017, pp. 4920–4925.

[23] J.Lee, B.Park, and I.Yun, “Cumulative Travel-Time Re-sponsive Real-Time Intersection Control Algorithm in theConnected Vehicle Environment,” J. Transp. Eng., vol.139, no. 10, pp. 1020–1029, 2013.

[24] S.Lam and J.Katupitiya, “Cooperative intersection nego-tiation for multiple autonomous platoons,” IFAC Proc.Vol., vol. 46, no. 10, pp. 48–53, 2013.

[25] E. R.Muller, R. C.Carlson, and W. K.Junior, “Intersec-tion control for automated vehicles with MILP,” IFAC-PapersOnLine, vol. 49, no. 3, pp. 37–42, 2016.

[26] Q.Jin et al., “Platoon-based multi-agent intersection man-agement for connected vehicle,” in 16th IEEE Int. Conf.Intell. Transp. Syst. (ITSC), 2013, pp. 1462–1467.

[27] D.Fajardo et al., “Automated Intersection ControlPerformance of Future Innovation versus CurrentTraffic Signal Control,” Transp. Res. Record, vol.2259, pp. 223–232, 2011. [Online]. Available: http://trrjournalonline.trb.org/doi/10.3141/2259-21

[28] D.Carlino, S. D.Boyles, and P.Stone, “Auction-basedautonomous intersection management,” in 16th IEEE Int.Conf. Intell. Transp. Syst. (ITSC), 2013, pp. 529–534.

[29] Y.Bichiou and H. A.Rakha, “Developing an OptimalIntersection Control System for Automated ConnectedVehicles,” IEEE Trans. Intell. Transp. Syst., 2018.

[30] Y.Jiang et al., “Distributed algorithm for optimalvehicle coordination at traffic intersections,” IFAC-PapersOnLine, vol. 50, no. 1, pp. 11 577–11 582, 2017.

[31] M. W.Levin and D.Rey, “Conflict-point formulation ofintersection control for autonomous vehicles,” Transp.Res. Part C Emerg. Technol., vol. 85, pp. 528–547, 2017.

[32] Z.Du, B.HomChaudhuri, and P.Pisu, “Hierarchical dis-tributed coordination strategy of connected and au-tomated vehicles at multiple intersections,” J. Intell.Transp. Syst., vol. 22, no. 2, pp. 144–158, 2018.

[33] M.Bashiri, H.Jafarzadeh, and C. H.Fleming, “PAIM:platoon-based autonomous intersection management,” in21st IEEE Int. Conf. Intell. Transp. Syst. (ITSC), 2018,pp. 374–380.

[34] A.Mirheli, L.Hajibabai, and A.Hajbabaie, “Development

of a signal-head-free intersection control logic in afully connected and autonomous vehicle environment,”Transp. Res. Part C Emerg. Technol., vol. 92, pp. 412–425, 2018.

[35] B.Li and Y.Zhang, “Fault-tolerant cooperative motionplanning of connected and automated vehicles at a signal-free and lane-free intersection,” IFAC-PapersOnLine,vol. 51, no. 24, pp. 60–67, 2018.

[36] C.Wuthishuwong, A.Traechtler, and T.Bruns, “Safe tra-jectory planning for autonomous intersection manage-ment by using vehicle to infrastructure communication,”EURASIP J. Wireless Commun. & Netw., vol. 2015, no. 1,p. 33, 2015.

[37] X.Zhao et al., “Multi-objective cooperative schedulingof CAVs at non-signalized intersection,” pp. 3314–3319,2018.

[38] A. A.Hassan and H. A.Rakha, “A fully-distributed heuris-tic algorithm for control of autonomous vehicle move-ments at isolated intersections,” Int. J. Transp. Sci. &Technol., vol. 3, no. 4, pp. 297–309, 2014.

[39] P.Stone, S.Zhang, and T.-C.Au, “Autonomous intersec-tion management for semi-autonomous vehicles,” inRoutledge Handbook of Transportation. Routledge,2015, pp. 116–132.

[40] F.Creemers et al., “Design of a supervisory controller forcooperative intersection control using model predictivecontrol,” IFAC-PapersOnLine, vol. 51, no. 33, pp. 74–79, 2018.

[41] Y.Wu, H.Chen, and F.Zhu, “DCL-AIM: Decentralizedcoordination learning of autonomous intersection man-agement for connected and automated vehicles,” Transp.Res. Part C Emerg. Technol., vol. 103, pp. 246–260,2019.

[42] K.Dresner and P.Stone, “Human-usable and emergencyvehicle-aware control policies for autonomous intersec-tion management,” in 4th Int. Workshop on Agents inTraffic & Transp. (ATT), 2006.

[43] M.Elhenawy et al., “An intersection game-theory-basedtraffic control algorithm in a connected vehicle environ-ment,” in 18th IEEE Int. Conf. Intell. Transp. Syst., 2015,pp. 343–347.

[44] Q.He, K. L.Head, and J.Ding, “PAMSCOD: Platoon-based arterial multi-modal signal control with onlinedata,” Transp. Res. Part C Emerg. Technol., vol. 20, no. 1,pp. 164–184, 2012.

[45] M. W.Levin, S. D.Boyles, and R.Patel, “Paradoxes ofreservation-based intersection controls in traffic net-works,” Transp. Res. Part A: Policy & Pract., vol. 90,pp. 14–25, 2016.

[46] Highway Capacity Manual, 6th ed. TransportationResearch Board, 2016.

[47] Y.Zhang and C. G.Cassandras, “The penetration effect ofconnected automated vehicles in urban traffic: an energyimpact study,” in 2018 IEEE Conference on ControlTechnology and Applications (CCTA). IEEE, 2018, pp.620–625.

[48] A.Mirheli et al., “A consensus-based distributed trajec-tory control in a signal-free intersection,” Transp. Res.

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 15

Part C Emerg. Technol., vol. 100, pp. 161–176, 2019.[49] F.Zhu and S. V.Ukkusuri, “A linear programming for-

mulation for autonomous intersection control within adynamic traffic assignment and connected vehicle envi-ronment,” Transp. Res. Part C Emerg. Technol., vol. 55,pp. 363–378, 2015.

[50] Y.Zhang, C. G.Cassandras, and A. A.Malikopoulos, “Op-timal control of connected automated vehicles at urbantraffic intersections: A feasibility enforcement analysis,”in 2017 Amer. Control Conf. (ACC). IEEE, 2017, pp.3548–3553.

[51] D.Tonon, M. S.Aronna, and D.Kalise, Optimal control:Novel directions and applications. Springer, 2017, vol.2180.

[52] N.Yao and F.Zhang, “Resolving contentions for intelli-gent traffic intersections using optimal priority assign-ment and model predictive control,” in 2018 IEEE Conf.Control Technol. & Appl. (CCTA), 2018.

[53] D. S.Naidu, Optimal control systems. CRC press, 2002.[54] Y.Zhang, A. A.Malikopoulos, and C. G.Cassandras, “De-

centralized optimal control for connected automated ve-hicles at intersections including left and right turns,” inIEEE 56th Annu. Conf. Decis. & Control (CDC), 2017,pp. 4428–4433.

[55] A. A.Malikopoulos, C. G.Cassandras, and Y. J.Zhang, “Adecentralized energy-optimal control framework for con-nected automated vehicles at signal-free intersections,”Automatica, vol. 93, pp. 244–256, 2018.

[56] I. A.Ntousakis, I. K.Nikolos, and M.Papageorgiou, “Op-timal vehicle trajectory planning in the context of cooper-ative merging on highways,” Transp. Res. Part C Emerg.Technol., vol. 71, pp. 464–488, 2016.

[57] Y.Wang, P.Cai, and G.Lu, “Cooperative autonomous traf-fic organization method for connected automated vehiclesin multi-intersection road networks,” Transp. Res. Part CEmerg. Technol., vol. 111, pp. 458–476, 2020.

[58] B.HomChaudhuri, A.Vahidi, and P.Pisu, “Fast modelpredictive control-based fuel efficient control strategy fora group of connected vehicles in urban road conditions,”IEEE Trans. Control Syst. Technol., vol. 25, no. 2, pp.760–767, 2017.

[59] P.Koonce and L.Rodegerdts, “Traffic signal timingmanual,” Federal Highway Administration, Tech. Rep.FHWA-HOP-08-024, 2008.

[60] T.Urbanik et al., Signal Timing Manual. Transp. Res.Board, 2015.

[61] Manual on Uniform Traffic Control Devices for Streetsand Highways. Federal Highway Administration, 2009.

[62] T. V.Mathew, “Transportation systems engineering,” CellTransmiss. Model., IIT Bombay, 2014.

[63] H.Rakha, K.Ahn, and A.Trani, “Development of VT-micro model for estimating hot stabilized light duty ve-hicle and truck emissions,” Transp. Res. Part D: Transp.& Environ, vol. 9, no. 1, pp. 49–74, 2004.

[64] C.-S.Chou and A. P.Nichols, “Evaluation of triangaboutas alternative for intersection with nonthrough arterialmovement,” Transp. Res. Rec., vol. 2404, no. 1, pp. 38–48, 2014.

[65] T. E.Hildebrand, “Unconventional intersection designsfor improving through traffic along the arterial road,”Master’s thesis, Dept. Civil & Environ. Eng, Florida StateUniversity, Tallahassee, FL, USA, 2007.

[66] J. D.Reid and J. E.Hummer, “Travel time comparisonsbetween seven unconventional arterial intersection de-signs,” Transp. Res. Rec., vol. 1751, no. 1, pp. 56–66,2001.

[67] A.Rohatgi, “Webplotdigitizer,” 2020. [Online]. Available:https://github.com/ankitrohatgi/WebPlotDigitizer

[68] Volpe National Transportation Systems Center, “Vehicle-infrastructure integration (VII) initiative benefit-costanalysis version 2.3 (draft),” 2008.

[69] K. M.Dresner and P.Stone, “Sharing the road: Au-tonomous vehicles meet human drivers.” in Int. JointConf. Artif. Intell. (IJCAI 2007), vol. 7, 2007, pp. 1263–1268.

[70] Z.Zhong et al., “Unconventional arterial intersection de-signs under connected and automated vehicle environ-ment: A survey,” arXiv preprint arXiv:1811.03074, 2018.

[71] D.Elliott, W.Keen, and L.Miao, “Recent advances inconnected and automated vehicles,” J. Traffic & Transp.Eng. (English Edition), 2019.

[72] Q. A.Chen et al., “Exposing congestion attack on emerg-ing connected vehicle based traffic signal control,” inNetw. & Distrib. Syst. Secur. (NDSS) Symp., 2018.

[73] K.Ahn et al., “Multimodal intelligent traffic signal sys-tem simulation model development and assessment,”Transp. Res. Rec., vol. 2558, no. 1, pp. 92–102, 2016.

[74] B.Leiding, P.Memarmoshrefi, and D.Hogrefe, “Self-managed and blockchain-based vehicular ad-hoc net-works,” in 2016 ACM Int Joint Conf. Pervasive andUbiquitous Computing: Adjunct. ACM, 2016, pp. 137–140.

[75] W.Li, M.Nejad, and R.Zhang, “A blockchain-based archi-tecture for traffic signal control systems,” in 2019 IEEEInt. Congr. Internet Things (ICIOT), 2019, pp. 33–40.

DOI:10.1109/MITS.2020.3014074 (IN PRESS) 16

Zijia (Gary) Zhong is a Postdoctoral Researcherin the Center of Integrated Mobility Sciences atthe National Renewable Energy Laboratory, UnitedStates. He received his Ph.D degree in Transporta-tion Engineering and Master’s degrees in Civil Engi-neering from the New Jersey Institute of Technologyin 2018 and 2011 respectively. His research inter-ests include deployment of intelligent transportationsystems (ITS), emerging mobility, high-performancecomputing for transportation modeling, data ana-lytics, vehicle platooning, highway automation, and

human factor study for ADAS.

Mark Nejad is an Assistant Professor in the De-partment of Civil and Environmental Engineering atthe University of Delaware. His research interestsinclude connected and automated vehicles, networkoptimization and control, and game theory. He haspublished more than thirty peer-reviewed papersin venues such as Transportation Science, IEEETransactions on Parallel and Distributed Systems,and IEEE Transactions on Computers. He receivedseveral publication awards including the 2016 In-stitute of Industrial and Systems Engineers (IISE)

Pritzker Best Doctoral Dissertation Award and the INFORMS ENRE beststudent paper award. He is a member of the IEEE and INFORMS.

Earl (Rusty) E. Lee, II received his Bachelorsdegree in Nuclear Engineering, Masters degree inManagement, and Ph.D. degree in Decision Sciencesand Engineering Systems from Rensselaer Polytech-nic Institute in 1978, 2004, and 2006, respectively.Currently, he is an Assistant Professor in the Depart-ment of Civil and Environmental Engineering at theUniversity of Delaware. He is also the Director ofthe Delaware Technology Transfer (T2) Center and aCore Faculty member in the Disaster Research Cen-ter. His research interests include disaster manage-

ment, infrastructure system modeling, and transportation system operations.