A Next-Generation Intersection Control Algorithm for Autonomous Vehicles

1 A reservation-based autonomous intersection control system, named Autonomous Control of 2 Urban TrAffic (ACUTA) is presented in this paper. ACUTA manages autonomous vehicles in 3 the vicinity of an intersection to allow them to pass the intersection without any conflict and few 4 stops. To address the operational issues identified in previous studies on reservation-based 5 autonomous intersection management, three operational enhancement strategies were introduced 6 and incorporated in ACUTA. Along with operational enhancements offered by ACUTA, its 7 implementation in the standard simulation platform VISSIM is significant. The enhancement 8 strategies were evaluated and shown to be effective in reducing intersection delay. ACUTA was 9 modeled as a single-tile and a multi-tile system and simulation experiments were conducted in 10 VISSIM to evaluate operational performance of both. Performance of single and multi-tile 11 ACUTA was compared with operational performance of an optimized signalized intersection, 12 and a four-way stop intersection. Evaluation results demonstrated that compared with the 13 optimized signal control, Multi-Tile ACUTA increased left turn, right turn and through 14 capacities by 37%, 32%, and 31%, respectively. As a result, the Multi-Tile ACUTA intersection 15 caused considerably less delay than the optimized signalized intersection. Single-Tile ACUTA 16 also resulted in significantly less delay than four-way stop control, when the approach traffic 17 demand was less than 300 veh/hr. Finally, sensitivity analyses were conducted on ACUTA’s 18 configurable parameters, identifying the parameters that the intersection delay is sensitive to, 19 along with their trends in impacting intersection delay. Results of the sensitivity analyses can be 20 used to optimize the operational performance of ACUTA in future research. 21 TRB 2013 Annual Meeting Paper revised from original submittal. Li, Chitturi, Zheng, Bill, and Noyce 2 INTRODUCTION 22 Traffic congestion is a global issue with increasing traffic demand every year. Federal Highway 23 Administration (FHWA) estimates that by 2020, 29% of urban National Highway System (NHS) 24 routes will be congested for much of the day, and 42 percent of NHS routes will be congested 25 during peak periods (1). A key solution to alleviate future traffic congestion lies in better 26 management of the existing network to process traffic more efficiently. One of the key 27 bottlenecks in the transportation system is the signalized intersection. 28 The application of autonomous vehicles makes it possible to eliminate traditional traffic 29 signals from the intersection, and hence has the potential to maximize intersection capacity, 30 significantly enhancing intersection mobility. From a safety perspective, considering that 90% of 31 road crashes are attributed to driver errors (2), use of autonomous vehicles, is potentially 32 effective in reducing intersection related crashes. Therefore, autonomous vehicles (vehicles 33 without human intervention) offer an unprecedented opportunity to address the twin issues of 34 traffic operations and safety dodging the society today. Autonomous vehicles are under 35 development by many automotive manufacturers and their wide usage on highway systems is 36 expected to become reality in the near future. Although potential benefits are expected, how to 37 take full advantage of autonomous vehicles, and maximize operational performance of 38 autonomous vehicles at intersections is not fully understood. 39 Previous studies have investigated both centralized and decentralized strategies for 40 managing autonomous vehicles at intersections (3-17, 19-20). In fact, the research on the 41 autonomous vehicles can be dated back to 1990s (21-24). An evaluation study indicated that 42 among all possible solutions to autonomous intersection control, the reservation-based 43 centralized control had the best performance in terms of maximizing the intersection capacity 44 and reducing the delay (17). The mechanism of the reservation-based system is introduced in the 45 following section of Background and Literature. Another study found that starvation issues may 46 occur in the reservation-based system when traffic demands on the mainline and side road were 47 unbalanced (8). Starvation here reflects the scenario that approaching vehicles on the side street 48 cannot get reservations and form a queue at the entrance of the intersection. 49 According to a different comparison research, the reservation-based system was 50 outperformed by the traffic signal when the traffic demand was higher than a certain threshold 51 and indicated a further investigation on the robustness of reservation-based system is needed 52 (20). All these facts indicate that issues still exist in the reservation-based system although it has 53 potential to maximize intersection capacity among all possible solutions. It has to be noted that 54 none of the exiting studies on autonomous intersection control used standard commercial 55 microscopic simulation software, such as VISSIM or CORSIM. Customized simulation tools 56 were used in those studies, which cause that the results from different studies can not be 57 comparable to each other due to the ununiformed simulation platform. 58 Therefore, the objective of this research is three-fold: (1) develop an enhanced 59 reservation-based autonomous intersection control algorithm, named as Autonomous Control of 60 Urban TrAffic (ACUTA), with potential enhancements that address existing operational issues 61 and make the system more realistic; (2) develop a VISSIM-based simulation platform to evaluate 62 ACUTA; and (3) compare ACUTA with 4-way stop control and signal control, as well as 63 conduct sensitivity analysis to investigate avenues to maximize the performance of ACUTA. 64 65 TRB 2013 Annual Meeting Paper revised from original submittal. Li, Chitturi, Zheng, Bill, and Noyce 3 BACKGROUND AND LITERATURE REVIEW 66 Many researchers have explored ideas and algorithms for effective management of autonomous 67 vehicles at intersections. Both centralized and decentralized control strategies were investigated 68 in previous studies. 69 Centralized control features an intersection controller that regulates the entire 70 intersection. Vehicles only communicate with the central controller to get passing instructions. 71 Dresnser and Stone were the first to introduce a reservation-based multi-agent system, named as 72 Autonomous Intersection Management (AIM) (3). In reservation-based system, intersection is 73 divided into a grid of n by n tiles. When a vehicle approaches an intersection, the driver agent 74 that represents the vehicle communicates with the intersection manager. Basic mechanism of 75 AIM is that driver agent sends requests to intersection manager to reserve the intersection for 76 certain time-spaces needed for traversing the intersection based on vehicle’s estimated arrival 77 and departure time. Intersection manager checks what and how much resource (tiles) will be 78 occupied by arequesting vehicle, and identifies whether these requested tiles have already been 79 reserved by other vehicles. If the tiles are already reserved,, the request will be rejected. 80 Otherwise a reservation will be made. Vehicle agent is notified by intersection manager whether 81 the request is approved or rejected. The instruction of travel will be sent to vehicle agent by 82 intersection manager with approval notice. 83 In the prototype version of Dresner and Stone’s system, left and right turns were not 84 allowed and all vehicles traveled at the same speed (3). Dresner and Stone validated their 85 algorithm using a simulation that they developed, in which they defined certain lane-change and 86 car following behaviors, signal and stop control operations for comparison purpose, and methods 87 for estimating throughput volume and delay. The second version of their system was much more 88 comprehensive by allowing turns and acceleration in the intersection (4, 5). The improved 89 system was evaluated in their own simulation environment with comparison to stop-control and 90 signal-control scenarios. The impact of restricting left and right turns being made from 91 designated lanes rather than from any lanes was also analyzed. Theoretically, in a reservation92 based system, the restriction was not necessary. Relaxing the restriction was supposed to provide 93 more flexibility to drivers. However, results showed that restricted turn conditions resulted in 94 lower delay than allowing turns from any lane. Dresner and Stone further stated that the results 95 might be misleading, because the delay incurred by vehicles from lane change maneuvers can 96 cause longer delay (6). 97 In later versions of AIM, safety issues were addressed by adding a safety net in the 98 system (7). Batch processing of reservation requests were also realized to address the starvation 99 issue due to unbalanced traffic demands on mainline and side road (8, 9). AIM was finally tested 100 in a mixed reality platform (10). Most of Stone’s studies resulted in an exceptionally low delay 101 (< 5 s/veh) at even extremely high traffic demand (i.e. 2100 veh/hr/ln) which even exceeds the 102 typical saturation flow rate (10). All these results indicate their algorithm performed very well 103 under high demand. However, these results were obtained using their own simulation tool, rather 104 than standard commercial simulation packages like VISSIM or CORSIM. 105 In addition to Stone et al., centralized control system was also investigated by researchers 106 from France. Wu et al. (12) and Yan et al. (13) studied a theoretical approach to control 107 autonomous vehicles at an isolated intersection through V2I communications. In their system, 108 the intersection has only two directions. Yan et al. (14) improved the system by generalizing the 109 intersection into a common four-way intersection. Approaching vehicles inform the intersection 110 controller of their position and routing information. The intersection controller decides the 111 TRB 2013 Annual Meeting Paper revised from original submittal. Li, Chitturi, Zheng, Bill, and Noyce 4 passing sequence of vehicles. The decision by the controller was optimized. The objective of the 112 optimization was to minimize total time of clearing all autonomous vehicles at the intersection. 113 The key point was to decide an optimal vehicle passing sequence. A dynamic programming 114 algorithm was used to solve this problem. Vehicle passing sequence could dynamically change 115 when new vehicles enter the control range. No simulation or validation was performed in their 116 research. 117 Wu et al. compared both of their centralized control strategies based on dynamic 118 programing and their negotiation-based decentralized control strategy to an adaptive traffic 119 controller and reservation-based traffic system developed by Dresner and Stone (3) in terms of 120 operational performance (19). Results indicated that the reservation-based system performed best 121 while their centralized and decentralized systems had similar operational performance. They 122 concluded that despite the fact that reservation-based system maximizes use of space of the 123 intersection, it lacks considerations of safe distance between two vehicles in both non-conflicting 124 and conflicting movements. 125 Vasirani and Ossowski evaluated reservation-based system and compared it to signal 126 control system (20). They found that reservation-based system only outperformed traffic signal 127 when traffic demand is below a certain threshold of about 555 veh/hr/ln. Reservation-based 128 approach performed worse than traffic signal when traffic volume was higher than a certain 129 threshold. They concluded that this was because a reservation-based intersection is less robust 130 than a signal-controlled intersection and performance is very sensitive to traffic demand. 131 In summary, centralized control can achieve better efficiency by maximizing the use of 132 all available resources, and is more reliable and safer. However, it will also cost more to deploy 133 in the field. Decentralized control has lower cost to implement when compared with centralized 134 control. Therefore, centralized control is more suitable for urban intersections with heavy traffic, 135 while the decentralized control works better for rural intersections with light traffic. Among all 136 centralized control strategies, reservation-based system is the simplest one with the highest 137 efficiency, although it has some potential issues like starvation and lower performance under 138 high traffic demand. 139 140 THE ENHANCED RESERVATION-BASED ALGORITHM 141 Working Mechanism of ACUTA 142 Considering the superiority of reservation-based system in terms of maximizing intersection 143 capacity, the next-generation intersection control system developed in this project was based on 144 First-Come-First-Serve (FCFS) reservation-based protocol (2), with enhancements to improve 145 some operational issues identified in previous studies (2, 9). The system was named Autonomous 146 Control of Urban TrAffic (ACUTA). Note that ACUTA only applies to the condition that 100% 147 of the vehicles on the road are autonomous vehicles. 148 ACUTA utilizes a centralized control strategy for managing fully-autonomous vehicles at 149 an intersection. All vehicles in ACUTA are autonomous and communicate only to an intersection 150 controller, namely, intersection manager (IM). An IM regulates the intersection by determining 151 the passing sequence of all approaching vehicles. Specifically, intersection is divided into a mesh 152 of n by n tiles, as shown in Figure 1, where “n” is termed as granularity, which is tile density of 153 the intersection mesh. 154 TRB 2013 Annual Meeting Paper revised from original submittal. Li, Chitturi, Zheng, Bill, and Noyce 5