Towards a robust and efficient traffic junction management

Abstract

As autonomous vehicles (AVs) are becoming more advanced, the future where all vehicles are connected and autonomous vehicles (CAVs) is highly likely to be a reality. To realise their potential, existing literature has proposed approaches for non-signalised or autonomous junction management to improve the traffic flow efficiency at junctions by utilising the communication capability of CAVs. However, existing methods entail important limitations arising from several simplifying assumptions that are often made. For example, most methods are centralised, which makes the control more straightforward but introduces a computational bottleneck issue, raising concerns about scalability and latency. Also, dynamic features, such as obstructions that could lead to a deadlock, are typically not considered. Moreover, several approaches adopt platooning to improve traffic efficiency, whereby vehicles maintain a close gap between each other and cross the junction as a group. This is usually operated in a restricted and static manner (e.g. using pre-generated and fixed-sized platoons), which can be inefficient. Additionally, only a few approaches have considered key practical constraints, such as the presence of pedestrians at the junction in urban areas, which introduces considerable complexity due to their shared road usage with vehicles. Furthermore, while all existing approaches have been validated at the individual junction level, only a few have tested their performance and impacts at the corridor or network level. It is imperative that these challenges are addressed if CAV-enable signal-less traffic management at junctions in urban areas is to be implemented in practice. The present thesis addresses the challenges identified in a number of ways. Firstly, a novel computationally decentralised signal-less traffic management approach has been introduced and formulated as a multi-agent system consisting of a manager agent and driver agents. Specifically, the main reliance of the manager agent is alleviated by transferring most of the computation to the driver agents, thereby addressing the bottleneck issue and improving scalability & latency. By having minimal information provided by the manager agent, the driver agents can perform local calculations which are the prediction of crossing paths and resolving conflicts between each other. With similar settings from the state-of-the-art method, i.e. a 3-lane-4-wayjunction, our method can address the challenge of the bottleneck at the manager agent and reduce the manager's computation burden, the number of exchanged messages between the manager and driver agents required to perform automation control, and as well as enabling parallel system operation. Secondly, with possible road obstruction, e.g., a construction site, delaying traffic flow at the junction, we propose a multi-vehicle collision avoidance approach that guarantees vehicles safe crossing and alleviates delays. In particular, vehicles intelligently calculate a path that safely avoids obstruction while using the least possible space. A microscopic traffic simulation, Simulation of Urban MObility (SUMO), is used to model an environment after a practical junction in Manhattan, producing two obstruction scenarios: obstacle at the entry and junction area. By addressing this, our method is more robust to possible obstructions located in the junction vicinity. To further improve traffic efficiency, we introduce the use of platooning at the signal-less junction, utilising the close gap between vehicles and group crossing that majorly increases traffic capacity and reduces delays due to fewer stop-and-go movements. Specifically, we propose an agent-based dynamic platoon formation mechanism, where the manager agent calculates the benefits of forming each platoon in terms of waiting time to determine the optimal platoons' size dynamically. This is to speed up platoon members' crossing movements while minimising the delays of several vehicles waiting for such lengthy vehicles, i.e. platoons, to cross the junction. More importantly, the group's leaders are responsible for members' path prediction and reservation requests, thereby reducing computation load and the number of exchanged messages with the manager agent. Moreover, to realise the performance of platooning in a more realistic environment, we expand our study to the network level, covering multiple junctions. A real-world case study network from Athens, Greece, is considered, comprising real vehicle movement data from an extensive drone dataset. We calibrate the simulation after such dataset, reproducing the ground-truth traffic demand with the practical corridor geometry and heterogeneous vehicle types, e.g. buses, taxis, and motorcycles, to ensure the realism of the environment as much as possible. Additionally, the randomness of vehicle generation is also introduced in order to reduce the wave-like bias from the dataset due to the use of conventional traffic lights. In this way, our platooning method can be evaluated extensively and realistically. Lastly, we address the crucial aspect of pedestrian considerations in autonomous junctions. This is to anticipate the future usage of autonomous junctions in urban areas where pedestrians are a crucial system element, especially from a safety and complexity viewpoint. To this end, we introduce a waiting-time-driven approach that dynamically switches between different operational phases at the junction, including pedestrian and freely automated phases. By taking advantage of the uneven traffic flow on different inbound roads, we can maximise traffic throughput while balancing pedestrian waiting time. These operational phases can be switched dynamically to accommodate varying traffic conditions taking the vehicles and pedestrian waiting time into account, ensuring that the system can adapt to changing circumstances and optimise performance in real time. In conclusion, our decentralised junction management model can still maintain a similar performance to the state-of-the-art approach. The results show that with single junction scenarios, our approach can reduce the number of exchanged messages by up to ≈40%. For obstruction avoidance, the simulation results show that whenever obstructions exist in the junction area and at the entry our model can maintain the throughput up to 94% and 99% respectively, compared to the no-obstruction baseline. Moreover, with our dynamic platoon formation, the evaluation with a single junction shows that the throughput can be increased by up to ≈12%, and the average travel time can be shortened by up to ≈31% compared to a non-platoon-based state of the art. Furthermore, with a highly realistic corridor environment, the simulations with light and heavy traffic scenarios how that our platooning can reduce the trip duration by up to ≈22% and ≈45%compared to conventional traffic lights and state-of-the-art approach. Additionally, with pedestrians in the system circulation, at the network level with light traffic volumes, our approach can reduce the vehicle trip duration by up to ≈12% and ≈21% with and without platooning compared to traffic light controls. However, the out-performance disappears in heavy traffic scenarios. 

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata