李怡乐

入学时间：2018级

答辩时间：2021年

论文题目：城市道路交叉口网联自动驾驶车辆通行权分配方法研究

中文摘要

摘要

随着自动驾驶车辆、智能网联车辆的出现，道路交通系统将会发生较大的变化，构成由传统人类驾驶车辆、智能网联车辆、网联自动驾驶车辆、行人、非机动车等组成的新型混合交通。在新型混合交通场景下，道路交叉口通行权分配问题亟须重构与深入研究。

本文作为探索研究的第一步，主要针对网联自动驾驶车辆通过城市道路交叉口的通行权分配方法展开研究。要解决的问题是，城市道路交叉口网联自动驾驶车辆通行权分配方法研究。沿用前辈对通行权分配的研究与定义，本文在信号控制交叉口、无信号交叉口两个场景下，分别研究了网联自动驾驶车辆的通行权分配。

在信号控制交叉口，网联自动驾驶车辆的通行权分配，实质上是信号控制器根据交叉口内交通流的状态信息，考虑延误及停车率等因素，以最大化交通效益为目标，对交通参与主体分配通行次序。已有的研究有固定配时、感应控制等方法，针对交通流的宏观特征进行信号配时，存在灵活性不足的问题。强化学习用于解决不确定性顺序决策问题，在模型训练好后可以直接调用参数进行实时预测，因此本文用强化学习的方法建模来弥补不足。研究结果表明，可以用强化学习的方法解决网联自动驾驶车辆在城市道路信号控制交叉口的通行权分配问题，并且这样的方法具有更好的灵活性。

在无信号交叉口，网联自动驾驶车辆的通行权分配，已有的研究主要聚焦于集中式分配方法，由于计算复杂度高，存在难以扩展到车辆数量较多的情况的不足。本文的创新点在于研究分布式的通行权分配方法。亦即将车辆在交叉口内部通行时的决策权分发到每辆车自身上，车辆与车辆之间是相对独立的。由于多智能体系统里智能体之间是相互独立的，因此引入多智能体的概念。由于强化学习的决策方法在分布式分配通行权方面有优势，因此本文用强化学习的方法构建了网联自动驾驶车辆在无信号交叉口的通行权分配模型。在分布式分配通行权时需要车辆独立考虑自身与周围车辆的关系来满足安全约束，而图网络可以实现这一目标，因此引入图建模方法，建立了车辆之间的互动关联模型，从而使得分布式的通行权分配更具合理性。研究结果表明，本文提出的通行权分配方法具有可行性。

最后，作为拓展研究，本文研究了一种在普适环境下，网联自动驾驶车辆通行权分配的可能性。研究灵感起源于不同交叉口的设计规则的差异性，为了使在一个交叉口学到的通行权分配方法，可扩展到其他的交叉口，因此提出考虑环境动力学特征的模仿学习方法用于解决网联自动驾驶车辆通行权分配问题。研究结果表明，这种方法为解决通行权分配问题提供了新的可能。

综上所述，本文主要研究了网联自动驾驶车辆在信号控制交叉口和无信号交叉口两个场景下的通行权分配方法，并扩展研究了在同种类型但几何设计不同的交叉口的通行权策略的可迁移性问题。研究结果表明，可以使用人工智能技术进行网联自动驾驶车辆通行权分配。

关键词：通行权分配，强化学习，信号控制，多智能体，图神经网络

英文摘要

ABSTRACT

With the emergence of autonomous vehicles and intelligent connected vehicles, the participants in the urban transportation system will undergo major changes, consisting of traditional human-driven vehicles, intelligent connected vehicles, connected autonomous vehicles, pedestrians, and non-motorized vehicles. In the scenario with new mixed traffic, the issue of control method is facing new opportunities and challenges.

The problem to be solved in this paper is to study the control method of connected autonomous vehicles at an urban road intersection. Following the research of the predecessors, this paper studies the control method of the connected autonomous vehicles in the two scenarios which are signal-controlled intersections and unsignalized intersections.

At a signal-controlled intersection, the control method of connected autonomous vehicles is to assign the traffic order of the traffic participants based on the status information of the traffic flow in the intersection. Existing research methods such as fixed timing and adaptive signal control have the problem of insufficient flexibility. This paper uses reinforcement learning method to make up for the shortcomings. The research results show that the method of reinforcement learning can be used to solve the problem of the control method of connected autonomous vehicles at a signal-controlled intersection on urban roads, and this method has better flexibility.

At a signal-free intersection, the control method of connected autonomous vehicles mainly focuses on the centralized distribution method. However, due to the high computational complexity, it is difficult to expand to the situation of a large number of vehicles. The innovative point of this article is to study the distributed control method of connected autonomous vehicles at an intersection and vehicles pass through the intersection by themselves. Therefore, the concept of multi-agent is introduced, and each vehicle is regarded as an independent agent. Because the prediction method of reinforcement learning has advantages in distributed modeling, this paper uses reinforcement learning to model the problem of control method of connected autonomous vehicles at an unsignalized intersection. In the distributed control method, the vehicle needs to independently consider the relationship between itself and the surrounding vehicles to meet the safety constraints. The graph technology in artificial intelligence can meet this requirement. Therefore, the interactive association between the vehicles is introduced to model the interactive relationship between the vehicles so that the control method of connected autonomous vehicles is more reasonable. The research results show that the distributed control method of connected autonomous vehicles proposed in this paper is feasible and achieves the goal of distributed method.

Finally, this paper studies the control method of connected autonomous vehicles from a new point of view. The research inspiration originates from the difference of design rules of different intersections. In order to make the control method learned at one intersection can be extended to other intersections, an imitation learning method that takes into account the dynamics of the environment is proposed to solve the problem.

In conclusion, this article mainly studies the control methods of connected autonomous vehicles in signal-controlled intersections and unsignalized intersections, and expands the research on the same type of intersections with different dynamics. The research results show that it is feasible to use artificial intelligence technology to control the connected autonomous vehicles at an intersection.

Key Words:Control Methods, Reinforcement Learning, Traffic Signal Control, Multi-agent Learning, Graph Neural Network

龙可可

入学时间：2018级

答辩时间：2021年

论文题目：单交叉口网联混合交通流通行间隔计算方法

中文摘要

摘要

智能网联汽车在降低驾驶人工作负荷、提高交通效率和安全、减少油耗和尾气排放等方面体现出巨大潜力，被广泛认为是未来交通发展的新方向。在网联环境下，车路协同有望提升交叉口交通安全和效率，具有广阔的发展和应用前景。

在此背景下，本文旨在研究混合交通流条件下的通行间隔设计方案，并将其应用在典型单交叉口场景，研究该方案对交叉口通行效率、安全等方面的影响。

首先，在网联自动驾驶车辆渗透率为100%的交通环境下，提出基于动态通行间隔的交叉口信号-轨迹协同控制方案。该方案制定了通行间隔动态设计方法，同时将其应用在交叉口的信号-轨迹协同优化中，从而研究通行间隔动态设计对交叉口效率及安全的影响。本章首先阐述了通行间隔的计算原理，然后选取典型交叉口场景进行建模，构造了信号-轨迹协同优化模型，仿真结果表明，所提出的方法通过对通行间隔的动态设计满足安全要求。且相对于传统感应式控制可以降低24.9%的延误以及提高5.5%的燃油效率，且该方法计算时间短，满足落地应用的需求。

在此基础上，将场景扩展到网联混合交通流环境下的单交叉口。本章旨在提出混合交通流场景下的通行间隔计算方法。首先分析通行间隔的影响因素，综合考虑了车辆类型以及人类驾驶车辆的不确定性，随后，提出通行间隔动态设计方法，设计了其决策时间与结束时间。最后提出了针对行人及非机动车等慢行交通的信号灯设计方法，并从慢行交通安全的角度对信号配时提出要求。

最后，搭建仿真平台对所提出的方法是否能有效规避交叉口内部冲突车辆之间的碰撞进行验证。结果显示，与传统固定通行间隔的交叉口信号控制相比，本文提出的动态通行间隔设计方案可以在一定程度上合理减少陷入两难区的当量车辆数，在安全方面，能保证冲突车辆之间满足安全要求，在效率方面，本章提出的动态通行间隔设计方案对车辆延误没有明显负面影响。

本论文从效率安全油耗等多个方面对单交叉口网联混合交通流通行间隔计算方法进行了探索性研究，研究成果对网联混合流环境下的信号灯控制方案设计具有一定的实际指导意义。

关键词：车路协同，网联自动驾驶车辆，通行间隔，交叉口安全，慢行交通

英文摘要

ABSTRACT

Connected and automated vehicles (CAV) technique embody great potential in reducing driver workload, improving traffic efficiency and safety, and reducing fuel consumption and emissions. This technique is widely regarded as a new direction for future transportation development. At intersections, vehicle-road collaboration is expected to improve intersection safety and efficiency and has broad effect and application prospects.

This paper aims to study the right-of-way switching interval design scheme in a partially connected and automated traffic environment and apply it to a typical single intersection scenario to investigate its impact on intersection traffic efficiency, safety, and fuel consumption.

First, an integrated optimization of traffic signals and vehicle trajectories based on the dynamic green interval is proposed in a traffic environment with 100% penetration of CAV. The signal is optimized to improve intersection efficiency, and the calculation of green interval (GI) serves as constrains to guarantee vehicle safety during signal changing. Then the vehicle trajectory in the approach lane and inside intersection is optimized to increase fuel efficiency. The proposed method is evaluated by microscopic simulation, comparing with the actuated signal control (ASC) method and an ad hoc cooperation method between traffic signals and vehicles. Results indicate that the proposed control algorithm is effective in preventing conflicts during signal changing periods. Efficiency and fuel efficiency are improved. The benefit is 24.9% on vehicle delay and is 5.5% on fuel efficiency. The proposed system can potentially be used in real-time.

Then, the scenario is extended to the partially connected and automated traffic environment. This chapter aims to propose a dynamic calculation method for the right-of-way switching interval. First, the influencing factors of the right-of-way switching interval are analyzed. The uncertainty of human driven vehicles is considered comprehensively. Then, the dynamic design method the right-of-way switching interval is proposed. Besides, a detailed signal timing design for non-motorized vehicles is proposed from the perspective of slow-moving traffic safety.

In the last chapter, a simulation platform is built to verify the improvement of the proposed method in intersection safety and efficiency. The results show that compared with the traditional intersection signal control with a fixed yellow signal, the proposed method reduces the number of equivalent vehicles driving into the dilemma zone. In terms of safety, the PET values of potential conflicts are guaranteed to meet the safety requirements. The proposed method has no significant negative impact on efficiency.

This thesis conducts an exploratory study on the dynamic calculation method of green interval in a partially connected and automated traffic environment from several aspects such as efficiency, safety, and fuel consumption. The results provide practical guidance for the signal timing in a partially connected and automated traffic environment.

Key Words:Cooperative Vehicle-Infrastructure Systems, Connected and Automated Vehicle, Green Interval, Intersection Safety, Non-motorized Traffic

梁海伦

入学时间：2018级

答辩时间：2021年

论文题目：信号控制交叉口新型混合交通流延误计算方法研究

中文摘要

摘要

信号控制交叉口作为城市路网的关键节点，是制约道路交通运行效率和安全的瓶颈所在。车辆延误是评价城市道路交叉口交通运行状态的重要指标，其实时计算亦能为信号控制策略提供反馈与输入。随着自动驾驶技术的发展，亟需回答如何基于个体级的轨迹信息计算车辆通过交叉口产生的实时延误的问题。论文的目的是在新型混合交通流环境下提出一种有效的交叉口延误计算方法，旨在基于高精度的采样网联自动驾驶车辆数据，对数据进行深度挖掘，力图仅通过数据进行全样本延误的推演，实现对交叉口车辆延误的实时感知计算。

首先，论文对延误计算相关研究进行了详细分析，从其采用的基础理论角度，将模型划分为基于累积车辆到达-驶离过程、冲击波理论两大类。论文后续的模型建立亦从两类理论出发，根据轨迹数据的特征提出定制化的处理和建模思路。

其次，论文对交叉口车辆延误机理及CAV、HDV车辆运行影响因素进行分析，建立了交叉口信号控制方案与车辆未来运行状态的解析关系。在此基础上，基于NGSIM Lankershim路段三个交叉口的数据，主要对比分析了交叉口中车辆的真实到达-驶离特征与延误模型假定特征的差异。通过结果可以看出，时变到达过程、交通需求趋于饱和时交通状态的波动及下游排队溢出是造成延误估计误差的主要原因。

考虑实际的道路交通中车辆到达的时变性，提出了新型混合交通流环境下基于车辆到达-驶离过程的实时交通延误计算模型。通过每周期不少于两辆的采样CAV的时空行驶状态对全样本车辆到达和驶离过程进行重构，并基于马尔可夫模型对关键排队点进行动态预测，计算交叉口延误。实证分析结果表明，模型预测排队曲线很好地匹配了实际值的变化。当CAV渗透率为60%时，延误估计误差在5％左右。但该模型在欠饱和场景下对渗透率的要求较高。

最后，基于冲击波理论构建了适用于低渗透率场景的延误计算模型。通过对采样CAV数据点进行分类来获得车辆状态转换的轨迹点，并通过凸优化模型计算轨迹点构成的排队形成和消散的分段曲线方程，从而计算延误时间。验证实验表明，对于欠饱和场景，当渗透率为10%时，延误计算错误率仅为7.06%。此外，考虑低渗透率、数据低质量的场景，使用轨迹点中的加减速信息来对数据集进行补充，在渗透率为20%时，延误计算误差改进幅度为0.2s左右。

论文所建立的模型旨在通过数据挖掘方法对新型混合交通流环境下的交叉口延误进行计算而不过度依赖于模型假设，论文提出的两种方法可在不同的场景下互为验证补充，同时，也是在新型混合交通流环境下对于基于车辆到达-驶离特征及冲击波理论两种延误计算思想的沿袭与展开。

关键词：交通动态，延误计算，信控交叉口，混合交通流，网联自动驾驶

英文摘要

ABSTRACT

As a key node of the urban road network, signal-controlled intersections are the bottleneck that restricts the efficiency and safety of road traffic.Besides, delay is an important indicator to evaluate the traffic operation status of urban road intersections, and its real-time calculation result can also provide feedback and input for signal control strategies. The purpose of the paper is to propose an effective method for calculating intersection delays in the mixed traffic environment. It aims to conduct in-depth data mining based on high-precision CAV data, and strive to perform full sample delays only through data.

Firstly, the paper analyzes the related research of delay calculation. From the perspective of the basic theory adopted, the model is divided into two categories based on the cumulative vehicle arrival-departure process and shock wave theory. The subsequent model establishment of the paper also starts from two types of theories.

Secondly, the paper analyzes the blocking process of vehicles at the intersection and the factors affecting the operation of CAV and HDV vehicles, and establishes an analytical relationship between the intersection signal control plan and the vehicle's future operating status. On this basis, based on the data of three intersections on the NGSIM Lankershim section, the difference between the actual arrival-departure characteristics of the vehicles at the intersection and the assumed characteristics of the delay model are analyzed. The results show that the time-varying arrival process, the fluctuation of the traffic state when the traffic demand is saturated, and the downstream queue overflow are the main reasons for the delay estimation error.

Considering the time variability of vehicle arrival in actual road traffic, a real-time traffic delay calculation model based on the arrival-departure process of vehicles in a new mixed traffic flow environment is proposed. The arrival and departure process of the full sample of vehicles is reconstructed through the spatiotemporal driving state of the sampled CAV with no less than two vehicles per cycle, and the key queuing points are dynamically predicted based on the Markov model, and the total delay at the intersection is calculated. The empirical analysis results show that when the CAV penetration rate is 60%, the delay estimation error is about 5%. However, this model has higher requirements for permeability in under saturated scenarios.

Finally, a delay calculation model suitable for low permeability scenarios is constructed based on shock wave theory. The trajectory points of the vehicle state transition are obtained by classifying the sampled CAV data points, and the segmented curve equations of queue formation and dissipation formed by the trajectory points are calculated by the convex optimization model, and the delay time can be calculated by the intersection of the vehicle trajectory and the shock wave curve. Validation experiments show that for under saturated scenarios, when the penetration rate is 10%, the delay calculation error rate is only 7.06%. In addition, considering scenarios with low penetration rate and low data quality, the acceleration and deceleration information in the track points are used to supplement the data set. When the penetration rate is 20%, the delay calculation error is improved by about 0.2s.

The model established in this paper aims to use data mining methods to calculate intersection delays in a new mixed traffic flow environment without excessively relying on model assumptions. The two methods proposed in this paper can be complementary to each other in different scenarios, and at the same time. It is also the inheritance and development of two delay calculation ideas based on vehicle arrival-departure characteristics and shock wave theory in a new mixed traffic flow environment.

Key words:Traffic dynamics,Delay calculation, Intersection, Mixed traffic, Connected and automated vehicle