钟心志
为推进智能网联车辆的研发与应用，实现高智能性与高安全性的目标，本文开展了智能网联环境下高速公路驾驶行为建模与快速仿真测试研究。本研究以高速公路为主要交通环境，针对智能网联车辆的高智能性特征，提出在其驾驶行为建模中考虑人工驾驶车辆驾驶行为典型特征；针对智能网联车辆的安全性能，提出采用考虑驾驶行为特征的快速仿真测试方法，加快智能网联车辆在极端高风险事件下的安全性能验证。
首先，本研究根据高速公路自然驾驶数据集挖掘人工驾驶车辆驾驶行为——跟驰行为以及变道行为典型特征，为实现智能网联车辆的拟人化行为决策，提高车辆的智能性提供理论基础。通过相关系数计算，显著性分析等统计方法与仿真试验相结合，验证了跟驰模型各参数间以及变道行为各特征变量间存在不可忽略的相关性且存在不同的边缘分布特征，这两类特征对于确保驾驶行为建模与仿真精度极其重要。
其后，本研究围绕如何将人工驾驶车辆跟驰行为的典型特征纳入到智能网联车辆的建模与仿真中，提高智能网联车辆的智能性，实现拟人化的跟驰决策的目标展开，在进一步验证跟驰模型参数间存在不可忽略的相关性与异质性的基础上，提出了基于HD-GMCM聚类特征的耦合模型作为跟驰行为建模与仿真的核心算法，采用HD-GMCM聚类方法对整体的跟驰行为特征进行分解，消除参数异质性特征的影响，挖掘参数间真实的耦合关系与边缘分布特征，使各类别跟驰行为特征更为突出，再通过耦合模型进行跟驰行为建模与仿真。同时从参数仿真以及交通流仿真两个角度验证基于HD-GMCM聚类特征的耦合模型在跟驰行为建模与仿真的精度，各评价指标结果显示该方法能够有效的生成符合实际交通流特征的跟驰模型参数，准确模拟实际交通流运行状态，在把握人工驾驶车辆跟驰行为特征上具备极高的优越性。
最后，针对智能网联车辆安全性能仿真测试研究，提出了采用考虑驾驶行为特征的重要性抽样测试方法代替目前常用的蒙特卡洛测试方法，加快车辆安全性能验证。通过选取智能网联环境下较为复杂的两种典型交通事件，即在智能网联车辆与人工驾驶车辆混合行驶的交通流状态下的变道切入以及紧密变道事件进行仿真测试试验，确定了两种极端事件极限状态：碰撞事件以及伤亡事件，以验证考虑驾驶行为特征的重要性抽样测试方法在提高测试效率上的优越性。通过模型加速效用对比可得，基于高斯耦合混合模型的重要性抽样方法在描述各特征变量关系和边缘分布上更具备灵活性，能够快速适应不同的极限状态，全面捕捉不同事件特征，模拟到特定事件所需的样本数最少，能够极大程度上提高智能网联车辆在极端高风险事件下的安全性能测试效率。
关键词：智能网联车辆，驾驶行为，耦合性，异质性，快速仿真测试
ABSTRACT
To promote the research and application of connected and autonomous vehicles and achieve the goals of high-level intelligence and security of the transportation system, the thesis carried out a research on travel behavior modeling and accelerated evaluation on vehicle safety on motorway in the connected and autonomous environment. The study is conducted in a motorway environment. To make connected and autonomous vehicles more intelligent, we purposed that travel behavior characteristics of human driven vehicles should be considered when modeling travel behavior of the connected and autonomous vehicles. To improve the safety performance of connected and autonomous vehicles, we proposed that accelerated simulation and evaluation method considering travel behavior characteristics of human-driven vehicles can be an alternative method of the traditional simulation and evaluation method based on Monte Carlo.
Firstly, we focused on the travel behavior characteristics of human-driven vehicles, including car-following behavior and lane-changing behavior using NGSIM vehicle trajectory datasets. We explored two features of the travel behavior. The parameters in car-following and lane-changing behavior are correlated according to the results of correlation coefficient calculations and significance tests. Besides, the marginal distributions of these parameters are different. Accommodating the two features in travel behavior modeling is important to enhance the modeling and simulation accuracy.
Secondly, we validated that overlooking the correlation and heterogeneity and the interaction between them in parameters of car-following behavior of human-driven vehicles would impose negative influence on modeling and simulation. To realize the anthropomorphic car-following decisions in connected and autonomous vehicles, we proposed that HD-GMCM–based copula method to model and simulate the car-following behavior of connected and autonomous vehicles. The HD-GMCM clustering algorithm is applied to decompose the overall travel behavior and eliminate the influence of heterogeneity. It uncovers the real correlations among parameters of car-following behaviors and individual marginal distributions of the parameters. Then the copula model is constructed to accommodate the correlations and individual marginal distribution in modeling and simulation of car-following behaviors. From the results of modeling and simulation, the new parameters generated by the HD-GMCM copula method can replicate actual traffic state. The method shows great outperformance in describing car-following characteristics of human-driven vehicles. Appling the HD-GMCM copula method to model and simulate of connectesd and autonomous vehicles will help them to imitate the behavioral characteristics of human-driven vehicle and improve their intelligence of car-following decision.
Finally, the importance sampling method considering the characteristics of travel behavior is proposed to accelerate the evaluation on safety performance of connected and autonomous vehicles. To evaluate the performance of the proposed method in acceleration of safety evaluation, especially in rare events, the model of common use-Monte Carlo is selected for comparison. Correspondingly, two typical scenarios in connected and autonomous environment are selected, cut-in event and close lane-changing event. Two rare events are considered- crash event and injury event and then limit states of the rare events can be defined. From the modeling results, the importance sampling method based on Gaussian copula mixture model can capture all the travel behavior characteristics and is flexible in different scenarios. It can accelerate the safety evaluation by reducing the samples needed to simulate predefined rare events so that the performance of connected and autonomous vehicles in rare events can be obtained with lower simulation cost.
Key words: Connected and Autonomous Vehicles, Travel Behavior, Copula, Heterogeneity, Accelerated evaluation.
