李杨

自动驾驶技术作为高端制造业的重要组成部分推动着中国经济从“规模大”向“质量强”的方向转变。随着高级别自动驾驶技术逐步进入大规模路测阶段，高级别自动驾驶车辆将会很快与人工驾驶车辆共同出现在道路上，构成新型的混合交通流。自主换道建模旨在提高自动驾驶车辆的智能换道决策能力，确保其在与人工驾驶车辆混行的复杂多变环境中获得安全、高效、舒适、经济的换道行为。当前自主换道建模已取得大量成果，但也存在一些不足，如忽视换道交互行为的连续性和持续性、换道决策与换道执行的建模缺乏统一性、换道建模未充分考虑换道行为对周边车辆的广泛影响、对周边人工驾驶车辆行为不确定性的考虑有所欠缺、以及缺少宏微观结合的仿真测试等等。为了克服这些不足，本研究围绕混合交通流中自主协调换道行为的建模技术展开研究，通过有效提升自主换道算法的交互性能，构建一体化交互策略，实现对自主协调换道算法进行“系统最优建模”，对人工驾驶行为进行“不确定性建模”的两大关键技术突破。

鉴于数据的可获得性，以及建模理念与方法无本质差异的情况下，本研究题目中的“快速路”遂指“高速公路主线段”，重点探讨高速公路主线段上自动驾驶车辆在混合交通流环境中的自主协调换道行为建模技术。

本研究主要对五个方面的内容展开探索：1)基于HighD车辆轨迹数据集对换道轨迹进行提取，建立量化时空交互行为的换道车辆综合生存分析框架，对影响交互行为的关键影响因素进行识别；2)引入考虑优先经验回放的强化学习算法和考虑动态步长的换道轨迹规划算法，构建包含交互决策与运动规划的一体化交互策略，提升换道车辆与周边车辆的交互能力；3)建立自主协调换道算法框架，构建考虑换道车辆与周边车辆驾驶感受的多目标优化模型，实现自主换道策略从“个体最优”至“系统最优”的转变，降低换道行为所带来的负面影响；4)构建基于模糊任务困难度的人因模型，推动自主协调换道算法对于人工驾驶行为的不确定性建模，并采用轨迹数据集对于不确定性模型进行标定与验证，提高自主协调换道算法应对人工驾驶车辆的适应性；5)构建混合交通环境下自主协调换道算法的集成仿真测试平台，从宏微观视角对自主协调换道算法成果展开对比分析与实际轨迹数据验证，并探讨混合交通环境下自动驾驶策略、自动驾驶渗透率、以及混合车辆排列组合对自主协调换道算法的影响。

研究成果对于高级辅助驾驶系统中自主换道算法设计具有重要指导意义，可直接应用于仿真软件中的底层换道模型，为自主换道技术提供新的视角和实用方法，为智能交通系统的发展做出重要贡献。

限于数据和现场实验的局限性，研究尚有诸多不足，如本研究主要集中在单车的换道探索，未考虑多车同时换道的情形；尚未进行实车测试验证；对于非高速公路环境中的自主换道行为，未能充分验证换道算法的效果等等。因此，未来研究的一个重要方向是扩展换道行为的研究范围，涵盖多车同时换道的情况，并在更广泛的道路环境中进行实车测试，以进一步完善和验证自主换道算法的效果和适用性。

关键词：混合交通流，自主协调换道，交互决策，帕累托最优、不确定性建模

As an integral part of the high-end manufacturing industry, autonomous driving technology not only symbolizes cutting-edge engineering and technological progress but also propels the Chinese economy's shift from "scale expansion" to "quality enhancement." With high-level autonomous driving technologies gradually entering the large-scale road-testing phase, vehicles equipped with such technologies will soon share the roads with manually driven vehicles, creating a new form of mixed traffic flow. Autonomous lane-changing modeling aims to enhance the decision-making capability for lane-changing vehicles, ensuring their ability to perform safe, efficient, comfortable, and economical lane changes in complex environments shared with manually driven vehicles. Despite significant achievements in this research field, there remain shortcomings such as overlooking the continuity and persistence of lane-changing interaction behaviors, a lack of unity between modeling of lane-changing decisions and execution, insufficient consideration of the broad impact of lane-changing behaviors on surrounding traffic, neglecting the uncertainty in the behaviors of surrounding manually driven vehicles, and the absence of simulation testing that integrates both macro and micro perspectives.

To address these gaps, this study explores modeling techniques for lane-changing behaviors in mixed traffic, aiming to enhance the interactive performance of lane-changing algorithms. The study introduces two key technological breakthroughs: "system optimum " and "uncertainty modeling" of human driving behaviors.

Given the availability of data and the non-differentiation of modeling concepts and methods，the term "expressways" in this study specifically refers to the mainline segments of highways, with a focus on modeling autonomous coordinated lane-changing behaviors in this context.

This study delves into five key aspects: 1) Extraction of lane-changing trajectories from HighD trajectory dataset, establishing a comprehensive survival analysis framework that quantifies spatio-temporal interaction behaviors of lane-changing vehicles and identifies key influencing factors; 2) Integration of reinforced learning algorithms with priority experience replay and trajectory planning with dynamic step sizes to develop an integrated interaction strategy that enhances the interactive capability of lane-changing vehicles with surrounding vehicles; 3) Construction of a coordinated lane-changing framework that formulates a multi-objective optimization problem considering the driving experiences of lane-changing vehicles and surrounding vehicles, achieving a strategic shift from "individual optimum" to "system optimum"; 4) Development of an uncertainty model based on fuzzy task difficulty, advancing the research on uncertainty modeling in coordinated lane-changing algorithms and using trajectory datasets for calibration and validation of the uncertainty model to improve the adaptability of the coordinated lane-changing algorithm to manually driven vehicles; 5) Creation of an integrated simulation testing platform in a mixed traffic environment, conducting comparative analysis and validation with actual trajectory data from both macro and micro perspectives, and exploring the impact of automated driving strategies, penetration rates, and vehicle arrangements on the coordinated lane-changing algorithm.

The findings of this study contribute significantly to the theoretical and methodological advancement of lane-changing algorithms, enhancing the interactive capability of lane-changing algorithms and promoting research on lane-changing algorithms. Our results have important implications for the design of autonomous lane-changing algorithms in Advanced Driver Assistance Systems and can be directly applied to the underlying lane-changing models in simulation software, offering new perspectives and practical approaches for autonomous lane-changing technology and contributing to the development of intelligent transportation systems.

Due to limitations in data and field experiments, this study has several shortcomings, such as focusing primarily on individual lane-changing without considering scenarios involving multiple vehicles simultaneously changing lanes, the absence of real-vehicle testing validation, and the lack of comprehensive verification of the lane-changing algorithm's effectiveness in non-highway environments. Therefore, a crucial direction for future research is to expand the scope of lane-changing behavior exploration to include multiple vehicles changing lanes simultaneously and conduct real-vehicle testing in a broader range of road environments to further refine and validate the effectiveness and applicability of the autonomous lane-changing algorithm.

Key Words: Mixed Traffic Flow, Autonomous Coordinated Lane Changing, Interaction Decision-making Model, Pareto Optimal, Uncertainty modeling