数据驱动跟驰模型综述

车辆跟驰模型是被交通科学与交通工程领域广泛认可的微观交通流模型，是交通流理论的基础。近年来，信息感知与获取、大数据、人工智能等技术快速发展，推动了数据驱动跟驰模型的快速发展。数据驱动跟驰模型，是以真实的车辆行驶数据为基础，利用数据科学与机器学习等理论和方法，通过样本数据的训练、学习、迭代、进化，挖掘车辆跟驰行为的内在规律。本文系统回顾了数据驱动跟驰模型在过去20余年的发展历程以及由神经网络和深度学习带动的两次研究热潮，归纳了基于传统机器学习理论的跟驰模型、基于深度学习的跟驰模型、模型与数据混合驱动的跟驰模型3类数据驱动跟驰模型，并分别介绍了其中的典型代表。分析数据源发现，尽管各种高精度轨迹数据不断涌现，目前研究仍多使用美国于2006年发布的Next Generation Simulation(NGSIM)高精度车辆轨迹数据，模型的可移植性和泛化能力值得思考与研究。提出关于模型输入、输出的3个问题：如何考虑更多驾驶行为变量，是否有必要考虑更多行为变量，现有输入、输出是否可替换。在模型测试与验证方面，发现并讨论了目前测试不充分、对比不完整、缺少统一测试集与测试标准等问题。最后，探讨了数据驱动跟驰模型原创性与成功的关键因素等问题。期望通过本文的梳理，帮助研究者更好地了解数据驱动跟驰模型的过去与现状，促进相关研究的快速发展。

A Review of Data-driven Car-following Models

A car- following model is one of the microscopic traffic flow models that are widely focused on bytransportation research and engineering. In recent years, the rapid technological advancement in information perceptionand acquisition, big data, and artificial intelligence, etc., has promoted the great development of data- driven carfollowing models. Based on data science and machine learning theory, data-driven car- following models obtain theinherent law of car- following behaviors through the training, learning, iteration and evolution of real-world vehiclemotion data. This paper reviews the evolution of data-driven car-following models over the past 20 years and analyzesits two research waves driven by neural network and deep learning, respectively. Three typical types of data-driven carfollowing models and their representatives are reviewed, including traditional machine learning- based car- followingmodels, deep learning-based car- following models, and model-data hybrid driven car- following models. Data sourceanalysis indicates that, although a variety of high-fidelity trajectory datasets are constantly emerging, the NextGeneration Simulation (NGSIM) datasets released by the United States in 2006 are still the most widely used, inparticular in recent years. Therefore, the transferability and generalization of the models are worth investigating. Wealso discuss from the following aspects: model input and output including how to involve more driving behaviorvariables, whether it is necessary to consider more behavioral variables, and whether the existing input and output canbe replaced; Model testing and verification including insufficient testing, incomplete comparison, lack of unified testdataset and test standard. At last, the key factors regarding the originality and success of data- driven car- followingmodels are discussed. It is expected that this review can help researchers better understand the past and presentsituations of data-driven car-following models and promote the progress of related research.