车辆跟驰模型研究进展

梳理了近70年关于跟驰模型的研究, 根据建模方法将其分为理论驱动与数据驱动2类模型, 并归纳了跟驰模型的研究热点; 从人类因素、基础设施、交通信息、异质交通流、新建模型理论5个方面对理论驱动类跟驰模型的研究进行了综述; 根据所用机器学习算法的不同, 从模糊逻辑、人工神经网络、实例学习、支持向量回归、深度学习5个方面对数据驱动类跟驰模型的研究进行了综述。分析结果表明: 理论驱动类跟驰模型以理论推演交通现象, 对影响因素的考量难以全面, 部分人类因素难以量化, 驾驶人决策制定过程的解释不够准确, 异质交通流的跟驰模型缺乏一般交通条件下有效性的理论基础和形式化证明; 数据驱动类跟驰模型以交通现象归纳交通规律, 由于数据的来源、评价指标及评价方法不同, 导致应用机器学习算法得到的模型无法系统比较; 数据驱动类模型侧重于从微观角度研究驾驶行为特性, 对复杂交通现象(如交通震荡、迟滞等)的解释性不强; 跟驰模型的研究应创新数据采集方法, 捕捉驾驶人的心理倾向、感知特性和认知能力, 并量化人类因素的影响和充分利用大数据; 数据驱动类跟驰模型应为无人驾驶技术发展提供技术支持; 在自动驾驶完全普及之前, 人工驾驶与自动驾驶混合场景下的驾驶人跟驰行为特性尚待深入研究。 

Research progress on car-following models

The researches on the car-following models in the past 70 years were reviewed. According to the modeling methods, car-following models were divided into two types: theory-driven model and data-driven model, and the hotspots were summarized. The theory-driven car-following model was reviewed from five aspects: human factor, infrastructure, traffic information, heterogeneous traffic flow, and new modeling theory. According to different machine learning algorithms, the data-driven car-following model was also reviewed from five aspects: fuzzy logic, artificial neural network, instance learning, support vector regression, and deep learning. Analysis result shows that the theory-driven car-following model can theoretically deduces the traffic phenomenon. But it is difficult to comprehensively consider the influencing factors, and some human factors are difficult to quantify, and the explanation of driver decision-making process is not accurate enough. The car-following model of heterogeneous traffic flow lacks effective theoretical basis and formal proof under general traffic conditions. The data-driven car-following models summarize the traffic rules by traffic phenomenon. Due to different of data sources, evaluation indicators and methods, the models based on machine learning algorithms cannot be systematically compared. The data-driven models focuse on micro-angles to study driving behavior characteristics, but are not very explanatory for complex traffic phenomena(such as traffic oscillation, hysteresis, etc.). The research of the car-following models should innovate the data collection method, and capture the drivers' psychological tendencies, perceptual characteristics and cognitive abilities, as well as quantify the influence of human factors and make full use of big data. The data-driven car-following models should provide technical support for the development of driverless technology. Before the automatic driving is fully popularized, the characteristics of drivers' car-following behaviors in the mixed scene of manual driving and automatic driving need to be further studied.