朱婷

车辆跟驰过程中的交通安全问题日益严峻，降低人为因素带来的安全隐患是保障交通安全、提升通行效率的关键。随着智能网联技术的迅速发展，先进驾驶辅助系统可以在一定程度上弥补驾驶员感知、决策和控制上的不足，有效避免事故发生或者降低事故风险。现有的驾驶辅助系统普遍忽略了驾乘者的驾驶习惯和舒适度，因此市场接受度较低。为推进跟驰过程中驾驶辅助系统的研发与应用，本文开展了跟驰场景下的个性化驾驶风格分析及行为预测研究，主要研究内容和成果如下：

首先，基于人工驾驶车辆的行驶轨迹数据，提取稳定跟驰事件，分析数据分布特征，并应用统计方法初步验证了跟驰行为各特征变量之间存在不可忽略的相关性。当前针对驾驶行为的研究很少考虑特征变量之间的相互作用，可能导致驾驶行为语义划分不合理和驾驶风格理解不充分的问题。为解决上述问题，本研究应用耦合隐马尔可夫模型划分驾驶行为语义，通过行为语义精准解析和集计分析，归纳行为语义的出现及转换规律，充分解读驾驶环境信息与驾驶操作之间的对应关系，深入挖掘个性化的跟驰行为特性。结果表明，跟驰行为特性是多维度耦合的特征变量共同作用的结果，相较于传统的高斯混合隐马尔可夫模型，考虑特征变量相互作用的耦合隐马尔可夫模型可以得到更合理的驾驶行为语义分割结果。该模型可以更精准地捕捉数据的潜在特性，吻合数据波动特征且具有更强的抗干扰能力，划分得到的行为语义持续时长符合实际。

然后，基于个性化跟驰行为特征，对驾驶风格的相似性量化评估方法进行研究。针对现有驾驶风格相似性研究中综合评估维度不完整和驾驶操作评估维度不具体等问题，本研究从行为语义转移规律、行为语义选择偏好和驾驶操作激进程度三个维度出发，构建了驾驶风格相似性的综合评估体系，提出了基于特征变量分布特征的驾驶操作激进程度测评方法，实现了驾驶风格相似程度的综合量化评估。对照行为语义持续时长和特征变量的分布特征，发现本研究提出的驾驶风格相似性评估方法可以得到正确的判定结论，并且通过量化相似程度可以提供更加确切可用的评估结果。对比驾驶风格三个维度的相似性结果，发现相似性结论不总是一致，例如，一些驾驶员呈现出相似的行为语义转移规律，但是行为语义选择偏好存在明显的差异性。因此，单角度分析可能导致错误的相似性分析结论，融合三维度表现可以得到更全面准确的驾驶风格相似程度量化评估结果。

最后，结合驾驶风格相似性量化评估结果，对跟驰行为预测进行研究，提出了以驾驶风格为单位的跟驰行为预测方法，旨在提高预测的精度和稳定性。考虑到驾驶行为是驾驶风格和外界环境共同作用下的产物，本研究选取历史时刻的相对距离、相对速度和加速度指标描述动态复杂的驾驶环境，作为模型的输入进行建模，同时，根据驾驶风格的相似程度分析结果，调整模型训练集的提取方式，建立考虑驾驶风格的跟驰行为预测方法，并通过实测数据验证算法的预测精度和稳定性。结果显示：仅采用驾驶风格接近的车辆轨迹数据训练模型，可以提高加速度预测的精度和稳定性，LSTM模型可以更好地完成高难度的加速度预测任务，并且预测的均方误差和平均绝对误差分别降低了29.1%和3.85%。

关键词：稳定跟驰场景，驾驶行为，耦合隐马尔可夫，驾驶风格相似性评估，跟驰行为预测

The traffic safety problem is becoming more and more serious in China. Reducing the hidden safety hazards caused by human factors is the top priority to ensure traffic safety and improve the efficiency of road traffic. With the rapid development of intelligent and connected technology, personalized driving style analysis and behavior prediction can be utilized to guide the development of driver-centered driving assistance systems. The main research contents and results are as follows:

Firstly, based on the naturalistic driving data, the stable car-following events are extracted, the data distribution characteristics are analyzed, and statistical methods are used to preliminarily verify that there is a non-negligible correlation between the characteristic variables of car-following behavior.The current research on driving behavior rarely considered the interaction effects among driving behavior variables, which may lead to unreasonable car-following behavioral semantics and insufficient understanding of driving style. In order to solve these problems, this study constructed a multivariate coupled model framework to fully understand and extract the underlying primitive driving patterns. This paper introduces the coupled hidden Markov model (CHMM) and evaluates its applicability in obtaining the car-following behavioral semantics, summarizes the emergence and transformation rules of driving modes, and deeply understands the characteristics of personalized car-following behavior. The results show that the car-following behavior is the result of multi-dimensional coupled driving behavior variables. Compared with the GMMHMM model, CHMM with 3 hidden states provides the most appropriate driving behavior pattern segmentation results. It generates reasonable durations for each behavioral semantic fragment by accounting for the dependence relationship among variables in the analysis, and can better capture the potential characteristics of data, match the characteristics of data fluctuations and have stronger anti-interference ability.

Then, based on the characteristics of personalized car-following behavior, a quantitative evaluation method for similarity of driving styles is studied.Aiming at the problems of incomplete comprehensive evaluation dimension and unspecific driving operation evaluation dimension in the existing driving style similarity researches, a similarity analysis is conducted considering three aspects of a driving behavior pattern: driving behavior pattern transfer, driving behavior pattern selection, and aggressiveness behavior. Comparing with the original distribution characteristics of behavioral semantic duration and driving behavior variables, it is found that the driving style similarity evaluation method proposed in this study can obtain correct judgment conclusions, and can provide more accurate and usable evaluation results by quantifying the degree of similarity. Comparing the similarity results of different dimensions of driving style, it is found that the similarity conclusions are not always consistent. Therefore, single-angle analysis may lead to wrong similarity analysis conclusions, and fusion of three-dimensional results can obtain more comprehensive and accurate quantitative evaluation results of driving style similarity.

Finally, combined with the results of the quantitative evaluation of driving style similarity, the car-following behavior prediction is studied, and a driving style-based car-following behavior prediction method is proposed to improve the accuracy and stability of the prediction.Considering that driving behavior is the product of the driving style and the external environment, this study selects the relative distance, relative speed and acceleration at historical moments to describe the dynamic and complex driving environment, as the input of the model. According to the driving style similarity results, a prediction method of car-following behavior considering driving style was established, and the prediction accuracy and stability of the algorithm were verified in this study. The results show that: only using driving data with similar driving styles to train the model can improve the accuracy and stability of acceleration prediction, the model can better complete the difficult acceleration prediction task, and the mean square error and mean absolute error of the prediction are respectively decreased by 29.1% and 3.85%.

Key Words: Stable car-following scenario, Driving behavior, Coupled hidden Markov model, Driving style similarity evaluation, Car-following behavior prediction