基于多类型变量空间的驾驶行为基元提取

Extraction of Driving Behavior Primitives Based on Multi-type Variables Space

从驾驶数据中提取驾驶行为基元是实现驾驶行为高效准确分析的重要前提。为了更好地理解驾驶行为,使驾驶行为基元能够体现不同驾驶人的驾驶行为差异,考虑驾驶行为产生时所受驾驶人的主观约束,提出基于多类型变量空间的驾驶行为基元提取方法。使用驾驶行为原始变量反映车辆运行状态和驾驶操作,将所选择的原始变量定义为基础变量集;使用基础变量构造能够反映驾驶人对车辆运行状态主观期望的变量,并将其定义为构造变量集。利用基础变量集和构造变量集生成多类型变量空间,使用贝叶斯凝聚型序列分割算法分割数据以提取驾驶行为基元。针对多类型变量空间自调节问题,提出基于分割质量优选滑窗尺寸的方法,使多类型变量空间能够自适应不同数据集的数值特性,确保基元提取的准确度。对所得基元进行特征构造和提取,利用高斯混合聚类算法对直行和弯道路段的行为基元分别进行聚类,并通过分析各类基元的统计特征得到基元的语义描述。最后,通过实例分析验证基元提取和语义解释的准确性,以及多类型变量空间的优越性。研究结果表明:所提取的驾驶行为基元具有多角度语义,不仅能够反映车辆运行状态和驾驶操作,而且能够体现驾驶人对车辆操纵决策的主观期望,有利于从因果角度更全面地理解驾驶行为。

Efficient and accurate analysis of driving behaviors requires the extraction of driving behavior primitives from driving data.A method was developed to extract driving behavior primitives based on a multi-type variable space,considering the subjective constraints of drivers during driving.This was aimed at providing a comprehensive understanding of driving behaviors and ensure that driving behavior primitives reflect the diverse behaviors of different drivers.The original variables that capture driving behaviors,including the vehicle's running states and driving maneuvers were used.The selected original variables formed the base variable set.Additionally,basic variables were used to construct more involved variables representing drivers'subjective expectations of vehicle running states,known as constructed variable sets.By combining the base and constructed variable sets,a multi-type variables space was generated.The BMASS(Bayesian Model-based Agglomerative Sequence Segmentation)was employed to segment the data and extract driving behavior primitives.To address the self-adjusting problem of the multi-type variables space,a method that selects the sliding window size based on the segmentation quality was proposed.This method enables seamless adaptation of the variable space to diverse datasets with varying numerical characteristics,ensuring accurate extraction of driving behavior primitives.Furthermore,features of the driving behavior primitives extracted were constructed and employed for GMM to cluster driving behavior primitives in both straight and turning sections.As a result,a primitive semantic description was obtained by analyzing the statistical characteristics of different primitive clusters.Experimental verification was conducted to assess the accuracy of primitive extraction,semantic interpretation,and the advantages of the multi-type variables space.The extracted driving behavior primitives exhibit multi-angle semantics,reflecting not only the vehicle running states and driving maneuvers but also drivers'subjective expectations in vehicle control decisions.The study results highlight the accuracy and semantic richness of the extracted primitives,reinforcing the superiority of the proposed method.