交通运输系统工程与信息

交通运输系统工程与信息 ›› 2022, Vol. 22 ›› Issue (3): 286-292.DOI: 10.16097/j.cnki.1009-6744.2022.03.032

• 城市多模式交通网运行仿真 • 上一篇    下一篇

自适应期望跟车间距和行为习惯的驾驶人跟驰模型

倪捷*，张凯铎，刘志强，葛慧敏   

  1. 江苏大学，汽车与交通工程学院，江苏 镇江 212013
  • 收稿日期:2021-12-22 修回日期:2022-02-25 接受日期:2022-03-07 出版日期:2022-06-25 发布日期:2022-06-22
  • 作者简介:倪捷(1982- )，女，江苏启东人，副教授。
  • 基金资助:
    国家自然科学基金

Car-following Model with Adaptive Expected Driver's Following Distance and Behavior

NI Jie* , ZHANG Kai-duo, LIU Zhi-qiang, GE Hui-min   

  1. School of Automobile and Traffic Engineering, Jiangsu University, Zhenjiang 212013, Jiangsu, China
  • Received:2021-12-22 Revised:2022-02-25 Accepted:2022-03-07 Online:2022-06-25 Published:2022-06-22
  • Supported by:
    National Natural Science Foundation of China(51905224)。

PDF

摘要： 为满足智能车辆的个性化需求，提高智能车辆人-机交互协同的满意度和接受度，构筑双层驾驶人跟驰模型框架，提出自适应驾驶人期望跟车间距和行为习惯的个性化驾驶人跟驰模型。首先，提取个体驾驶人跟驰均衡状态的数据，采用高斯混合和概率密度函数(Gaussian Mixture Model and Probability Density Function, GMM-PDF)建立第 1 层模型，即驾驶人期望跟车距离模型。然后，将期望跟车距离参数引入模型，基于高斯混合-隐马尔可夫方法(Gaussian Mixture Model and Hidden Markov Model, GMM-HMM)学习驾驶习性，建立第2层模型预测加速度，即个性化驾驶人跟驰模型。其次，研究不同高斯分量个数对模型效果的影响，对比双层模型与 Gipps 模型、最优间距模型(Optimal Distance Model, ODM)、单层模型及通用模型的性能。最后，8位被试驾驶人的自然驾驶行为数据验证结果表明：高斯分量数量与模型性能存在一定的正相关性；在最优高斯分量数量下，8位被试驾驶人在训练集上预测误差均值为0.101 m·s-2，在测试 集上为0.123 m·s-2；随机选取其中1位驾驶人的2个跟车片段数据进行模型计算，结果显示，加速度的平均误差绝对值分别为0.087 m·s-2和0.096 m·s-2，预测效果优于Gipps模型、ODM模型、单层 模型及通用模型30%以上，与驾驶人实际跟驰行为的吻合度更高。

关键词: 交通工程, 驾驶行为, 跟驰模型, 高斯混合模型, 隐马尔科夫模型

Abstract: To satisfy the personalized demand of intelligent vehicles and to improve the satisfaction and acceptance of intelligent vehicle human-computer interaction, a two-layer driver car-following model framework was constructed, and a personalized driver car-following model was proposed. The models can adapt the driver's expected following distance and behavior. Firstly, the equilibrious car-following data was extracted. The first layer model was established by using the Gaussian Mixture Model (GMM) and Probability Density Function (PDF), which was the driver's expected following distance model. Then, the expected following distance parameter was introduced into the model, and the driving behavior was learned based on Gaussian Mixture Model (GMM) and Hidden Markov Model (HMM). The second layer model, i.e., the personalized car-following model, was established to predict the acceleration of future time. Next, the effect of different numbers of GMM components on the model performance was studied, and the comparison was made among the two-layer driver car-following model, the Gipps model, the optimal distance model (ODM), monolayer model and the general model. Finally, the results of the 8 drivers' naturalistic driving behavior data show that the number of GMM components is positively correlated with the model performance. Under the optimal Gaussian model component, the mean predictive deviation of 8 drivers in the training set is 0.101 m·s-2 , and 0.123 m·s-2 in the test set. The model calculation results of randomly selecting one of the drivers' experimental data show that the mean absolute deviation of acceleration is 0.087 m·s-2 and 0.096 m·s-2 , and the prediction results are better than that of the Gipps model, the ODM model, monolayer model and the general model by more than 30% , which is moreconsistent with the actual car-following behavior of the driver

Key words: traffic engineering, driving behavior, car- following model, Gaussian Mixture Model, Hidden Markov Model

中图分类号: