汽车工程 ›› 2022, Vol. 44 ›› Issue (12): 1825-1833.doi: 10.19562/j.chinasae.qcgc.2022.12.004

所属专题: 智能网联汽车技术专题-规划&控制2022年

• • 上一篇    下一篇

面向自动驾驶汽车测试的交通车辆交互过程建模

蒋渊德1,朱冰2(),赵祥模1,赵健2,郑兵兵3   

  1. 1.长安大学信息工程学院,西安  710018
    2.吉林大学,汽车仿真与控制国家重点实验室,长春  130025
    3.中航光电科技股份有限公司,洛阳  471000
  • 收稿日期:2022-06-16 修回日期:2022-07-23 出版日期:2022-12-25 发布日期:2022-12-22
  • 通讯作者: 朱冰 E-mail:zhubing@jlu.edu.cn

Modeling of Traffic Vehicle Interaction for Autonomous Vehicle Testing

Yuande Jiang1,Bing Zhu2(),Xiangmo Zhao1,Jian Zhao2,Bingbing Zheng3   

  1. 1.School of Information Engineering,Chang’an University,Xi'an  710018
    2.Jilin University,State Key Laboratory of Automotive Simulation and Control,Changchun  130025
    3.AVIC Jonhon Optronic Technology Co. ,Ltd. ,Luoyang  471000
  • Received:2022-06-16 Revised:2022-07-23 Online:2022-12-25 Published:2022-12-22
  • Contact: Bing Zhu E-mail:zhubing@jlu.edu.cn

摘要:

为满足自动驾驶汽车测试对场景真实度的要求,提出一种针对交通车辆交互关系的耦合特征建模方法。结合基于机理模型构建的虚拟数据和采集的真实场景数据建立交通车辆行为数据集;采用深度学习建立局部信息响应的交通车辆行为决策模型、跟驰模型和换道模型,结构相对简单的单体模型能提升场景模拟的可扩展性;针对自动驾驶汽车测试对模型鲁棒性要求高的问题,建立分布执行-集中对抗训练的架构进行交通车辆模型优化,提高模型对输入扰动的鲁棒性。构建交通车辆交互仿真环境对模型进行仿真,通过仿真数据与真实数据分布之间的对比和量化评价指标验证模型的有效性。

关键词: 自动驾驶车辆, 测试, 交互关系, 深度学习, 交通场景

Abstract:

In order to meet the requirements of autonomous vehicle testing on the realness of the test scenes, a coupled feature modeling method is proposed, aiming at the interaction relationship between traffic vehicles. The dataset of traffic vehicle behaviors is established by combining the virtual data built based on mechanism model and the real scene data. Deep learning is adopted to set up the behavior decision model, following model and lane change model of traffic vehicles with local information response. The single model with relatively simple structure can enhance the expandability of scene simulation. In view of the high requirements of autonomous vehicle testing on model robustness, a framework of distributed execution and centralized adversarial training is constructed to conduct traffic vehicle model optimization for enhancing its robustness to input disturbance. The simulation environment for traffic vehicle interaction is created with a simulation on the model performed, and the effectiveness of the model is verified by comparing simulation data distribution with real data one and quantifying evaluation indicators.

Key words: autonomous vehicles, testing, interaction relationship, deep learning, traffic scenes