《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (S2): 157-162.DOI: 10.11772/j.issn.1001-9081.2022121853

• 多媒体计算与计算机仿真 • 上一篇    

面向效率与多样性的自动驾驶关键场景生成方法

李爽1,2, 周悦2(), 张昕2, 张孟2   

  1. 1.复旦大学 计算机科学技术学院,上海 201112
    2.上海市计算机软件测评重点实验室(上海计算机软件技术开发中心),上海 201112
  • 收稿日期:2022-12-13 修回日期:2023-03-27 接受日期:2023-03-27 发布日期:2024-01-09 出版日期:2023-12-31
  • 通讯作者: 周悦
  • 作者简介:李爽(1983—),女,黑龙江海伦人,高级工程师,硕士,CCF会员,主要研究方向:网络安全、车联网安全、数据安全评估
    周悦(1994—),男,安徽合肥人,硕士,CCF会员,主要研究方向:入侵检测、机器学习、自动驾驶测试
    张昕(1995—),女,湖北孝感人,硕士,CCF会员,主要研究方向:网络攻防、车联网渗透测试
    张孟(1990—),男,安徽蚌埠人,高级工程师,博士,CCF会员,主要研究方向:网络安全、数据安全评估。

Critical scenario generation method oriented to efficiency and diversity for autonomous driving

Shuang LI1,2, Yue ZHOU2(), Xin ZHANG2, Meng ZHANG2   

  1. 1.School of Computer Science,Fudan University,Shanghai 201112,China
    2.Shanghai Key Laboratory of Computer Software Testing & Evaluating (Shanghai Development Center of Computer Software Technology),Shanghai 201112,China
  • Received:2022-12-13 Revised:2023-03-27 Accepted:2023-03-27 Online:2024-01-09 Published:2023-12-31
  • Contact: Yue ZHOU

摘要:

针对自动驾驶关键场景生成耗时长和容易生成大量极端场景的问题,提出一种面向效率和多样性的自动驾驶关键场景生成方法S-AD(Surrogate model-ADversarial example)。首先利用参数随机生成一批随机场景,通过模拟驾驶获得自动驾驶系统(ADS)在随机场景中的驾驶结果;接着根据随机场景和驾驶结果训练代理模型;随后使用对抗样本生成算法对代理模型发起攻击,生成代理模型认为能引发ADS碰撞的对抗场景;最后通过模拟驾驶对抗场景进行验证,确实能引发ADS碰撞的对抗场景为生成的关键场景。与当前研究较多的基于遗传算法的关键场景生成方法(GA)对比,S-AD在获得代理模型后(平均需要1 163.7 s)生成一个关键场景平均仅需要0.084 s,而GA需要95.12 s,在生成13个以上的场景时,S-AD的时间消耗远少于GA;针对切入场景中的变道距离(DLC)参数,S-AD生成的DLC有75%分布在8.5~11.5 m,而GA生成的DLC有75%集中在9~10 m,S-AD生成的场景多样性更好;在与非碰撞场景的距离上,75%的S-AD场景分布在10~20 m,75%的GA场景分布在15~23 m,GA生成的场景更加极端。实验结果表明,S-AD在生成效率、场景多样性和非极端性上要优于对比的GA方法。

关键词: 自动驾驶测试, 模拟驾驶, 关键场景生成, 代理模型, 对抗样本生成

Abstract:

In order to solve the problems of consuming too much time and generating a large number of extreme scenarios, an efficiency and diversity oriented generation method of critical scenarios for autonomous driving system called S-AD (Surrogate model-ADversarial example) was proposed. Firstly, a batch of random scenarios were generated by randomly choosing parameter values, and the driving results of Autonomous Driving System (ADS) in the random scenarios were obtained by driving simulations. Secondly, a surrogate model was trained with random scenarios and driving results. Thirdly, an adversarial example generation algorithm was used to attack the surrogate model, and generate adversarial scenarios where the surrogate model could cause ADS collision. Finally, driving simulation was used to verify the adversarial scenarios, and scenarios that could actually cause ADS collision were generated as critical scenarios. Compared with critical scenario generation method based on Genetic Algorithm (GA), which is widely studied at present, S-AD only needs 0.084 s on average to generate one scenario after obtaining the surrogate model (takes 1 163.7 s on average), while GA needs 95.12 s. When generating more than 13 scenarios, S-AD consumes far less time than GA. According to the Distance Lane Change (DLC) parameter in the cut-in scenarios, 75% of the DLCs generated by S-AD are distributed between 8.5 and 11.5 m, while 75% of the DLCs generated by GA are concentrated between 9 and 10 m, which means the diversity of the scenarios generated by S-AD is better. In the experiment of the distance from non-collision scenarios, 75% of S-AD scenarios are distributed between 10 and 20 m, and 75% of GA scenarios are distributed between 15 and 23 m, which means the scenarios generated by GA are more extreme. Experimental results show that S-AD is superior to the GA method in terms of generation efficiency, scenario diversity and non-extreme.

Key words: autonomous driving test, driving simulation, critical scenario generation, surrogate model, adversarial example generation

中图分类号: