欢迎访问《汽车安全与节能学报》,

汽车安全与节能学报 ›› 2022, Vol. 13 ›› Issue (4): 697-704.DOI: 10.3969/j.issn.1674-8484.2022.04.011

• 智能驾驶与智慧交通 • 上一篇    下一篇

用于自动驾驶仿真测试的车—车事故场景复杂度评价

李平飞1,2(), 金思雨1(), 胡文浩3, 高立1, 车瑶栎1, 谭正平1,2, 董小飞4   

  1. 1.西华大学 汽车与交通学院,成都 610039,中国
    2.四川西华交通司法鉴定中心,成都 610039,中国
    3.国家市场监督管理总局缺陷产品管理中心,北京 100191,中国
    4.上海机动车检测认证技术研究中心有限公司,上海 201805,中国
  • 收稿日期:2021-11-29 修回日期:2022-08-09 出版日期:2022-12-31 发布日期:2023-01-01
  • 作者简介:李平飞(1977—),男(彝),云南,副教授。E-mail:xhlpf12@mail.xhu.edu.cn
    金思雨(1997—),女(汉),四川,硕士研究生。E-mail:1404937658@qq.com
  • 基金资助:
    上海汽车工业科技发展基金项目(1828)

Complexity evaluation of vehicle-vehicle accident scenarios for autonomous driving simulation tests

LI Pingfei1,2(), JIN Siyu1(), HU Wenhao3, GAO Li1, CHE Yaoyu1, TAN Zhengping1,2, DONG Xiaofei4   

  1. 1. School of Automobile and Transportation, Xihua University, Chengdu 610039, China
    2. Sichuan Xihua Jiaotong Forensics Center, Chengdu 610039, China
    3. State Administration for Market Regulation, Defective Product Administrative Center, Beijing 100191, China
    4. Shanghai Motor Vehicle Inspection Certification &Tech Innovation Center, LTD, Shanghai 201805, China
  • Received:2021-11-29 Revised:2022-08-09 Online:2022-12-31 Published:2023-01-01

摘要:

为解决自动驾驶仿真测试场景选取依据缺乏问题,进行了车-车事故场景复杂度评价。从中国国家车辆事故深度调查体系(NAIS)数据库中,筛选出670例车-车事故;根据场景不同维度,提取13项变量;基于信息熵理论,建立了车-车事故场景复杂度评价模型。用逻辑回归得到变量水平优势比,来计算水平复杂度;使用反向传播(BP)神经网络算法,获取场景各维度及变量权重。用该模型计算各个事故案例的场景复杂度,通过K-means聚类将场景复杂度分为4个等级,得到各场景具有优势占比的特征及死亡情况。结果表明:4级复杂度场景占比1.6%,但死亡率高达90.9%,此类场景值得重点关注。

关键词: 自动驾驶, 车—车事故, 事故场景, 复杂度, 信息熵

Abstract:

The complexity of vehicle-vehicle accident scenarios was evaluated to solve the problem of lack of basis for selecting autonomous driving simulation test scenarios. 670 vehicle-vehicle accidents were selected from the National Automobile Accident In-depth Investigation System (NAIS) database of China. 13 variables were extracted according to different dimensions of the scenario. The complexity evaluation model of vehicle-vehicle accident scenarios was established based on information entropy theory. The logical regression was used to obtain the variable level odds ratio to calculate the level complexity, and the dimensions and variables weights of the scenario were obtained by Back Propagation (BP) neural network algorithms. The model was applied to calculate the complexity of the scenario for each accident case. The scenario complexity was divided into four levels by K-means clustering. The dominant characteristics and deaths of various scenarios were obtained. The results show that level 4 complexity scenarios account for 1.6 %, but with a mortality rate of 90.9 %, such scenarios deserve priority attention

Key words: autonomous driving, vehicle-vehicle accidents, accident scenarios, complexity, information entropy

中图分类号: