汽车工程 ›› 2024, Vol. 46 ›› Issue (7): 1239-1248.doi: 10.19562/j.chinasae.qcgc.2024.07.011

• • 上一篇    

面向自动驾驶道路场景中异常案例的多模态数据挖掘算法

王海1(),张桂荣1,罗彤3,邱梦2,蔡英凤2,陈龙2   

  1. 1.江苏大学汽车与交通工程学院,镇江 212013
    2.江苏大学汽车工程研究院,镇江 212013
    3.江苏理工学院,常州 213001
  • 收稿日期:2023-11-03 修回日期:2024-01-15 出版日期:2024-07-25 发布日期:2024-07-22
  • 通讯作者: 王海 E-mail:wanghai1019@163.com
  • 基金资助:
    国家自然科学基金(52225212);江苏省重点研发项目(BE2020083-2)

A Multi-modal Data Mining Algorithm for Corner Case of Automatic Driving Road Scene

Hai Wang1(),Guirong Zhang1,Tong Luo3,Meng Qiu2,Yingfeng Cai2,Long Chen2   

  1. 1.School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang  212013
    2.Institute of Automotive Engineering,Jiangsu University,Zhenjiang  212013
    3.Jiangsu Institute of Technology,Changzhou  213001
  • Received:2023-11-03 Revised:2024-01-15 Online:2024-07-25 Published:2024-07-22
  • Contact: Hai Wang E-mail:wanghai1019@163.com

摘要:

基于深度学习的视觉感知技术的发展有利于自动驾驶系统中环境感知技术的进步。然而,对于自动驾驶场景中的异常案例,目前的感知模型还存在一些问题。这是因为基于深度学习的感知模型的能力取决于训练数据集的分布。尤其是驾驶场景中的类别从未出现在训练集中,感知系统也往往很脆弱。因此识别未知类别和极端场景仍然是自动驾驶感知技术安全性的挑战。本文从处理数据集的角度出发,提出了一种新颖的多模态异常案例自动挖掘流程(corner case mining pipeline, CCMP)。为验证CCMP的有效性,在Waymo开放数据集的基础上构建了异常案例子集“Waymo-Anomaly”,该子集共有3 200个图像,每个图像都将包含文本中定义的异常案例场景。并且基于私有数据集Waymo-Anomaly,证明了CCMP针对异常案例场景挖掘的召回率可以达到91.7%。此外,还通过实验验证了目标检测器在包含异常案例的数据集中针对长尾分布的有效性。最终,希望从处理数据集的角度来提高自动驾驶感知模型在现实世界中的真实性。

关键词: 自动驾驶, 深度学习, 目标检测, 异常案例

Abstract:

The development of visual perception technology based on deep learning is beneficial for the advancement of environment perception technology in automatic driving systems. However, for corner cases of autonomous driving scenario, there are still some problems in the current perception model. This is because the ability of the perception model based on deep learning depends on the distribution of the training dataset. Especially when categories in the driving scene never appear in the training set, the perception system is often fragile. Therefore, identifying unknown categories and extreme scenarios remains a challenge for the safety of automatic driving perception technology. From the perspective of processing data sets, in this paper a novel multimodal automatic corner case mining process called "Corner Case Mining Pipeline (CCMP)" is proposed. In order to verify the effectiveness of "CCMP", the concern case subset "Waymo-Anomaly" on the basis of Waymo open datasets is established, with a total of 3 200 images, each of which will contain the corner case scene defined in the text. Then based on the private data set Waymo-Anomaly, it is proved that the recall rate of "CCMP" corner case mining can reach 91.7%. In addition, the effectiveness of object detectors targeting long-tailed distributions in datasets containing corner case is experimentally verified. Ultimately, the authenticity of the automatic driving perception model in the real world is expected to improve from the perspective of datasets processing.

Key words: autonomous vehicles, deep learning, object detection, corner case