汽车工程 ›› 2023, Vol. 45 ›› Issue (10): 1815-1823.doi: 10.19562/j.chinasae.qcgc.2023.10.004

所属专题: 智能网联汽车技术专题-感知&HMI&测评2023年

• • 上一篇    下一篇

基于D-S证据理论的多模态结果级融合框架研究

程腾1,2,3(),侯登超1,2,3,张强4,石琴1,2,3,郭利港1,2,3   

  1. 1.合肥工业大学汽车与交通工程学院,合肥 230000
    2.安徽省智慧交通车路协同工程研究中心,合肥 250000
    3.自动驾驶汽车安全技术安徽省重点实验室,合肥 230009
    4.奇瑞汽车股份有限公司,芜湖 241000
  • 收稿日期:2023-02-23 修回日期:2023-04-04 出版日期:2023-10-25 发布日期:2023-10-23
  • 通讯作者: 程腾 E-mail:cht616@hfut.edu.cn
  • 基金资助:
    国家自然科学基金(82171012);安徽省自然科学基金(2208085MF171);中央高校基本科研业务费专项资金(JZ2023YQTD0073);汽车标准化公益性开放课题资助项目(CATARC-Z-2022-01350)

Research on Multi-modal Late Fusion Framework Based on D-S Evidence Theory

Teng Cheng1,2,3(),Dengchao Hou1,2,3,Qiang Zhang4,Qin Shi1,2,3,Ligang Guo1,2,3   

  1. 1.School of Automotive and Transportation Engineering,Hefei University of Technology,Hefei  230000
    2.Engineering Research Center for Intelligent Transportation and Cooperative Vehicle-Infrastructure of Anhui Province,Hefei  250000
    3.Key Laboratory for Automated Vehicle Safety Technology of Anhui Province,Hefei  230009
    4.Chery Automobile Co. ,Ltd. ,Wuhu  241000
  • Received:2023-02-23 Revised:2023-04-04 Online:2023-10-25 Published:2023-10-23
  • Contact: Teng Cheng E-mail:cht616@hfut.edu.cn

摘要:

多模态融合感知是自动驾驶的研究热点之一,然而在复杂交通环境下由于天气、光照等外部因素干扰,目标识别可能出现错误,融合时会不可避免地出现分类冲突问题。为此,本文提出一种基于D-S证据理论的多模态结果级融合框架,将深度神经网络的置信度得分输出并作为D-S证据理论的概率密度函数,通过证据组合修正冲突的分类结果,该框架可以解决任意模态之间融合的分类冲突问题。基于KITTI数据集对该框架进行实验验证,实验测试的结果表明,框架输出的融合结果较单一感知网络mAP值均能提高8%左右,其中Yolov3与Pointpillar的融合结果相较于Pointpillar单一网络感知结果mAP值提高32%,且在复杂交通环境下能够有效解决多模态融合后的分类冲突问题。

关键词: D-S证据理论, 多模态融合, 目标识别, 分类冲突

Abstract:

Multi-modal fusion perception is one of the research hotspots of automatic driving. However, in complex traffic environment, due to the interference of weather, illumination and other external factors, the target recognition may be wrong, leading to inevitable classification conflict during fusion. Therefore, this paper proposes a multi-modal late fusion framework based on D-S Evidence Theory. The confidence score of deep neural network is output and used as the probability density function of D-S evidence theory. By modifying the classification result of conflict through evidence combination, this framework can solve the classification conflict problem of fusion between any mode. The framework is verified by experiments based on KITTI data set. The results show that the fusion result of the framework output can increase by about 8% compared with the mAP value of a single sensing network, with the fusion result of Yolov3 and Pointpillar increasing by 32% compared with the single sensing result of Pointpillar, which can effectively solve the classification conflict after multi-mode fusion in the complex traffic environment.

Key words: D-S evidence theory, multimodal fusion, object recognition, conflict classification