汽车工程 ›› 2021, Vol. 43 ›› Issue (11): 1602-1610.doi: 10.19562/j.chinasae.qcgc.2021.11.005

• • 上一篇    下一篇

基于特征融合的多层次多模态目标检测

程腾1(),孙磊1,侯登超1,石琴1,张峻宁2,陈炯3,黄鹤1   

  1. 1.合肥工业大学汽车与交通工程学院,合肥  230041
    2.国防科技大学电子对抗学院,合肥  230037
    3.蔚来汽车科技(安徽)有限公司,合肥  230071
  • 收稿日期:2021-07-05 修回日期:2021-08-02 出版日期:2021-11-25 发布日期:2021-11-22
  • 通讯作者: 程腾 E-mail:cht616@hfut.edu.cn
  • 基金资助:
    中央高校基本科研业务费专项资金(PA2021KCPY0041);安徽省新能源汽车暨智能网联汽车创新工程项目(GXXT-2020-076)

Multi-level and Multi-modal Target Detection Based on Feature Fusion

Teng Cheng1(),Lei Sun1,Dengchao Hou1,Qin Shi1,Junning Zhang2,Jiong Chen3,He Huang1   

  1. 1.School of Automobile and Traffic Engineering,Hefei University of Technology,Hefei  230041
    2.College of Electronic Engineering,National University of Defense Technology,Hefei  230037
    3.NIO Automotive Technology (Anhui) Company Limited,Hefei  230071
  • Received:2021-07-05 Revised:2021-08-02 Online:2021-11-25 Published:2021-11-22
  • Contact: Teng Cheng E-mail:cht616@hfut.edu.cn

摘要:

针对复杂环境下自动驾驶汽车环境感知的鲁棒性较低和小目标难以识别的问题,本文提出一种基于特征融合的多层次多模态融合方法。首先将图像与点云两种模态信息映射到同一维度,提取针对不同大小目标的层级特征,在此基础上进行多模态的多层次特征融合;并设计6次对比实验验证网络各模块的有效性。采用Waymo数据集结合蔚来汽车实车数据进行训练和测试,实验结果表明,该网络的检测MAP值较YOLO V3提升23.1%。

关键词: 自动驾驶, 环境感知, 层级特征融合, 多模态融合, 小目标检测

Abstract:

For the problems of low robustness of the environment perception and identification difficulty of small targets of autonomous driving in complex environment, a multi-level and multi-modal fusion method based on feature fusion is proposed in this paper. Firstly, the image and point cloud modal information are mapped to the same dimension, and the hierarchical features of different size targets are extracted. On this basis, the multi-modal multi-level feature fusion is carried out. Then, six comparative experiments are designed to verify the effectiveness of each module. Finally, the Waymo data set and NIO real car data are used for training and testing. The test results show that the detection MAP value of the network is improved by 23.1% compared with that of YOLO V3.

Key words: autonomous driving, environmental perception, hierarchical feature fusion, multimodal fusion, small target detection