汽车工程 ›› 2021, Vol. 43 ›› Issue (7): 1022-1029.doi: 10.19562/j.chinasae.qcgc.2021.07.009

• • 上一篇    下一篇

融合毫米波雷达与深度视觉的多目标检测与跟踪

甘耀东,郑玲(),张志达,李以农   

  1. 重庆大学,机械传动国家重点实验室,重庆 400044
  • 收稿日期:2021-01-12 修回日期:2021-03-04 出版日期:2021-07-25 发布日期:2021-07-20
  • 通讯作者: 郑玲 E-mail:zling@cqu.edu.cn
  • 基金资助:
    国家自然科学基金面上项目(51875061);重庆市经信委重点实验室开放课题(19AKC9);重庆市技术创新与应用发展专项(cstc2019jscx?zdztzxX0032)

Multi⁃target Detection and Tracking with Fusion of Millimeter⁃wave Radar and Deep Vision

Yaodong Gan,Ling Zheng(),Zhida Zhang,Yinong Li   

  1. Chongqing University,State Key Lab of Mechanical Transmissions,Chongqing 400044
  • Received:2021-01-12 Revised:2021-03-04 Online:2021-07-25 Published:2021-07-20
  • Contact: Ling Zheng E-mail:zling@cqu.edu.cn

摘要:

针对现有融合毫米波雷达与传统机器视觉的车辆检测算法准确率较低与实时性较差的问题,本文中对多目标检测与跟踪进行研究。首先,利用阈值筛选和前后帧数据关联方法对毫米波雷达数据进行预处理,进而提出一种用于毫米波雷达数据跟踪的自适应扩展卡尔曼滤波算法。然后,为提高目标检测精度与速度,基于采集到的实车数据集训练卷积神经网络,完成深度视觉的多车辆检测。最后,采用决策级融合策略融合毫米波雷达与深度视觉信息,设计了一种用于复杂交通环境下前方车辆多目标检测与跟踪的框架。为验证所设计的框架,进行了不同交通环境下的实车实验。结果表明:该方法可实时检测跟踪前方车辆,具有比融合毫米波雷达与传统机器视觉的车辆检测方法更好的可靠性与鲁棒性。

关键词: 车辆检测, 目标跟踪, 毫米波雷达, 深度视觉, 自适应卡尔曼滤波, 神经网络

Abstract:

Aiming at the problems of low accuracy and poor real?time performance of the existing vehicle detection algorithm fusing millimeter wave radar and traditional machine vision, multi?target detection and tracking are studied in this paper. Firstly, the millimeter?wave radar data is preprocessed by using threshold screening and the data association of adjacent frames, and an adaptive extended Kalman filter algorithm is proposed for millimeter?wave radar data tracking. Then, for enhancing the accuracy and speed of target detection, the convolutional neural network is trained based on the real?vehicle data set collected to complete multi?vehicle detection with deep vision. Finally, a decision?level fusion strategy is adopted to fuse the information of millimeter?wave radar and deep vision, and a framework for the multi?target detection and tracking of front vehicles is designed. Real vehicle tests under different traffic environment are carried out to verify the designed framework. The results show that the method adopted can detect and track the front vehicles in real time with higher reliability and robustness, compared with the vehicle detection method fusing millimeter wave radar and traditional machine vision.

Key words: vehicle detection, target tracking, millimeter?wave radar, depth vision, adaptive Kalman filter, neural network