图像与点云三维体信息交互的3D多目标跟踪网络
CSTR:
作者:
作者单位:

青岛科技大学

作者简介:

通讯作者:

中图分类号:

TP391.4

基金项目:

国家重点研发计划; 青岛市关键技术攻关及产业化示范类项目


3D Multi-Object Tracking Network based on 3D Volume Information Interaction between Image and Point Cloud
Author:
Affiliation:

Qingdao University of Science and Technology

Fund Project:

National Key Research and Development Program of China; Key Technology Research and Industrial Demonstration Projects in Qingdao City

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    多目标跟踪是自动驾驶领域中的一个关键问题.然而,仅依赖单一图像信息或点云信息难以克服复杂场景下的跟踪挑战,而目前多模态融合的跟踪方法在融合性能、数据关联、轨迹管理等方面仍存在许多问题.为此,本文提出图像与点云三维体信息交互的3D多目标跟踪网络.首先,设计三维体特征交互模块以获取目标的三维体形态信息,得到更有判别性的特征,提升复杂场景下的定位精度.其次,设计基于三维综合运动估计的数据关联,利用卡尔曼滤波以及目标在点云中的运动信息,获取目标在下一帧中的位置预测,从而提升目标在帧间的一致性.最后,为进一步增强轨迹关联的鲁棒性,设计一种基于三维体特征的轨迹管理模块,以更好地克服目标消失—重现的关联问题.在KITTI数据集上的实验结果表明,与其他方法相比,本文提出的跟踪方法具有更好的跟踪性能.

    Abstract:

    Multi-object tracking is a crucial area in autonomous driving. However, it is difficult to overcome the challenges of complex scenes by only relying on a single image or point cloud information, and the current multi-modal fusion tracking methods still have many problems in fusion strategy, data association and trajectory management. In this paper, we propose a 3D Multi-Object tracking network based on 3D volume information interaction between image and point cloud. Firstly, we design the 3D volume feature interaction module to obtain the 3D volume morphology information of the object, so as to get more discriminative features and improve the localisation accuracy in complex scenes. Secondly, we design the data association based on 3D integrated motion estimation, using Kalman filtering and the object"s motion information in the point cloud to obtain the prediction of the object"s position in the next frame, so as to enhance the consistency of the object between frames. Finally, to further enhance the robustness of trajectory association, we design a 3D volume feature-based trajectory management module to better overcome the object disappearance-reappearance association problem. Experimental results on the KITTI dataset show that our method has better tracking performance compared with other methods.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2023-11-30
  • 最后修改日期:2024-06-17
  • 录用日期:2024-03-26
  • 在线发布日期: 2024-04-10
  • 出版日期:
文章二维码