设为首页 联系我们 加入收藏

当前位置: 网站首页 期刊分类目录 2025第3期 正文

基于多模态特征动态融合与信息交叉关联的 3D 多目标跟踪

作者:时间:2025-07-04点击数:



全文下载: 1.pdf


文章编号:1672-6987(2025)03-0152-07;DOI:10. 16351/j. 1672-6987. 2025. 03. 020


祥,秦 素,李 * ,吴依凡,杨浩冉 (青岛科技大学 信息科学技术学院,山东 青岛 266061)


要:自动驾驶场景的多样性以及目标交互的复杂性,使得 3D 多目标跟踪仍是一项艰巨的 任务。现有方法在构建跟踪模型时,存在多模态特征未对齐、亲和度矩阵匹配度较低、目标信 息利用不充分等问题,导致跟踪过程中存在漏检、轨迹碎片化、身份切换等现象。为解决上述 问题,提出一种基于多模态特征动态融合和信息交叉关联的 3D 多目标跟踪方法。首先,提出 多模态特征动态融合模块,将点云和图像进行充分融合,以获取更具代表性的融合特征。其 次,利用检测和轨迹之间的距离以及目标自身的属性构建更具有辨别性的位置亲和度矩阵。 最后,设计信息交叉关联模块,采用过滤的方式将外观和位置信息用于数据关联。KITTI 数 据集上的实验结果表明,与其他方法相比,提出的方法取得了先进的跟踪性能,具有更好的跟 踪鲁棒性。


关键词:多模态融合;亲和度矩阵;数据关联;3D 多目标跟踪


中图分类号:TP 391 文献标志码:A


引用格式:刘祥,秦素,李辉,等 . 基于多模态特征动态融合与信息交叉关联的 3D 多目标跟 踪[J]. 青岛科技大学学报(自然科学版),2025,46(3):152-158.


LIU Xiang, QIN Su, LI Hui, et al. 3D Multi-object tracking based on multi-modal feature dynamic fusion and information cross-association[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition),2025,46(3):152-158.


3D Multi-Object Tracking Based on Multi-Modal Feature Dynamic Fusion and Information Cross-Association


LIU Xiang,QIN Su,LI Hui,WU Yifan,YANG Haoran (College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China)


Abstract:The diversity of autonomous driving scenarios and the complexity of target interac tions make 3D multi-object tracking still a daunting task. When building a tracking model in existing methods, there are problems such as misalignment of multimodal features, low match ing degree of affinity matrix, and insufficient utilization of target information, which lead to missed detection, trajectory fragmentation, and identity switching in the tracking process. To solve the above problems, a 3D multi-object tracking method based on dynamic fusion of multi

modal features and information cross-association is proposed. First, a multi-modal feature dynamic fusion module is proposed to fully fuse point clouds and images to obtain more repre sentative fusion features. Second, a more discriminative position affinity matrix is constructed using the distance between detections and trajectories as well as the attributes of the target itself. Finally, the information cross-association module is designed, and the appearance and

position information are used for data association by filtering. Experimental results on the KITTI dataset show that the proposed method achieves state-of-the-art tracking performance with better tracking robustness compared to other methods.


Key words:multi-modal fusion; affinity matrix; data association;3D multi-object tracking


收稿日期:2024-05-10

基金项目:国家自然科学基金项目(61702295);中国高校产学研创新基金项目(2021ITA05047)。

作者简介:刘 祥(1996—),男,硕士研究生 . * 通信联系人 .第 3 期

祥等:基于多模态特征动态融合与信息交叉关联的 3D 多目标跟踪


Copyright © 2011-2017 青岛科技大学学报 (自然科学版)