航空学报 > 2025, Vol. 46 Issue (17): 331643-331643   doi: 10.7527/S1000-6893.2024.31643

基于Transformer的异类目标智能关联跟踪

鹿瑶1,2, 李子豪1,3, 刘准钆1,3(), 杨衍波1,3   

  1. 1.西北工业大学 自动化学院,西安 710129
    2.中国电子科技集团公司第二十研究所 多域数据协同处理与控制全国 重点实验室(二十所分部),西安 710068
    3.西北工业大学 信息融合技术教育部重点实验室,西安 710129
  • 收稿日期:2024-12-10 修回日期:2025-01-08 接受日期:2025-03-17 出版日期:2025-04-01 发布日期:2025-03-31
  • 通讯作者: 刘准钆 E-mail:liuzhunga@nwpu.edu.cn
  • 基金资助:
    国家自然科学基金(U20B2067);国家自然科学基金(62425308);国家自然科学基金(U24B20178);陕西省自然科学基础研究计划(2024JC-YBMS-457)

Intelligent data association and tracking of different types of multiple targets based on Transformer

Yao LU1,2, Zihao LI1,3, Zhunga LIU1,3(), Yanbo YANG1,3   

  1. 1.School of Automation,Northwestern Polytechnical University,Xi’an 710129,China
    2.National Key Laboratory of Multi-domain Data Collaborative Processing and Control (Division of the 20th Research Institute),The 20th Research Institute of China Electronics Technology Group Corporation,Xi’an 710068,China
    3.Key Laboratory of Information Fusion Technology,Northwestern Polytechnical University,Xi’an 710129,China
  • Received:2024-12-10 Revised:2025-01-08 Accepted:2025-03-17 Online:2025-04-01 Published:2025-03-31
  • Contact: Zhunga LIU E-mail:liuzhunga@nwpu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(U20B2067);Natural Science Basic Research Program of Shaanxi Province(2024JC-YBMS-457)

摘要:

考虑异类多目标跟踪,现有跟踪系统模型参数固定,未能充分挖掘异类目标运动特性差异并将其体现在模型参数上,从而造成跟踪精度下降、甚至出现失跟或误跟现象。针对这一问题,提出一种基于Transformer的异类目标智能关联跟踪算法。在时间滑窗下,结合稀疏注意力和自注意力机制,提取历史雷达点迹数据/短时航迹数据中的目标类型特征、点迹航迹关联特征,动态调整系统过程噪声协方差矩阵进行异类目标精细化运动建模与雷达点迹/目标航迹的一对一可靠关联,以获得高质量异类目标跟踪输出。不同杂波密度下异类目标跟踪典型场景仿真结果表明,相较于联合概率数据关联滤波、多假设跟踪、概率假设密度滤波和模型参数固定的智能关联跟踪方法,所提方法具有更高的跟踪精度和稳定性。

关键词: 异类多目标跟踪, 数据关联, 短时航迹分类, 过程噪声协方差动态调整, Transformer网络

Abstract:

Considering tracking different types of multiple targets, the current tracking system use fixed model parameters, which fail to fully exploit motion characteristics from different types of targets. This limitation can lead to decreased tracking accuracy and even track loss or mis-tracking. To address this issue, a Transformer-based intelligent data association and tracking method is proposed for different types of multiple targets. In the case of time sliding window, by integrating sparse attention and self-attention mechanisms, the target-type features and track association features are deeply explored and exploited from the historical radar data and short-time track data, to adaptively adjust the process noise covariance matrix. This is beneficial to realize the refined motion modeling of different types of targets and acquire the reliable association between every radar echo and each target track, thereby producing high-quality tracking outputs for different types of multiple targets. Simulation results in the typical scenario of tracking different types of multiple targets with varying clutter densities demonstrate that, compared with the joint probabilistic data association filter, multiple hypothesis tracking, probability hypothesis density filter, and intelligent data association tracking with fixed model parameters, the proposed method achieves superior tracking accuracy and stability.

Key words: different types of multiple targets tracking, data association, short-time track classification, process noise covariance adjustment, Transformer network

中图分类号: