航空学报 > 2025, Vol. 46 Issue (16): 331687-331687   doi: 10.7527/S1000-6893.2025.31687

基于自适应模板更新的Transformer无人机目标跟踪算法

刘芳1, 卢晨阳1(), 路言1, 王鑫2   

  1. 1.北京工业大学 信息科学技术学院,北京 100124
    2.国网北京丰台供电公司,北京 100161
  • 收稿日期:2024-12-19 修回日期:2025-02-13 接受日期:2025-04-11 出版日期:2025-04-27 发布日期:2025-04-25
  • 通讯作者: 卢晨阳 E-mail:lcy0213@emails.bjut.edu.cn
  • 基金资助:
    国家自然科学基金(61171119)

Adaptive template update-based Transformer algorithm for UAV target tracking

Fang LIU1, Chenyang LU1(), Yan LU1, Xin WANG2   

  1. 1.School of Information Science and Technology,Beijing University of Technology,Beijing 100124,China
    2.Fengtai Power Supply Bureau of Beijing Power Supply Bureau,Beijing 100161,China
  • Received:2024-12-19 Revised:2025-02-13 Accepted:2025-04-11 Online:2025-04-27 Published:2025-04-25
  • Contact: Chenyang LU E-mail:lcy0213@emails.bjut.edu.cn
  • Supported by:
    National Natural Science Foundation of China(61171119)

摘要:

无人机已被广泛应用在军事和民用领域,目标跟踪是无人机应用的关键技术之一。针对无人机目标跟踪中目标易发生形变、遮挡、尺度变化以及环境复杂等问题,提出一种基于自适应模板更新的Transformer无人机目标跟踪算法。首先,基于改进非对称注意力机制构建Transformer骨干网络,有效提取了图像特征并增强了特征对目标的表达能力。其次,提出一种基于外观变化系数的自适应模板更新策略,通过计算外观变化系数,自适应进行模板更新,提升了跟踪网络处理目标变化的能力。最后,根据搜索区域图像特征响应图计算最大置信分数,确定目标位置。仿真实验结果表明,所提算法能够有效提升无人机目标跟踪的精度,具有较好的鲁棒性。

关键词: 机器视觉, 无人机, 目标跟踪, 模板更新, 深度网络

Abstract:

Unmanned Aerial Vehicles (UAVs) have been extensively deployed in both military and civilian applications, where target tracking plays a critical role. To address challenges such as target deformation, occlusion, scale variation, and complex environmental conditions during UAV target tracking, a adaptive template update-based Transformer algorithm for UAV target tracking is proposed. Specifically, a Transformer backbone network is constructed using an improved asymmetric attention mechanism to effectively extract image features and enhance the representation of target-related information. Furthermore, an adaptive template updating strategy based on an appearance variation coefficient is introduced. By dynamically computing this coefficient, the template is updated adaptively to improve the ability of network to cope with appearance changes of the target. Finally, the target position is determined by calculating the maximum confidence score from the response map of the search region. Experimental results demonstrate that the proposed algorithm significantly improves the accuracy of UAV target tracking and exhibits strong robustness.

Key words: machine vision, Unmanned Aerial Vehicle (UAV), target tracking, template update, deep network

中图分类号: