航空学报 > 2023, Vol. 44 Issue (1): 326107-326107   doi: 10.7527/S1000-6893.2021.26107

基于多尺度融合的自适应无人机目标跟踪算法

薛远亮, 金国栋(), 谭力宁, 许剑锟   

  1. 火箭军工程大学 核工程学院,西安 710025
  • 收稿日期:2021-07-15 修回日期:2021-08-03 接受日期:2021-08-23 出版日期:2023-01-15 发布日期:2021-08-25
  • 通讯作者: 金国栋 E-mail:641797825@qq.com
  • 基金资助:
    国家自然科学基金(61673017)

Adaptive UAV target tracking algorithm based on multi-scale fusion

Yuanliang XUE, Guodong JIN(), Lining TAN, Jiankun XU   

  1. School of Nuclear Engineering,Rocket Force University of Engineering,Xi’an 710025,China
  • Received:2021-07-15 Revised:2021-08-03 Accepted:2021-08-23 Online:2023-01-15 Published:2021-08-25
  • Contact: Guodong JIN E-mail:641797825@qq.com
  • Supported by:
    National Natural Science Foundation of China(61673017)

摘要:

针对无人机(UAV)跟踪过程中目标的尺寸小、尺度变化大和相似物干扰等问题,提出了一种基于多尺度注意力和特征融合的自适应无人机航拍目标跟踪算法。首先,考虑到无人机视角下干扰信息多,构建了深层多样化特征提取网络,提供鲁棒表征目标的语义特征和多样化特征;其次,设计的多尺度注意力模块,抑制干扰信息的同时保留了不同尺度的目标信息;然后特征融合模块将不同层特征进行融合,有效整合了细节信息和语义信息;最后,使用多个基于无锚框策略的区域建议模块自适应感知目标的尺度变化,充分利用整合的特征信息实现对目标的精准定位与稳定跟踪。实验结果表明:该算法在数据集上的成功率和准确率为61.7%和81.5%,速度为40.5 frame/s。该算法对目标的辨别能力、尺度感知能力和抗干扰能力明显增强,能有效应对无人机跟踪过程中的常见挑战。

关键词: 无人机, 目标跟踪, 多尺度, 特征融合, 无锚框

Abstract:

To overcome the problems of small size, large variation of scale and interference of similar objects in the Unmanned Aerial Vehicle (UAV) tracking process, an adaptive UAV aerial target tracking algorithm is proposed based on multi-scale attention and feature fusion. Firstly, considering the abundant interference information in the UAV view, a deep and diverse feature extraction network is constructed to provide robust characterization of semantic features and diverse features of the target. Secondly, a multi-scale attention module is designed to suppress interference information while retaining target information at different scales. Then, the feature fusion module is used to fuse different layers of features to effectively integrate detailed and semantic information. Finally, multiple region proposal modules based on anchor-free strategy are used to adaptively perceive the scale variation of the target and make full use of the integrated feature information to achieve accurate localization and robust tracking of the target. The experiments show that the success and precision of the algorithm on the dataset are 61.7% and 81.5% with a speed of 40.5 frame/s. The algorithm has significantly enhanced target discrimination, scale perception and anti-interference capability, and can effectively cope with common challenges in the UAV tracking process.

Key words: unmanned aerial vehicle, object tracking, multi-scale, feature fusion, anchor free

中图分类号: