航空学报 > 2020, Vol. 41 Issue (1): 323423-323423   doi: 10.7527/S1000-6893.2019.23423

基于自适应Siamese网络的无人机目标跟踪算法

刘芳, 杨安喆, 吴志威   

  1. 北京工业大学 信息学部, 北京 100124
  • 收稿日期:2019-09-02 修回日期:2019-09-17 出版日期:2020-01-15 发布日期:2019-10-17
  • 通讯作者: 杨安喆 E-mail:anzheyang@emails.bjut.edu.cn
  • 基金资助:
    国家自然科学基金(61171119)

Adaptive Siamese network based UAV target tracking algorithm

LIU Fang, YANG Anzhe, WU Zhiwei   

  1. Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
  • Received:2019-09-02 Revised:2019-09-17 Online:2020-01-15 Published:2019-10-17
  • Supported by:
    National Natural Science Foundation of China (61171119)

摘要: 无人机已被广泛应用到军事和民用领域,目标跟踪是无人机应用的关键技术之一。针对无人机跟踪过程中目标易发生形变、遮挡等问题,提出一种基于自适应Siamese网络的无人机目标跟踪算法。首先,利用2个卷积网络构建一个5层Siamese网络,通过对模板特征与当前帧图像特征进行卷积得到目标位置;其次,利用高斯混合模型对以往的预测结果进行建模并建立目标模板库;然后,从模板库中挑选出最可靠的目标模板并以此更新Siamese网络的匹配模板,使Siamese网络能够自适应目标的外观变化;最后,引入回归模型进一步精确目标位置,降低背景对网络性能的影响。仿真实验结果表明:该算法有效降低了形变、遮挡等情况对跟踪性能的影响,具有较高的准确率。

关键词: 无人机, 目标跟踪, Siamese网络, 模板匹配, 自适应更新

Abstract: UAVs have been widely used in military and civilian applications. Target tracking is one of the key technologies for UAV applications. Aiming at the problem that the target is prone to deformation and occlusion during the tracking process of the UAV, a target tracking algorithm for UAV based on adaptive Siamese network is proposed. Firstly, using two convolution networks, a 5-layer Siamese network is constructed. The target location is obtained by convolving the template features with the current frame image features. Secondly, the Gaussian mixture model is used to model the previous prediction results and establish the target template library. Thirdly, the most reliable target template is selected from the template library to update the matching template of the Siamese network, so that the Siamese network can adapt to the target. Finally, a regression model is introduced to further pinpoint the target location and reduce the impact of background on network performance. The results show that the algorithm effectively reduces the influence of deformation and occlusion on tracking performance and are highly accurate.

Key words: UAV, target tracking, Siamese network, template matching, adaptive update

中图分类号: