利用深度卷积特征的无人机视觉跟踪
DOI:
作者:
作者单位:

西南石油大学

作者简介:

通讯作者:

中图分类号:

TP242.62

基金项目:

国家自然科学基金项目:开方动态环境下不确定数据的主动学习方法研究(基金号:62006200);四川省级重点研发项目(2022YFG0117);南充市校科技战略合作项 目(SXHZ026,SXJBS002,SXHZ053).


UAV visual tracking using deep convolutional feature
Author:
Affiliation:

Southwest Petroleum University

Fund Project:

Research on active learning method of uncertain data in square dynamic environment(Nos.62006200)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    由于无人机视觉跟踪视角范围广且环境复杂, 常遇到无人机飞行震动、目标遮挡、相似目标等问题, 导致无人机跟踪目标发生漂移. 因此, 本文对具有回归计算的全卷积孪生网络跟踪算法(SiamRPN)进行改进, 提出一种加强深度特征相关性的无人机视觉跟踪算法(SiamDFT). 首先, 将全卷积神经网络后三层卷积的网络宽度提升一倍, 充分利用目标的外观信息, 完成对模板帧和检测帧的特征提取;然后, 在检测帧和模板帧分别提出注意力信息融合模块和特征深度卷积模块, 两个深度的特征相关性计算方法能够有效抑制背景信息, 增强像素对之间的关联性, 高效完成分类和回归任务;最后, 采用深度互相关运算完成相似性计算, 并引入距离交并比的计算方法完成对目标的定位;实验结果表明, SiamDFT在无人机短时跟踪场景下精确率和成功率分别达到79.8%和58.3%, 在无人机长时跟踪场景下精确率和成功率分别达到73.4%和55.2%. 最后完成实景测试, 充分验证本文算法的有效性.

    Abstract:

    Due to the wide visual angle range and complex environment of UAV visual tracking, UAV flight vibration, object occlusion, similar objects and other problems are often encountered, resulting in the drift of UAV tracking targets. Therefore, this paper improves the full convolution siamese network tracking algorithm (SiamRPN) with regression calculation, and proposes a UAV visual tracking algorithm (SiamDFT) to enhance the correlation of depth features. Firstly, the network width of three-layer convolution after full convolution neural network is doubled, and the feature extraction of template frame and detection frame is completed by making full use of the appearance information of the object. Then, the attention information fusion module and feature deep convolution module are proposed in the detection frame and template frame respectively. The feature correlation calculation methods of the two depths can effectively suppress the background information, enhance the correlation between pixel pairs, and efficiently complete the tasks of classification and regression. Finally, the depth cross-correlation operation is used to calculate the similarity, and the calculation method of distance intersection over union is introduced to locate the target. The experimental results show that the accuracy and success rate of SiamDFT are 79.8% and 58.3% respectively in UAV short-term tracking scene, and 73.4% and 55.2% respectively in UAV long-term tracking scene. Finally, the real-world test is completed to fully verify the effectiveness of the algorithm.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-12-05
  • 最后修改日期:2022-04-20
  • 录用日期:2022-04-27
  • 在线发布日期:
  • 出版日期: