• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学 ›› 2020, Vol. 42 ›› Issue (08): 1406-1413.

• 图形与图像 • 上一篇    下一篇

基于VGG网络的鲁棒目标跟踪算法

徐亮1,张江4,张晶1,2,3,杨亚琦5   

  1. (1.昆明理工大学信息工程与自动化学院,云南 昆明 650500;2.云南枭润科技服务有限公司,云南 昆明 650500;

    3.昆明理工大学云南省人工智能重点实验室,云南 昆明 650500;

    4.中国船舶集团有限公司第七〇五研究所昆明分部,云南 昆明 650102;5.云南省市场监督管理局,云南 昆明 650228)

  • 收稿日期:2019-12-09 修回日期:2020-02-22 接受日期:2020-08-25 出版日期:2020-08-25 发布日期:2020-08-29
  • 基金资助:
    云南省技术创新人才资助项目(2019HB113);云南省“万人计划”产业技术领军人才资助项目(云发改人事[2019]1096号)

A robust target tracking algorithm based on VGG network

XU Liang1,ZHANG Jiang4,ZHANG Jing1,2,3,YANG Ya-qi5   

  1. (1.Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500;

    2.Yunnan Xiaorun Technology Service Co.,Ltd.,Kunming 650500;

    3.Yunnan Key Laboratory of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650500;

    4.Kunming Branch of the 705th Research Institute of China State ShipBuilding Co.,Ltd.,Kunming 650102;

    5.Yunnan Administration for Market Regulation,Kunming 650228,China)

  • Received:2019-12-09 Revised:2020-02-22 Accepted:2020-08-25 Online:2020-08-25 Published:2020-08-29

摘要: 针对传统目标跟踪算法中当目标被遮挡和受光照强度变化等多种因素干扰时,相关滤波器模板更新不准确,误差逐帧累积最终导致目标跟踪失败,提出了一种基于VGG网络的鲁棒目标跟踪算法。首先通过VGG网络对第1帧输入图像中的局部上下文区域提取平均特征图来建立相关滤波器模板;然后通过VGG网络对后续帧输入图像中的局部上下文区域提取平均特征图和仿射变换平均特征图;其次与核相关滤波跟踪算法相结合,自适应确定目标位置和最终目标位置;最后自适应更新最终平均特征图和最终相关滤波器模板。实验结果表明,本文算法在目标被遮挡和受光照强度变化等多种因素干扰时,仍具有较高的目标跟踪精度和较强的鲁棒性。


关键词: 目标跟踪, VGG网络, 核相关滤波, 特征图更新, 模板更新

Abstract: In the traditional target tracking algorithm, when the target is disturbed by various factors such as occlusion and light intensity changes, the correlation filter template updates incorrectly and the error accumulates frame by frame, eventually causing the target tracking failure. Therefore, this paper proposes a robust object tracking algorithm based on VGG network. Firstly, the VGG network is used to extract the average feature map of the local context area image to establish a correlation filter template in the first frame of the input image. Secondly, the VGG network is used to extract the average feature map and the affine transformation average feature map of the local context area image in the subsequent frame of the input image. Thirdly, combining the kernel correlation filter tracking algorithm, the target position and the final target position are adaptively determined. Finally, the algorithm adaptively updates the final average feature map and the final correlation filter template in the current frame of the input image. Experimental results show that the proposed algorithm still has high target tracking accuracy and robustness when the target is disturbed by various factors such as occlusion and light intensity changes.

Key words: target tracking, Visual Geometry Group (VGG) network, kernel correlation filter, feature map update, template update