欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (4): 563-571.DOI: 10.11996/JG.j.2095-302X.2021040563

• 图像处理与计算机视觉 • 上一篇    下一篇

基于深度估计和特征融合的尺度自适应目标跟踪算法

  

  1. 1. 西安科技大学计算机科学与技术学院，陕西 西安 710054；
    2. 西安科技大学机械工程学院，陕西 西安 710054
  • 出版日期:2021-08-31 发布日期:2021-08-05
  • 基金资助:
    国家重点研发计划项目(2019YFB1405000)；陕西省自然科学基础研究计划项目(2019JM-162，2019JM-348)

Scale adaptive target tracking algorithm based on depth estimation and#br# feature fusion

  1. 1. College of Computer Science and Technology, Xi’an University of Science and Technology, Xi’an Shaanxi 710054, China;
    2. College of Mechanical Engineering, Xi’an University of Science and Technology, Xi’an Shaanxi 710054, China
  • Online:2021-08-31 Published:2021-08-05
  • Supported by:
    National Key Research and Development Program (2019YFB1405000); Natural Science Basic Research Program of Shaanxi Province
    (2019JM-162, 2019JM-348)

PDF (PC)

274

摘要： 针对核相关滤波目标跟踪算法(KCF)使用单特征来描述所跟踪的目标，在复杂环境下，目标尺
度发生较大变化时，无法准确跟踪目标的问题，提出基于深度估计和特征融合的尺度自适应目标跟踪算法。首
先利用深度神经网络估计视频序列中目标的深度，建立并训练深度-尺度估计模型；在跟踪过程中，融合目标
方向梯度直方图(HOG)特征和 CN (Color Name)特征训练相关滤波器，利用深度估计网络得到目标深度值，并
利用深度-尺度估计模型得到目标的尺度值，从而在目标尺度发生变化时，能够调整目标框大小，实现尺度自
适应的目标跟踪算法。实验结果表明，与经典的 KCF 算法相比，可获得更高的精度，与尺度自适应的判别型
尺度空间跟踪(DSST)算法相比，在尺度变化较大时，跟踪速度更快；在环境复杂、目标被遮挡时，鲁棒性更好。

关键词: 目标跟踪, 相关滤波, 特征融合, 深度估计网络, 深度-尺度估计模型, 尺度自适应

Abstract: The correlation filter target tracking algorithm kernelized correlation filters (KCF) utilizes a single feature
to describe the target, which cannot accurately track the target in complex environments, or when the target scale
changes significantly. To address these problems, a scale adaptive target tracking algorithm based on depth estimation
and feature fusion was proposed. Firstly, the depth of the target in the video was estimated using the depth estimation
network, and the depth-scale estimation model was built and trained. In the tracking, the histogram of oriented
gradient (HOG) features and color name (CN) features of the target were fused to train the correlation filter, and the
depth of the target was obtained using the depth estimation network. In addition, the depth-scale estimation model was
employed to obtain the scale of the target. Thus, the target size can be adjusted when the target scale changes, and the
scale adaptive target tracking algorithm is realized. The experimental results show that compared with the classical
algorithm KCF, higher accuracy can be obtained, and compared with the scale adaptive algorithm discriminative scale space tracking (DSST), the tracking speed is faster when the scale changes greatly, and the robustness is better in
cases of complex environments and obscured targets.

Key words: target tracking, correlation filter, feature fusion, depth estimation network, depth-scale estimation model;
scale adaptation

中图分类号: