电子学报

• •    

基于双模型竞争机制的目标跟踪算法

林彬1,2, 王华通3, 封全喜1,2   

  1. 1.桂林理工大学理学院,广西桂林 541004
    2.广西高校应用统计重点实验室,广西桂林 541004
    3.广东外语外贸大学信息科学与技术学院,广东广州 510006
  • 收稿日期:2022-12-07 修回日期:2023-02-19 出版日期:2023-03-02
    • 作者简介:
    • 林彬 男,1985年4月出生于广西壮族自治区桂林市.现为桂林理工大学理学院副教授. 主要研究方向为计算机视觉. E-mail: linbin@glut.edu.cn
      王华通 男,1997年7月出生于广西壮族自治区玉林市.现为广东外语外贸大学硕士研究生.主要研究方向为计算机视觉. E-mail: 20211050021@gdufs.edu.cn
      封全喜 男,1980年2月出生于湖南省衡阳市.现为桂林理工大学理学院教授.主要研究方向为机器学习及其应用. Email: fqx9904@163.com
    • 基金资助:
    • 国家自然科学基金(62166015);广西自然科学基金(2019GXNSFBA245056)

Object Tracking Algorithm Based on Dual-Model Competition Mechanism

LIN Bin1,2, WANG Hua-tong3, FENG Quan-xi1,2   

  1. 1.College of Science,Guilin University of Technology,Guilin,Guangxi 541004,China
    2.Guangxi Colleges and Universities Key Laboratory of Applied Statistics,Guilin,Guangxi 541004,China
    3.School of Information Science and Technology,Guangdong University of Foreign Studies,Guangzhou,Guangdong 510006,China
  • Received:2022-12-07 Revised:2023-02-19 Online:2023-03-02
    • Supported by:
    • National Natural Science Foundation of China(62166015);Guangxi Natural Science Foundation(2019GXNSFBA245056)

摘要:

为解决背景感知相关滤波器存在的特征表达能力不足和模型漂移问题,本文提出了一种基于双模型竞争机制的目标跟踪算法.一方面,本文基于颜色和梯度信息设计了一种简单高效的特征描述子,以实现更鲁棒的目标表观建模.另一方面,本文分别构建初始模型和变化模型作用于目标搜索区域,并根据两者的跟踪响应图置信度来决定跟踪结果.跟踪过程中,随着双模型主导地位不断地动态切换,变化模型也被赋予了可逆向学习的能力,从而达到缓解模型漂移的效果.实验结果表明,相比于基准算法,本文算法在OTB2015、TinyTLP和UAV20L三个数据集的跟踪精度分别提升5.0%、1.3%和4.1%,跟踪成功率分别提升3.8%、2.8%和1.7%,且在对不同跟踪场景实现稳定跟踪的同时能够保持25.5 fps的实时跟踪速度.

关键词: 目标跟踪, 相关滤波, 双模型竞争机制, 特征描述子, 跟踪置信度, 模型漂移

Abstract:

To solve the problems of insufficient feature expression ability and model drift in the background-aware correlation filters, this paper proposes an object tracking algorithm based on a dual-model competition mechanism.On the one hand, a simple and efficient feature descriptor that integrates color and gradient information is designed to achieve more robust target appearance modeling.On the other hand, we construct two filter models to describe the object's initial appearance and its variations, and then apply them to the target searching area respectively.The tracking results are determined by the confidence of the tracking response maps corresponding to these two models.During the tracking process, with the dynamic switching of the dominant position of the two models, the filter model for adapting to object variations is also endowed with the ability of reversible learning to alleviate the model drift.The experimental results show that, compared with the baseline tracker, the tracking precision of the proposed algorithm on OTB2015, TinyTLP and UAV20L datasets is improved by 5.0%, 1.3% and 4.1%, and the tracking success rate is improved by 3.8%,2.8% and 1.7%.The proposed algorithm can also achieve stable tracking performance for different tracking scenarios while maintaining a running speed of 25.5 fps.

Key words: object tracking, correlation filter, dual-model competition mechanism, feature descriptor, tracking confidence, model drift

中图分类号: