电子学报 ›› 2020, Vol. 48 ›› Issue (9): 1762-1768.DOI: 10.3969/j.issn.0372-2112.2020.09.014

• 学术论文 • 上一篇    下一篇

基于注意力学习的正则化相关滤波跟踪算法

仇祝令1, 查宇飞1,2, 吴敏3, 王青3   

  1. 1. 中国人民解放军空军工程大学航空工程学院, 陕西西安 710038;
    2. 西北工业大学计算机学院, 陕西西安 710072;
    3. 95894 部队, 北京 102211
  • 收稿日期:2019-12-26 修回日期:2020-01-26 出版日期:2020-09-25 发布日期:2020-09-25
  • 通讯作者: 查宇飞
  • 作者简介:仇祝令 男,1995年出生,山东人.空军工程大学在读硕士研究生,主要研究方向:目标跟踪、机器学习以及计算机视觉.
  • 基金资助:
    国家自然科学基金(No.61472442,No.61773397,No.61701524);航空科学基金(No.2020-HT-XG);中央高校基本科研业务费专项资金资助(No.3102019ZY1003,No.3102019ZY1004)

Learning Attentional Regularized Correlation Filter for Visual Tracking

QIU Zhu-ling1, ZHA Yu-fei1,2, WU Min3, WANG Qing3   

  1. 1. Aeronautics Engineering College, Air Force Engineering University, Xi'an, Shaanxi 710038, China;
    2. School of Computer Science, Northwestern Polytechnical University, Xi'an, Shaanxi 710072, China;
    3. Unit 95972 of the PLA, Beijing 102211, China
  • Received:2019-12-26 Revised:2020-01-26 Online:2020-09-25 Published:2020-09-25

摘要: 边界效应是制约相关滤波跟踪性能的一个重要因素.目前大多数方法只是简单地采用先验知识,如逆高斯分布,预设掩模等,或者分割前景目标作为正则化项,进行约束求解,并没有考虑目标的空时域特性.针对这一问题,本文提出一种基于注意力学习的正则化相关滤波跟踪算法.该方法考虑目标在空间中的分布特性,利用注意力机制学习目标的特定空间权重,适应目标在空域中的变化;同时利用目标在时域中的连续性,通过对注意力权重矩阵的约束来间接调整滤波器;最后通过交替方向乘子(ADMM)算法迭代优化模型.我们在标准的数据库上进行大量实验,结果表明本文算法能实时跟踪目标,并且在精确度和成功率上都有了一定的提升.

关键词: 单目标, 视觉跟踪, 机器学习, 正则化, 相关滤波, 注意力学习

Abstract: Boundary effect is an important factor which restricts the performance of correlation filter.At present,most methods simply use the prior knowledge,such as inverse Gaussian distribution,preset masks,etc.,or segment the foreground target to constrain solving as the regularization term,which do not consider characteristics of the target in the spatial and temporal domain.To address this problem,this paper proposes a learning attention regularized correlation filter for visual tracking.The method uses the attention mechanism to learn the specific spatial weight of the target,which can adapt to the variations of target in the spatial domain by considering the spatial distribution characteristics of the target.At the same time,this paper uses the continuity of the target in the temporal domain.The filter is indirectly adjusted by constraining the attention weight matrix.Finally,the alternating direction method of multipliers (ADMM) is employed to iteratively optimize the model.We conduct extensive experiments on the proposed method in the standard tracking database.The results show that the proposed algorithm can track the target in real time,and has a certain improvement in precision and success rate.

Key words: single target, visual tracking, machine learning, regularization, correlation filter, attention mechanism

中图分类号: