电子学报 ›› 2016, Vol. 44 ›› Issue (4): 838-845.DOI: 10.3969/j.issn.0372-2112.2016.04.012

• 学术论文 • 上一篇    下一篇

在线鲁棒判别式字典学习视觉跟踪

薛模根1, 朱虹1, 袁广林2   

  1. 1. 陆军军官学院偏振光成像探测技术安徽省重点实验室, 安徽合肥 230031;
    2. 陆军军官学院十一系, 安徽合肥 230031
  • 收稿日期:2014-09-16 修回日期:2015-01-20 出版日期:2016-04-25 发布日期:2016-04-25
  • 作者简介:薛模根 男,1964年10月出生于安徽合肥.博士,教授,合肥工业大学博士生导师.主要研究方向为图像处理、计算机视觉、光电防御等. E-mail:xuemogen@126.com;朱 虹 女,1987年10月出生于新疆乌鲁木齐.现为陆军军官学院硕士研究生.主要研究方向为图像处理、计算机视觉等. E-mail:729039126@qq.com
  • 基金资助:

    国家自然科学基金(No.61175035,No.61379105);中国博士后科学基金(No.2014M562535);安徽省自然科学基金(No.1508085QF114)

Online Robust Discrimination Dictionary Learning for Visual Tracking

XUE Mo-gen1, ZHU Hong1, YUAN Guang-lin2   

  1. 1. Anhui Province Key Laboratory of Polarization Imaging Detection Technology, Army Officer Academy of PLA, Hefei, Anhui 230031, China; (;
    2. Eleventh Department, Army Officer Academy of PLA, Hefei, Anhui 230031, China
  • Received:2014-09-16 Revised:2015-01-20 Online:2016-04-25 Published:2016-04-25

摘要:

传统子空间跟踪较好解决了目标表观变化和遮挡问题,但其仍存在对复杂背景下目标跟踪鲁棒性不足和模型漂移等问题.针对这两个问题,本文首先通过增大背景样本的重构误差和利用L1范数损失函数建立一种在线鲁棒判别式字典学习模型;其次,利用块坐标下降设计了该模型的在线学习算法用于视觉跟踪模板更新;最后,以粒子滤波为框架,结合提出的模板更新方法实现了鲁棒的视觉跟踪.实验结果表明:与IVT(Incremental Visual Tracking)、L1APG(L1-tracker using Accelerated Proximal Gradient)、ONNDL(Online Non-Negative Dictionary Learning)和PCOM(Probability Continuous Outlier Model)等典型跟踪方法相比,本文方法具有较强的鲁棒性和较高的跟踪精度.

关键词: 视觉跟踪, 模板更新, 字典学习, 粒子滤波

Abstract:

The traditional subspaces based visual trackers well solved appearance changes and occlusions.However,they were weakly robust for complex background and prone to model drifting.To deal with these two problems,this paper enlarges reconstruction errors of the background samples and uses L1-norm loss function to establish an online robust discrimination dictionary learning model.Then an online robust discrimination dictionary learning algorithm for template updating in visual tracking is designed via the block coordinate descent (BCD).Finally,robust visual tracking is achieved with the proposed template updating method in particle filter framework.The experimental results show that the proposed method has better performance in robustness and accuracy than the state-of-the-art trackers such as IVT(Incremental Visual Tracking),L1APG(L1-tracker using Accelerated Proximal Gradient),ONNDL(Online Non-Negative Dictionary Learning) and PCOM(Probability Continuous Outlier Model).

Key words: visual tracking, template updating, dictionary learning, particle filter

中图分类号: