电子学报 ›› 2017, Vol. 45 ›› Issue (2): 384-393.DOI: 10.3969/j.issn.0372-2112.2017.02.017

• 学术论文 • 上一篇    下一篇

基于改进协作目标外观模型的在线视觉跟踪

宋涛1, 李鸥1, 刘广怡1, 崔弘亮2   

  1. 1. 信息工程大学信息系统工程学院, 河南郑州 450002;
    2. 南京理工大学自动化学院, 江苏南京 210094
  • 收稿日期:2015-07-07 修回日期:2015-10-13 出版日期:2017-02-25 发布日期:2017-02-25
  • 作者简介:宋涛,男,1983年出生于河南焦作,现为解放军信息工程大学信息系统工程学院博士研究生,主要研究方向为计算机视觉和目标跟踪技术.E-mail:taosong_1983@126.com;李鸥,男,1961年出生于河南郑州,现为解放军信息工程大学信息系统工程学院教授、博士生导师,主要研究方向为无线通信网络、信息融合和目标跟踪技术.E-mail:zzliou@126.com
  • 基金资助:

    国家科技重大专项(No.2014ZX03006003)

Online Visual Tracking Based on Improved Collaborative Appearance Model

SONG Tao1, LI Ou1, LIU Guang-yi1, CUI Hong-liang2   

  1. 1. Institute of Information System Engineering, Information Engineering University, Zhengzhou, Henan 450002, China;
    2. School of Automation, Nanjing University of Science and Technology, Nanjing, Jiangsu 210094, China
  • Received:2015-07-07 Revised:2015-10-13 Online:2017-02-25 Published:2017-02-25

摘要:

在不受限制的复杂环境中在线跟踪任意类型的感兴趣目标仍是一项极具挑战的难题.本文在无模型跟踪框架基础上提出一种基于改进协作目标外观模型的在线视觉跟踪方法,解决了大多数协作模型类跟踪算法在学习阶段无法有效选择正、负样本的问题.该方法根据人类视觉感知准则将目标边缘信息视为最具区分度的目标特征,提出边缘判别模型并结合动态模型和检测模块建立二级似然匹配空间,为生成模型的似然匹配去除了背景干扰;采用分块策略建立目标生成模型,为模型引入空间结构信息;利用Mean-Shift计算各子块的最终位置和匹配系数,并根据子块匹配系数为遮挡处理和模型更新提供依据.在公开视频序列上同几种流行视觉跟踪算法的对比实验结果证明了本文算法的有效性和优越性.

关键词: 在线视觉跟踪, 协作外观模型, 人类视觉感知, 二级似然匹配空间, 模型更新

Abstract:

It is still a very challenging issue to online track arbitrary targets in the unrestricted complex environment.This paper presents an online visual tracking method with improved collaborative appearance model based on model-free framework,solving the problem of most other tracking algorithms with collaborative model,which is unable to effectively select the positive and negative samples.According to the human visual perception rules,object edge information is regarded as the most discriminative feature,on which an edge discriminative appearance model is proposed.In order to remove background interference in likelihood matching space for generative model,a two-stage matching space is put forward via integrating dynamic model,detection module and edge discriminative model.The generative model based on partition strategy is constructed for space and appearance information.The final position and matching coefficient of each sub-block are calculated by mean-shift,as a basis for occlusion handling and model update.Experimental results using challenging public video sequences show the effectiveness and superiority of the proposed method,compared with other state-of-the-art visual tracking approaches.

Key words: online visual tracking, collaborative appearance model, human visual perception, two-stage likelihood matching space, model update

中图分类号: