电子学报 ›› 2018, Vol. 46 ›› Issue (2): 440-446.DOI: 10.3969/j.issn.0372-2112.2018.02.025

• 学术论文 • 上一篇    下一篇

基于多表观特征子模型更新的鲁棒视觉跟踪

范舜奕, 管桦, 侯志强, 余旺盛, 戴铂   

  1. 空军工程大学信息与导航学院, 陕西西安 710077
  • 收稿日期:2016-09-26 修回日期:2017-02-15 出版日期:2018-02-25 发布日期:2018-02-25
  • 作者简介:范舜奕,男,1992年生于陕西西安,硕士研究生,主要研究方向为图像处理与视觉跟踪.E-mail:704398924@qq.com;管桦,男,1962年生于湖北孝感,教授,硕士生导师,主要研究方向为图像处理与计算机视觉.
  • 基金资助:
    国家自然科学基金(No.61473309,No.61703423);陕西省自然科学基金(No.2015JM6269,No.2016JM6050)

Robust Visual Tracking Based on Sub-model Updating of Multiple Apparent Features

FAN Shun-yi, GUAN Hua, HOU Zhi-qiang, YU Wang-sheng, DAI Bo   

  1. The Information and Navigation Institute of Air Force Engineering University, Xi'an, Shaanxi 710077, China
  • Received:2016-09-26 Revised:2017-02-15 Online:2018-02-25 Published:2018-02-25

摘要: 在视觉跟踪中,传统模型更新算法在遮挡、光照变化及自身旋转等情况下通常存在鲁棒性较差的问题.为改善该性能,提出一种对多表观特征相应子模型进行选择性更新的鲁棒视觉跟踪算法.该算法首先建立候选子模型库,然后通过三个互补特征融合的粒子滤波跟踪确定当前帧目标位置和信息,最后将当前帧三种特征直方图信息与候选库中各子模型分别计算加权相似度,更新候选库后与阈值比较,判断是否更新当前子模型.实验结果表明:本文算法能够对特征相应子模型进行有效的选择性更新,与对比算法比较,在多种复杂变化的跟踪条件下,总体上能够具有更好的跟踪鲁棒性.

关键词: 视觉跟踪, 粒子滤波, 模型更新, 多特征融合, 候选子模型库, 加权相似度

Abstract: In computer vision tracking,the traditional model updating has poor robustness in solving the problem of occlusion,illumination change and self rotation.To improve these problems,this study proposes a new visual object tracking method.The algorithm firstly builds a candidate update sub-model library.Secondly,it determines the position and information of the current target by fusing the three complementary features of the tracking based on Particle Filter.Finally,the algorithm divides the three characteristic histogram of the target and the candidate model library to calculate the similarity of the reliability weights,then determines whether the candidate sub-model library and current sub-model can be updated.Results show that the algorithm can effectively select to update the sub-model.Rather than the contrast algorithms,our method can achieve a better tracking accuracy to deal with the situation of occlusion,illumination change and self rotation.The proposed method updates the target model effectively and keeps the good robustness under various tracking scenarios.

Key words: visual tracking, particle filter, model update, multi feature fusion, candidate sub-model library, weighted similarity

中图分类号: