电子学报 ›› 2016, Vol. 44 ›› Issue (6): 1355-1361.DOI: 10.3969/j.issn.0372-2112.2016.06.014

• 学术论文 • 上一篇    下一篇

基于置信区域内多级动态层表达的类贯序蒙特卡洛视觉跟踪方法

宋涛, 李鸥, 刘广怡   

  1. 信息工程大学信息系统工程学院, 河南郑州 450002
  • 收稿日期:2014-10-30 修回日期:2015-03-23 出版日期:2016-06-25 发布日期:2016-06-25
  • 作者简介:宋涛 男,1983年9月出生于河南省焦作市.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年出生于河南省郑州市.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)

Quasi-Sequential Monte Carlo Visual Tracking Based on Multilevel Dynamic Layer Representations in Confidence Region

SONG Tao, LI Ou, LIU Guang-yi   

  1. Institute of Information System Engineering, Information Engineering University, Zhengzhou, Henan 450002, China
  • Received:2014-10-30 Revised:2015-03-23 Online:2016-06-25 Published:2016-06-25

摘要:

视觉跟踪是智能监控、机器人和视觉导航等领域的核心技术.针对现有类贯序蒙特卡洛跟踪方法复杂度高、实时性差的问题,本文提出了一种融合置信区域内多级动态层表达的跟踪方法,采用更加可靠、有效的粒子模拟状态后验概率.该方法利用检测模块得到目标可能出现的置信区域,根据真实目标尺寸给出一种粒子采样策略,每个粒子代表一级动态层表达,并为每个粒子建立双层运动模型;构建Mean-Shift分块观测模型以引入空间和外观信息,同时根据子块的匹配程度计算粒子权值、检测目标遮挡状态并提出模型更新策略.在公开视频序列上同经典粒子滤波和Mean-Shift等算法的实验对比结果证明了本文算法具有较优的跟踪准确度和实时性.

关键词: 视觉跟踪, 置信区域, 双层多级运动模型, 分块观测模型, 模型更新

Abstract:

Visual tracking is a core technology for the application domains of intelligent monitoring, robotics and vision navigation, etc.Aiming at the problem of high complexity and poor real-time performance in the existing quasi-sequential Monte Carlo tracking algorithms, this paper presents a method based on multilevel dynamic layer representations, which simulates the posteriori probability of a state using more reliable and effective particles.Then a sampling strategy is proposed in confidence areas derived from the detector, in which each particle represents a dynamic representation and has a two-layer motion model.The observation model based on parted-mean-shift is constructed for space and appearance information.Depending on the degree of matching sub-blocks, the weight of particles is calculated and a way to detect the occlusion state of an object is put forward for realtime model update.Experimental results using challenging public video sequences show better accuracy and efficiency of the proposed method, compared with the classical particle filter and mean-shift algorithms, etc.

Key words: visual tracking, confidence region, two-layer multilevel motion model, block-based observation model, model update

中图分类号: