北京航空航天大学学报 ›› 2015, Vol. 41 ›› Issue (4): 713-720.doi: 10.13700/j.bh.1001-5965.2014.0472

• 论文 • 上一篇    下一篇

基于时空关联图模型的视频监控目标跟踪

张诚, 马华东, 傅慧源   

  1. 北京邮电大学 智能通信软件与多媒体北京市重点实验室, 北京 100876
  • 收稿日期:2014-04-28 修回日期:2014-08-01 出版日期:2015-04-20 发布日期:2015-05-08
  • 通讯作者: 马华东(1964—),男,河南南阳人,教授,mhd@bupt.edu.cn,主要研究方向为物联网与传感网、多媒体系统与网络. E-mail:mhd@bupt.edu.cn
  • 作者简介:张诚(1990—),男,湖北黄冈人,硕士生,zhangcheng@bupt.edu.cn
  • 基金资助:

    国家863计划资助项目(2014AA015101); 国家自然科学基金资助项目(61402048); 国家工信部物联网发展专项资金; 北京市教育委员会共建项目

Object tracking in surveillance videos using spatial-temporal correlation graph model

ZHANG Cheng, MA Huadong, FU Huiyuan   

  1. Beijing Key Laboratory of Intelligent Telecommunications Software and Multimedia, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • Received:2014-04-28 Revised:2014-08-01 Online:2015-04-20 Published:2015-05-08

摘要:

多摄像机监控环境下的无重叠视域目标跟踪问题十分具有挑战性,其原因在于跟踪目标在网络中的转移与运动规律往往具有不确定性.目标跟踪的关键问题在于摄像机之间的目标关联以及如何依据网络拓扑结构来找到目标之间的对应关系.提出了一种图模型来对摄像机网络中的时空关联关系进行表达.图模型中的节点表示目标在摄像机视域中的出现区域和消失区域,边由时间与空间关系进行约束.提出了一种将目标外观模型与图模型相融合的跟踪方法,其中外观模型通过协方差描述子进行特征融合,同时,结合二部图匹配策略来解决多摄像头目标跟踪中的识别与匹配问题.在真实监控视频上的实验验证了该方法的有效性.

关键词: 目标跟踪, 多摄像机网络, 时空关联图模型, 二部图匹配, 数据关联

Abstract:

Object tracking in non-overlapping multi-camera surveillance is a challenging problem since the transition time between cameras varies greatly from individual to individual with uncertainty. The key problem of object tracking in wide areas is data association and how to find correspondences between objects via camera topology. A novel graph model was proposed to capture the spatial-temporal correlation among objects, which are moving in the camera network. Source/sink regions are graph nodes, and the graph edges are constructed by the spatial and temporal constrains. Specifically, a tracking method combining appearance model and graph model was proposed to solve the problem of object re-identification and data association via bipartite matching in multi-camera object tracking. In addition, region covariance descriptor was utilized to fuse the appearance feature. Experiments with real videos validate the proposed approach.

Key words: object tracking, multi-camera network, spatial-temporal graph model, bipartite matching, data association

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发