计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (9): 201-207.DOI: 10.3778/j.issn.1002-8331.1511-0237

• 图形图像处理 • 上一篇    下一篇

基于子区域匹配的稀疏表示跟踪算法

费博雯,邵良杉,刘万军   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
  • 出版日期:2017-05-01 发布日期:2017-05-15

Tracking algorithm based on sparse representation of sub-region matching

FEI Bowen, SHAO Liangshan, LIU Wanjun   

  1. School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2017-05-01 Published:2017-05-15

摘要: 经典稀疏表示目标跟踪算法在处理复杂视频时不免出现跟踪不稳定情况且当目标发生遮挡时易发生漂移现象。针对这一问题,提出一种基于子区域匹配的稀疏表示跟踪算法。首先,将初始目标模板划分为若干子区域,利用LK图像配准算法建立观测模型预测下一帧目标运动状态。然后,对预测的目标模型区域进行同等划分,并在匹配过程中寻找最优子区域。最后,在模板更新过程中引入一种新的模板校正机制,能够有效克服漂移现象。将该算法与多种目标跟踪算法在不同视频序列下进行对比,实验结果表明在目标发生遮挡、运动、光照影响及复杂背景等情况下该算法具有较为理想的跟踪效果,并与经典稀疏表示跟踪算法相比具有较好的跟踪性能。

关键词: 稀疏表示, 观测模型, 子区域匹配, 模板校正

Abstract: The traditional tracking method based on sparse representation may tend to be unstable when processing challenging videos and the phenomenon of tracking drifting may occur when the target is occluded during the process of object tracking. To solve the problem, a novel approach of object tracking based on sparse representation of sub-region matching is proposed. Firstly, the object template is divided into several sub-regions and the observation model is established by LK image registration algorithm to predict the object motion state of next frame. Then, the region of observation model of prediction is equally divided and the optimal sub-region in the matching process is searched. Finally, by introducing a new template correction mechanism in the process of template update, it overcomes the phenomenon of tracking drifting effectively. The experimental results demonstrate that the proposed tracking algorithm has ideal effect of tracking in the case of object occlusion, motion, illumination and complex background when the multiple object tracking algorithm is tested under different video sequences, and it has well tracking performance compared to traditional tracking method based on sparse representation.

Key words: sparse representation, observation model, sub-region matching, template correction