一种基于加权时空上下文的鲁棒视觉跟踪算法  

Robust Visual Tracking via Weighted Spatio-temporal Context Learning

作　　者：徐建强[1,2] 陆耀[1,2] 

机构地区：[1]北京理工大学计算机学院,北京100081 [2]智能信息技术北京市重点实验室,北京100081

出　　处：《自动化学报》2015年第11期1901-1912,共12页Acta Automatica Sinica

基　　金：国家自然科学基金(61273273;61271374);高等学校博士学科点专项科研基金(20121101110034)资助~~

摘　　要：由于光照及外观变化、复杂背景、目标旋转与遮挡等因素的影响,给实现鲁棒的视觉跟踪带来困难.有效利用上下文(Context)中包含的有用信息有助于提升上述条件下视觉跟踪的鲁棒性.时空上下文(Spatio-temporal context,STC)算法是新近提出的一种基于时空上下文的目标跟踪算法,它利用目标周围的稠密上下文信息,取得了良好的跟踪效果.STC的不足是其同等对待整个上下文区域,没有对上下文做进一步的区分,减弱了上下文的作用.本文采用动态分区处理思想,根据上下文中不同区域与跟踪目标运动相似度大小,赋予不同权值,提出了基于加权时空上下文(Weighted spatio-temporal context,WSTC)的鲁棒视觉跟踪算法.最后在公共数据集上进行的对比实验表明,本文所提出的算法具有更好的跟踪效果和鲁棒性.Implementing a robust visual tracker is a challenging task due to many disturbing factors such as illumination changes, appearance changes, rotation, partial or full occlusion, etc. The local context surrounding of the target could provide much effective information in getting a robust tracker. The spatio-temporal context（STC） learning algorithm proposed recently considers the information of the dense context around the target and has achieved a better performance.However, STC treats the whole region of the context equally, which weakens the effectiveness of the context information.In this paper, we propose a novel weighted spatio-temporal context（WSTC） learning algorithm. Our algorithm considers the surrounding context discriminatively and incorporates a weighted matrix by evaluating the motion consistencies of different regions with the tracking target. Extensive experimental results on public benchmark databases show that our algorithm outperforms the original STC algorithm and other state-of-the-art algorithms.

关 键 词：视觉跟踪 上下文 时空上下文 加权时空上下文 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]