基于多任务结构化稀疏模型的目标跟踪  

Object tracking via multi-task structured sparse model

作　　者：陈芸[1,2] 吴飞[2] 荆晓远[2] 

机构地区：[1]江苏信息职业技术学院物联网工程系,江苏无锡214153 [2]南京邮电大学自动化学院,江苏南京210003

出　　处：《计算机工程与设计》2016年第6期1663-1667,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61272273;61170305);江苏省普通高校研究生科研创新计划基金项目(CXLX13_465);江苏省高等职业院校国内高级访问学者计划基金项目(2014FX034)

摘　　要：为提高目标跟踪算法的鲁棒性,提出一种基于多任务学习框架下结构化稀疏模型的目标跟踪算法(SSMTL)。对目标模板和观测模板构建结构化稀疏模型,充分利用目标的内部结构化信息,有效应对严重遮挡、光线变化及背景杂乱等因素;结合多任务学习思想,将密集采样的粒子视为子任务,有效地利用粒子间的相互关系,改进加速近似梯度(APG)算法获得最优目标候选;提出利用遮挡检测及在线更新模板策略进一步提高目标跟踪性能。实验结果表明,相对于其它目标跟踪算法,SSMTL算法可以更准确地跟踪目标,应对复杂背景下的严重遮挡性能更优。To improve the robustness of object tracking algorithm,a tracking algorithm based on multi-task structured sparse learning（SSMTL）was proposed.Target templates and observation objects were built as structured sparse model,the internal structure information of the target was fully utilized,and the serious occlusion,illumination change and cluttered background etc.were effectively dealt with.Combined with the method of multi-task learning,intensive sampling particles were regarded as sub tasks,and the dependence between the particles and the same features between different particles were effectively used.To obtain the optimal target candidate,the approximate gradient（APG）algorithm was improved.To further improve the trac-king performance,occlusion detection and online template updates were proposed.Compared to other target tracking algorithms,experimental results show that the proposed algorithm can track the target more accurately,and the performance is better in severe occlusion complex background.

关 键 词：结构化稀疏 多任务学习 字典更新 目标跟踪 粒子滤波 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]