基于在线判别式字典学习的鲁棒视觉跟踪  

Robust Visual Tracking Based on Online Discrimination Dictionary Learning

作　　者：薛模根[1] 朱虹 袁广林[2] 

机构地区：[1]陆军军官学院偏振光成像探测技术安徽省重点实验室,合肥230031 [2]陆军军官学院十一系,合肥230031

出　　处：《电子与信息学报》2015年第7期1654-1659,共6页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61175035;61379105);中国博士后科学基金(2014M562535);安徽省自然科学基金(1508085QF114)资助课题

摘　　要：现有子空间跟踪方法较好地解决了目标表观变化和遮挡问题,但是它对复杂背景下目标跟踪的鲁棒性较差。针对此问题,该文首先提出一种基于Fisher准则的在线判别式字典学习模型,利用块坐标下降和替换操作设计了该模型的在线学习算法用于视觉跟踪模板更新。其次,定义候选目标编码系数与目标样本编码系数均值之间的距离为系数误差,提出以候选目标的重构误差与系数误差的组合作为粒子滤波的观测似然跟踪目标。实验结果表明:与现有跟踪方法相比,该文跟踪方法具有较强的鲁棒性和较高的跟踪精度。The existing subspace tracking methods have well solved appearance changes and occlusions. However, they are weakly robust to complex background. To deal with this problem, firstly, this paper proposes an online discrimination dictionary learning model based on the Fisher criterion. The online discrimination dictionary learning algorithm for template updating in visual tracking is designed by using the block coordinate descent and replacing operations. Secondly, the distance between the target candidate coding coefficient and the mean of target samples coding coefficients is defined as the coefficient error. The robust visual tracking is achieved by taking the combination of the reconstruction error and the coefficient error as observation likelihood in particle filter framework. The experimental results show that the proposed method has better robustness and accuracy than the state-of-the-art trackers.

关 键 词：视觉跟踪 模板更新 字典学习 观测似然 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]