多特征融合的混合模型视频跟踪算法  

Hybrid Model Visual Tracking Algorithm Based on Multi-feature Fusion

作　　者：王琳[1] 陈志国[1] 孙俊[1] 

机构地区：[1]江南大学物联网工程学院,江苏无锡214122

出　　处：《小型微型计算机系统》2017年第12期2689-2693,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61502203;61300149)资助;江苏省自然科学基金项目(BK20150122)资助

摘　　要：为克服传统目标跟踪算法特征表示单一、局部结构信息利用不足等问题,提出了一种多特征融合的混合模型跟踪算法.算法首先将像素的局部外观模型和全局颜色直方图、方向梯度直方图模板进行融合,构建一种鲁棒性强的混合外观模型,然后提出一种新的离群检测策略,该策略将系数矩阵分成两个相关的部分并采用l2,1规范求解.标准测试集上的实验结果表明,本文算法在处理光照变化和遮挡等场景时具有更高的跟踪精度和鲁棒性.In order to overcome the disadvantages of just adopting single feature and lacking of local structure information in traditional tracking algorithms, a hybrid model visual tracking algorithm based on multi-feature fusion is proposed. Firstly, it structures a robust appearance model via integrating the local appearance model of intensity together with the global templates of color histograms and histogram of oriented gradient （HOG ）. Then, a strategy of outlier rejection is also proposed, which divides the sparse coefficient into two collaborative components and imposes the l 2 ,1 mixed-norm regularization. The experimental results on benchmark dataset show that the proposed method is more accurate and robust in dealing with illumination change and occlusion.

关 键 词：目标跟踪 特征融合 外观模型 离群检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]