多特征融合的在线更新目标跟踪算法  

A Novel Tracking Approach Based on Multi-feature Fusion and Feature Space Online Updating

作　　者：孟钢[1] 姜志国[1] 赵丹培[1] 高越[1] 

机构地区：[1]北京航空航天大学宇航学院图像处理中心,北京100191

出　　处：《计算机辅助设计与图形学学报》2010年第10期1788-1795,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60802043)

摘　　要：为了适应跟踪过程中目标光照条件的变化,并对目标特征进行在线更新,提出一种将局部二元模式(LBP)特征与图像灰度信息相融合,同时结合增量线性判别分析对目标进行跟踪的算法.跟踪开始前,为了获得比较准确的目标描述,使用混合高斯模型和期望最大化算法对目标进行分割;跟踪过程中,通过蒙特卡罗方法对目标区域和背景区域进行采样,并更新特征空间参数,得到目标和背景的最优分类面;最后使用粒子滤波器结合最优分类面对目标状态进行预测.通过光照变化的仿真视频和自然场景视频的跟踪实验,验证了文中算法的有效性.In this paper,we propose a novel tracking approach by combining the local binary pattern（LBP） features with the gray level information under the incremental Fisher linear discriminant to online update feature space and to deal with the problems with illumination changes in object tracking.At the beginning of tracking,to get a more accurate description of the object,expectation-maximization is employed for the segmentation.Then,Monte Carlo method is used to estimate the parameters of feature space based on sampling in the regions of object and background,and the optimal hyperplane is updated.After that,particle Filtering is used to estimate the states of the object.Experimental results show that our proposed approach is effective and outperforms the traditional tracking approaches,such as the color-based particle Filter.

关 键 词：跟踪 局部二元模式 增量线性判别分析 蒙特卡罗方法 粒子滤波 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]