基于动态目标建模的粒子滤波视觉跟踪算法  

Visual Tracking Based on Dynamic Object Modeling in Particle Filter

作　　者：刘洋[1] 李玉山[1] 张大朴[1] 邱家涛[1] 

机构地区：[1]西安电子科技大学电路CAD研究所,西安710071

出　　处：《光子学报》2008年第2期375-380,共6页Acta Photonica Sinica

基　　金：国家自然科学基金(60172004);国家教育部博士点基金(20010701003)资助

摘　　要：提出一种根据场景变化动态建立目标模型的粒子滤波视觉跟踪算法.该方法首先选择简单且具有互补性的特征描述当前图像,并统一采用直方图法对这些特征进行建模;然后在粒子滤波框架下,根据巴塔恰里亚测度评价各个目标特征和背景特征之间的可区分程度,动态调整特征间的置信度;并对各个特征似然函数的噪音参量进行在线估计和更新,使其似然函数的度量标准达到统一.分析和实验表明,该算法性能优于仅仅采用多特征融合进行粒子滤波视觉跟踪的方法,对摄像机运动、混淆干扰、遮挡及目标外观大小的改变具有更强的鲁棒性.A visual tracking approach based on dynamic object modeling according to background variations in particle filter is proposed. The simple and complementary feature descriptions are used to represent the current image, and these features of the object and the background are all modeled by histograms. Then, the confidence for each feature is adjusted according to its contribution to the discrimination between the object and the background, where the contribution is estimated by Bhattacharyya distance measurement. Furthermore, The noise parameter of likelihood function of each feature is estimated and updated online in particle filter, and the measurement criterion for their likelihood functions are unified, in order to maintain the discrimination and further improve the performance of particle filter tracking. The analyses and experiments show that the performance of the proposed method is superior to that of particle filter tracking method which only based on multiple features fusion, and is more robust to camera motion, clutter, occlusion and size variation of object appearance.

关 键 词：动态目标建模 视觉跟踪 粒子滤波 巴塔恰里亚距离 克罗内克函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]