基于粒子滤波的多区域辨识性建模跟踪算法  

Robust Particle Tracking via Multi-region Based Discriminative Modeling

作　　者：季玉龙[1] 董天罡[1] 梁海军[1] 戈文一 

机构地区：[1]四川大学视觉合成图形图像技术国防重点学科实验室,四川成都610065

出　　处：《四川大学学报（工程科学版）》2013年第S1期79-83,共5页Journal of Sichuan University (Engineering Science Edition)

基　　金：国家"863"高技术研究发展计划资助项目(2012AA011804)

摘　　要：为使跟踪算法中的目标模型不被相同特征的背景像素所干扰,在粒子滤波的框架下引入多区域辨识性建模机制,提出1种新的鲁棒的目标跟踪算法。提出的算法侧重能够辨识前景与背景的最有效信息,把目标物体划分为多个子区域。通过统计子区域与背景类别的类内/类间特性选择出最具信息量的子区域。在选出的目标区域基础上,提出的算法以贝叶斯的方式同时考虑图像信息及空间信息,建立具有辨识性的目标表观模型,并在粒子滤波的更新阶段用以估量目标的所在区域。在一系列具有挑战性的视频序列上的实验结果证明了,在许多复杂场景下,提出的方法与传统的粒子滤波跟踪器相比的鲁棒性以及有效性。In order to reduce confusion from the similar background pixels in tracking object modeling procedure,a novel particle filter approach was presented for robust object tracking using a multi-region based discriminative modeling mechanism.The proposed method highlighted the most effective information of foreground/background discrimination and divided the object into a number of patches.By comparing the properties of within and between classes of the patch and background,the most informative patches were selected.Moreover,based on these selected patches,the proposed algorithm commanded both the photometric and spatial information in a Bayesian manner for target modeling.This discriminative appearance model was then used in updating step of the particle filter to measure the object region.Experimental results on several challenging video sequences verified that the proposed method is very robust and effective compared with the traditional particle filter in many complicated scenes.

关 键 词：目标跟踪 粒子滤波 辨识性 目标表观模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]