适用于多目标跟踪的MSPF算法研究  

Research on MSPF algorithm for multi-target tracking

作　　者：韩明[1] 刘教民[1,2] 王震洲[2] 王静云[3] 

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]河北科技大学信息科学与工程学院,河北石家庄050018 [3]河北省纺织纤维玖仟质量认证咨询中心,河北石家庄050091

出　　处：《河北工业大学学报》2012年第5期15-19,共5页Journal of Hebei University of Technology

基　　金：河北省自然科学基金(F2012208004)

摘　　要：针对背景发生变化和目标遮挡时多目标跟踪的实时效果差和准确性低的问题,提出了均值漂移与粒子滤波相融合(Mean shiftparticle filter,MSPF)多目标跟踪算法.在PF理论框架下,使用均值漂移对随机粒子样本进行重新分配,使粒子向目标状态的最大后验核密度估计方向移动,在均值漂移迭代过程中对样本权值进行更新;最后利用均值漂移算法的迭代运算和运动目标前一帧和当前帧的关系得到目标在下一帧的位置.实验结果表明该算法和传统的均值漂移算法相比具有更准确的跟踪结果,对于背景更新和目标遮挡的视频序列具有更强的鲁棒性.In view of the the low accuracy and poor real-time effect of multi-target tracking when the background changes with object occlusion,we propose a multi-target tracking algorithm based on fusion of Mean-Shift and Particle Filter(MSPF).With the theoretical framework of Particle Filter(PF),the mean shift is introduced to redistribute random sample particles,in which particles move toward the maximal posterior kernel density estimation of target state,the weights of particle samples are updated as the mean shift iterative operating.Finally,to obtain the position of the target in next frame,we choose to use mean shift iterative operation and the relationship of moving target in the prior and current frame.Experimental results show that our algorithm obtains more accurate tracking results than traditional mean-shift and has better robustness for background updating and occlusion.

关 键 词：多目标跟踪 均值漂移 粒子滤波 目标遮挡 最大后验核密度估计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]