摘要
针对粒子滤波跟踪过程中不精确的状态模型或观测模型会降低跟踪精度的问题,提出一种基于粒子滤波与在线随机森林分类的目标跟踪算法框架.通过在线样本学习,随机森林中的样本集可以准确地近似目标外观的概率分布;在粒子滤波跟踪中,采用随机森林分类结果及区域直方图相似度来估计粒子相似度,从而提高了观测模型的精度.当出现跟踪漂移时,通过随机森林检测目标来重新初始化粒子滤波器,可以防止由于误差积累而造成的跟踪失败.采用vc 6.0+opencv实现了本算法,并设计两类试验分别来验证算法的跟踪精度和抗漂移能力.结果表明,该算法跟踪正确率比粒子滤波提高23%,比随机森林提高16%,因此可以防止无规则运动等因素造成的跟踪漂移,实现了长序列可靠跟踪.
To avoid the performance degradation of the tracker caused by the inaccurate prediction of state model and inaccurate observation in particle filter, a new framework was proposed based on particle filter and online random forest for object tracking. The probability density of the object appearance was accurately approximated by the sample set which was collected by online learning method. In particle fil- ter scheme, the particle likelihood was measured by the combination of classified result of online random forest and region histogram likelihood to improve the accuracy of the observation model. While track drift occurred, the particle filter was re-initialized by the detection of random forest to prevent track failure caused by the accumulated error. The algorithm was realized by vc 6.0 ＋ opencv, and two experiments were designed to verify the tracking precision and the ability of drift resistance. The results show that the proposed algorithm can increase the ratio of correct tracking to 91% , while those are 68% and 75% for the particle filter and the random forest, respectively. The proposed approach can prevent track drift to achieve robust long sequences tracking.
出处
《江苏大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2014年第2期207-213,共7页
Journal of Jiangsu University：Natural Science Edition
基金
国家自然科学基金资助项目(61100139
61040009)
关键词
粒子滤波
随机森林
在线学习
运动跟踪
观测模型
particle filter
random forest
online learning
object tracking
observation model
