基于粒子滤波的on-line boosting目标跟踪算法

Object Tracking Algorithm of On-line Boosting Based on Particle Filter

基于Haar-like特征的on-line boosting跟踪算法(HBT)把目标跟踪看作是目标与背景的二分类问题,通过在候选区域搜索最大分类置信度的方法得到目标新的位置。但在获取最大置信度时选用的是区域穷举搜索法,当目标过大或者运动速度过快时,很难确保系统的实时性,且易造成跟踪丢失。本文将粒子滤波算法引入HBT目标跟踪框架中,通过建立目标运动模型,并把HBT目标分类置信度与粒子滤波的观测模型结合起来,提出了基于粒子滤波的on-line boosting目标跟踪算法(PFHBT)。与HBT算法相比,本文算法不仅加快了计算速度,而且很好地解决了目标速度过快造成跟踪丢失的问题,保证了系统的实时性和鲁棒性。

Object tracking is regarded as a classification between object and background in on-line boost-ing tracking algorithm （HBT） based on the Haar-like feature. The new position of the object is obtained by searching the maximum classification confidence in the candidate region. However, the exhaustive search procedure makes it difficult to ensure real-time property and result in tracking lost easily when the size of the object is too big or the speed of the object is too fast to get the maximum confidence in the candidate regions. In this paper,the particle filter is introduced into the HBT object tracking framework and an algorithm of on-line boosting object tracking based on particle filter （PFHBT） is proposed：the motion of the object is modeled and the object classification confidence is regarded as the observation of particle filter. Experimental results show that the algorithm not only improves the computing speed sig-nificantly, but also solves the problem of tracking lost caused by object fast moving effectively.