基于粒子滤波的多特征融合目标跟踪算法

针对复杂背景下单一特征目标跟踪鲁棒性不强的问题，提出了一种基于粒子滤波框架下的分级多特征融合的视频目标跟踪算法。在粒子滤波算法的框架下，选取颜色特征作为观测量进行一级滤波来构造粒子滤波器的提议分布，使得粒子集紧密分布在真实目标状态附近，通过对设计的提议分布进行重要性采样，并结合边缘分布特征进行二次滤波来获得跟踪目标状态的后验概率密度估计。为了克服模板漂移所引起的误跟踪，采用自适应模板更新策略对目标进行跟踪。实验结果表明，相对于单一特征的视频目标跟踪算法，本文所提出的算法可以有效地避免遮挡、姿态改变以及目标发生非平面旋转等对跟踪的影响，在复杂背景下的视颍目标跟踪具有较强的鲁棒性。

In view of using single feature may lead to poor robustness in tracking process under complex background, we proposed a new visual target tracking arithmetic with fusing color and edge feature by levels in particle filter (PF) frame work. The state of target was estimated approximately in frame of PF, choosing color feature or edge feature as observed value, constructing proposed distribution of the first level PF to let the particle set distributed closely around the real target state. By importance sampling on the proposed distribution, combined with the marginal feature, we got the posterior probability distribution of the second level PF. In order to overcome mistake-tracking of template drift, adaptive template update mechanism was used to tracking target. Experiment results show the proposed method, comparing with the single characteristics video object tracking, can effectively avoid the influence of blocking, attitude change and non-plane-rotation, having stronger robustness in complex sense. © 2012 Chinese Assoc of Automati.