基于核函数粒子滤波和多特征自适应融合的目标跟踪

Kernel-Based Particle Filter for Target Tracking with Adaptive Multiple Features Fusion

经典粒子滤波及其改进算法在观测模型与真实情况存在偏差时会导致滤波发散,针对这一问题,提出一种核函数粒子滤波算法.该算法根据目标状态与粒子状态之间的距离,利用核函数产生权值对粒子进行二次加权,根据粒子的二次加权结果进行粒子重采样;以改进的粒子滤波算法为框架,提出了一种自适应多特征融合目标跟踪方法,利用相似性度量动态地评价特征对目标与背景的区分能力,并自适应地计算特征融合权重,以适应目标跟踪过程中目标与背景的变化,提高目标跟踪的鲁棒性.实验结果表明,文中提出的目标跟踪方法比经典粒子滤波目标跟踪方法具有更强的抗干扰性能和较高的跟踪精度.

The particle filtering has been extensively used for visual tracking due to its flexibility.However the conventional particle filtering and its improved variants usually diverge when the measurement model is not accurate enough. To address this problem, a kernel-based particle filter algorithm is proposed. The algorithm reweighs the particles by weights which are produced by kernel function with the distance between target state and particles, and the particles are resampled according to the resultant weights. With the above improved particle filter algorithm, an adaptive multiple features fusion target tracking method is proposed. The proposed tracking method dynamically assesses the discriminability of each feature with respect to foreground to background separability and adaptively computes the feature＇s fusion weight by some similarity measure. Experimental results show that the proposed tracking method is superior over the conventional particle filter based tracking methods.