基于卡尔曼滤波器组的Mean Shift模板更新算法

针对Mean Shift算法缺乏必要的模板更新方法的缺陷,提出了一种基于卡尔曼滤波器组的Mean Shift模板更新算法。该算法首先将目标在特征空间中的特征值的概率作为模板信息;然后设计了一个滤波器组,其中每个滤波器用于估计特征子空间中一个子特征值概率的变化;最后将这些子特征值概率对应相乘就可以得到整个模板的更新值。由于滤波器的噪声参数是随着输入数据的变化随时动态确定的,因此,根据滤波器残差的变化就可以确定模板的更新策略。实验证明,该新算法不仅能够增强Mean Shift算法在目标姿态变化、光照变化下的跟踪效果,而且对阻挡时的鲁棒性也较好。

An Algorithm of Mean Shift Template Update Based a Group of Kalman Filters

To improve the limitation of Mean Shift lacks method of template update,an algorithm of template update based a group of Kalman filters is proposed.Probability of eigenvalue in feature space is taken as the template information.A group of filters are devised,where each filter is used to estimate the change of probability of sub-eigenvalue.All update value of template can be received by multiplying these corresponding probabilities in sub-feature space.The noise parameter of each filter would change with input data,so a novel strategy of template update according to change of residual of filters could be proposed.Experimental results show that the proposed algorithm can successfully track target under condition of changeable gesture of target and changeable illumination,and is robust to occlusion.