一种基于粒子滤波的自适应相关跟踪算法

Adaptive Correlation Tracking Algorithm Based on Particle Filter

相关跟踪是最常见的一种目标跟踪方法,但传统相关跟踪采取的"峰值"跟踪方法抛弃了所有小于峰值点相关值的位置点的信息,不够稳健,受遮挡影响大,并且很难求解相关模板的仿射变形参数。提出了一种改进的非线性相关跟踪算法,以改进的灰度模板作为目标表示方式,粒子的权值与相关值成比例,目标状态的后验概率由粒子加权表示。模板更新时根据粒子权值进行自适应调节,对所有粒子所在位置的区域进行加权更新,权值大的粒子具有更高的更新系数,避免了仅利用单一峰值点处的模板进行更新可能造成的误差累计。该算法大大提高了跟踪与模板更新的鲁棒性,同时也是一种在仿射空间进行运动参数搜索的实用方法。

Correlation tracking is the most commonly used method for visual tracking. The traditional ＂ peak - tracking＂ based method discarded the information of other points smaller than the peak value, which was not robust enough to disturbance. In addition; it was difficult to solve affine parameters. An improved nonlinear correlation tracking algorithm was proposed . The improved grey - level template is used for describing the objective. The weights of the particles are proportion to the correlation value. The posterior PDF （probability density function） of the object can be denoted by the sum of the weighted particles. The template is updated adaptively according to the old template and all of the particles. The bigger particles have higher updating weights. The method is practical and robust as a result of muhimodality. And it is also useful for solving affine parameters.