基于PF的窗口自适应Mean-Shift跟踪算法

Mean-Shift tracking algorithm with adaptive window based on PF

在视频目标跟踪过程中,Mean-Shift算法存在着核函数带宽固定不变的缺陷,对尺度大小发生变化的目标无法进行有效跟踪。提出一种多尺度理论与粒子滤波器(PF)相结合的改进算法。通过粒子滤波器对多尺度理论统计得到的跟踪窗信息量进行预测修正,据此计算核窗宽大小变化的比例系数,实现跟踪算法的窗口自适应能力。实验结果表明,改进的跟踪算法对尺寸逐渐减小和逐渐增大的目标均能自动选择合适的跟踪窗口大小。

The deficiency of Mean-Shift algorithm in the video tracking is that the bandwidth of kernel function is fixed,so the algorithm can＇t have an effective tracking when distinct sale of the object changes. Aiming to this problem, a modified method in this paper is proposed that is multi-scale space theory combined with Particle Filter（PF）. The information in the tracking window is calculated by the multi-scale space theory. The information which is predicted and modified by the particle filter is introduced to get the proportion of the object area and the tracking algorithm is achieved to have adaptive window ability. The experimental results show that the improved algorithm could track the target efficiently in the scenarios that not only the object scale increases but the scale decreases as well.