摘要
粒子的退化现象是粒子滤波算法中不容忽视的问题,在算法实现过程中需要考虑如何权衡粒子数和实时跟踪.提出一种结合粒子滤波和均值偏移2种算法的改进跟踪算法,只需要4个采样粒子,4个粒子通过均值偏移的方法收敛到局部最大值位置点,通过粒子加权的方式得到目标最终的位置.与传统的粒子滤波算法相比,该算法的计算复杂度大大降低,同时融合了结构信息这一特征,弥补了颜色信息的不足.由于文中算法的粒子总是能收敛到局部的极值,几乎不会出现退化现象,故无需进行重采样和粒子权值的更新.实验结果表明,该算法能在复杂背景下实现对目标的稳健跟踪,满足跟踪的实时性要求.
An inherent problem with particle filter is the so-called particle degeneracy, i. e. , now to balance the number of the particles and real-time tracking is a critical issue to consider. This paper presents an improved tracking algorithm by combining particle filter and mean shift. In our algorithm, only four particles are used and each one of them can converge to a local maximum. The final target position is a combination of weighted particles. The computational complexity decreased sharply in comparison with the conventional particle filtering. In addition, structural information is also employed to compensate for the insufficiency of pure color information. Because each particle can always converge to a local maximum, the particle degeneracy can rarely happen, as a result, the resampling and particle update steps are not used in our implementation. Experimental results show that our algorithm can robustly track the target even in cases where the color of tracked target is similar to that of the background, and it can also meet the demand of real-time tracking.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2008年第12期1583-1589,共7页
Journal of Computer-Aided Design & Computer Graphics
基金
四川省科技厅应用基础研究项目(2008ly0115-2)
四川省教育厅重点资助项目(2006A097)
国防基础科研项目(C1020060355)
关键词
粒子滤波
均值偏移
融合
结构信息
particle filter
mean-shift
fusion
structural information
