摘要
针对序贯蒙特卡罗(Sequential Monte Carlo,MC)算法存在的计算量大的缺点,提出了一种新的Mean Shift MonteCarlo(MSMC)目标跟踪算法。在传统的MC算法中采取Mean Shift这种梯度最优下降法来寻找局部最大样本值,以目标的颜色特征建立目标状态空间模型,并用Bhattacharyya系数作为评价函数给出样本的权值。算法以少于300个样本(实验用200个样本)来保持对目标运动状态预测的多样性,有效地克服了MC算法收敛速度较慢的弱点,将算法的计算时间从76 ms/frame降低到了35 ms/frame(跟踪窗口为28 pixel×26 pixel)。实验结果表明,提出的算法能够在发生遮挡的情况下实现较稳定的目标跟踪,使算法应用于实际工程成为可能。
A new Sequential Monte Carlo(MC) algorithm, Mean Shift Monte Carlo (MSMC) algorithm, is proposed for visual tracking in image sequences. The MSMC takes Mean Shift to converge the samples with smaller weight to look for the local maximum ones. A state space model of the object is established using the color cue, and the weights of samples are given by adopting the Bhattacharyya coefficients as the evaluation funtion. The MSMC algorithm can maintain the diversity less than 300 samples (200 samples arc used in experiments), so the consumed time can be decreased from 76 ms/ frame to 35 ms/frame (the tracking window is 28 pixel× 26 pixcl). The experimental results in real video data show that the algorithm can track objects stably in the case of obstruction and it is possible to be used in the practical projects.
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2008年第1期122-127,共6页
Optics and Precision Engineering
基金
中国科学院二期创新基金资助项目(No.C05T022)
关键词
目标跟踪
序贯蒙特卡罗算法
Mean
SHIFT
MC
局部最优
visual tracking
sequential Monte Carlo (MC) algorithm
Mean Shift Monte Carlo(MSMC) : local maximum