快速运动目标的mean shift跟踪算法 algorithm for tracking of fast motion objects with mean shift.pdfVIP

33 5 Vol.33, No.5 2006 5 Opto-Electronic Engineering May, 2006 文章编号 1003501X(2006 快速运动目标的 Mean shift 跟踪算法 ( 浙江大学 电气学院 浙江 杭州 310027 ) 摘要 Mean shift Mean shift Mean shift Mean shift 关键词 Mean shift 中图分类号 V556 文献标识码 A Algorithm for tracking of fast motion objects with Mean shift ZHU Sheng-li ZHU Shan-an LI Xu-chao ( College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China ) Abstract To improve theoretic limitation of Mean shift, an algorithm for tracking of fast motion objects, which combines Mean shift and Kalman filter, is proposed. At first, the starting position of Mean shift is found with Kalman filter in every frame, and then Mean shift is utilized to track the target position. When severe occlusion appears, filtering residuals is exploited to decide whether the Kalman filter works. At this moment, Kalman filter is replaced by linear prediction of object position. Experimental results show that the proposed algorithm can track fast moving objects successfully and have better robust for occlusion. Key words Mean shift; Kernel function; Kalman filter; Target tracking 引 言 Mean shift 是一种密度梯度的无参估计方法 于 1975 年由 Fukunaga[1] 提出 1995 年 Cheng 的文章[2] 将 [3 5] 它引入计算机视觉领域 直到近几年才引起国外学者们的广泛兴趣 Chmaniciu 和 Meer - 对 Mean shift 在图像滤波 分割和跟踪中的使用都做了较早的论述 Changjiang Yang[6] 对多维图像的 Mean shift 方法进 行了讨论 使用改进的快速高斯变换提高算法的速