Fast-moving target tracking based on mean shift and frame-difference methods

Abstract

Abstract: The mean shift tracker has difficulty in tracking fastmoving targets and suffers from tracking error accumulation prob-lem. To overcome the limitations of the mean shift method, a newapproach is proposed by integrating the mean shift algorithm andframe-difference methods. The rough position of the moving tar-get is first located by the direct frame-difference algorithm andthree-frame-difference algorithm for the immobile camera scenesand mobile camera scenes, respectively. Then, the mean shiftalgorithm is used to achieve precise tracking of the target. Severaltracking experiments show that the proposed method can effec-tively track first moving targets and overcome the tracking erroraccumulation problem. Keywords: mean shift, frame-difference method, target tracking,computer vision. DOI: 10.3969/j.iss.1004-4132.2011.04.006 1. Introduction Mean shift is a nonparametric kernel density estima-tor, which is based on the color kernel density appear-ance change. Fukunaga et al. [1] first proposed themean shift algorithm for clustering data in 1975. LaterCheng [2] introduced it to the image-processing com-munity in 1995. A remarkable work on mean shift al-gorithm for image segmentation and tracking is done in[3–7]. Later, many variants of the mean shift algorithmwereproposedfordifferentapplications[8–12]. Themeanshift tracker performswell with relatively low-speed mov-ing targets. However, its performance is not guaran-teed for fast-moving targets [6,9,10,13–15]. Because themean shift algorithm is an iterative algorithm, it suffersfromthetrackingerroraccumulationproblemandlacksanerror-eliminating mechanism. The Kalman filter has been