Fundamental study of moving object tracking using local moments

Abstract

In multimodal communication using visual data, a mechanism for automatically tracking the head of a user is necessary in order to provide an environment in which the user can focus on communication itself without needing to worry about the position and size of his own face or other obstacles in the visual field of the camera. In this paper, a method for such target tracking is proposed using observations based on local moments, with observation parameters regulated through α-β filters. The main features of the proposed method may be described as follows. Local moments employed as observed values are calculated for local areas of the image corresponding to the viewpoint and visual field. Due to the use of local moments, only local calculations are performed, which reduces the overall amount of computation and suppresses the influence of other objects or noise. In the problematic cases in which the target object approaches an obstacle, a mechanism is provided to recognize this fact and to avoid the obstacles. In particular, two alternative models are provided, one allowing for nearby obstacles and the other disregarding this possibility. One of the models is selected on the basis of a certain criterion, to be used for estimation of the target position. Experiments with model images and actual images confirmed the effectiveness of obstacle proximity recognition, and of the employment of alternative models, thus demonstrating that the proposed system ensures steady target tracking. © 1998 Scripta Technica, Electron Comm Jpn Pt 3, 82(3): 82–90, 1999