Target tracking by matching a shape represented by a tree of sigmoid functions

Abstract

This report describes a shape matching method which transforms a template shape represented by a tree of sigmoid functions so that it fits target shapes included in a sequence of image frames. The tree-based shape representation is automatically constructed by a training procedure in advance to the shape matching process. Through the matching process, the position and the posture of the target are computed by a gradient-descent-based searching procedure. The amount of computation is reduced by using a compact tree, which represents a shape roughly but with sufficient details, for determining the object postures. The search area is also limited assuming small changes of the object positions and postures in consecutive frames. The number of iterations in gradient-descent-based search is reduced using the previous position and posture as initial conditions for the next search. Some experiments were conducted and the method's sensitivity to noises and initial parameter values were examined. The results suggest the method's feasibility as a target tracking method.