Improved Kanade-Lucas-Tomasi tracker for images with scale changes

Abstract

To match two images with a large scale difference, a scale parameter can be introduced into the warp parameters of the KLT tracker. For some images, the KLT tracker with the scale warp parameter fails to converge. We assume that this result is caused by the singularity of the Hessian matrix. An improved KLT tracker is proposed to avoid this tracking failure. The proposed method introduces an index (scale invariance index) to determine how close the Hessian matrix is to a singular one with the scale warp parameter. According to the index, either of two different sets of warp parameters is selected: one is with the scale warp parameter and the other is without it.