A multi-scale approach to corner tracking

Abstract

This paper presents a multi-scale corner tracking algorithm based on a multi-scale corner detector. To extract corners from each frame of video sequences, the enhanced CSS corner detector using different scales of smoothing is applied. In the matching stage two-frame correspondence combined with three-frame based monitoring is considered. We monitor tracked corners from the third frame of input sequence in matching stage. Proposed three-frame monitoring helps to ensure that the number of tracked corners and their tracked positions among frames become more robust. Since the proposed corner tracker has enough robust corners based on multi-scale corner detector, it is practical and efficient. In matching stage, among similarity functions, standard cross-correlation, zero-mean cross-correlation, sum of squared differences, and x2-test measurements are tested. Well-known real video databases depicting translation, scaling, rotation and affine transformation with different lighting and different camera motions are used as experiments. All experiments confirm that the performance of the proposed tracker is high and reliable due to monitoring matched corners among frames.