Robust estimation of target scale by removing outlier motion vectors using MAD

Abstract

A robust estimation method for target scale using median absolute deviation (MAD) for reducing outlier motion vectors is proposed. To remove outlier motion vectors, the proposed method first uses normalised cross-correlation (NCC) and forward-backward (FB) filters. Then, the MAD filter removes the remaining outlier motion vectors. After filtering, inlier motion vectors are used to estimate a scale change ratio of the target. The estimated scale change ratio is evaluated by the MAD of scale change ratios to guarantee reliability of estimation.