On the accuracy of point curvature estimators in a discrete environment

摘要

Points of maximum curvature are commonly believed to be the most perceptually significant points on digital curves, and as such have been used as shape features in both 2D and 3D object recognition by many researchers. We have experimented with several different measures of digital curvature, and have found them to be unreliable in the presence of noise. We present the results of a series of experiments using geometric shapes which illustrate the relative performance of three different estimators of point curvature in varying degrees of noise. We make recommendations as to which estimator might be expected to achieve best results under given conditions. These results are extended and applied to real object boundaries.