Estimating curvature by Kalman filters

A problem which occurs frequently in computer vision research is that of determining reasonable curvature estimates, and is especially difficult in applications where noisy or imperfect data are used [1, 2]. Thus, estimating curve or surface points by an “edge detector” is not an end in itself, but only the first step of image understanding. These points are often noisy and the estimate must be improved before proceeding to subsequent stages such as grouping. The problem addressed in this paper is that of refining the estimated points of a smooth curve which have been corrupted by noise. The method presented here is expected to be of general interest for the edge detection problem by delivering a coherent set of putative curve points along with estimates of normals and curvatures. Here, we restrict the discussion to curves, but note that the method generalizes immediately to surfaces [3].