Curve segmentation using directional information, relation to pattern detection

We propose an extension of the conformal (or geodesic) active contour framework in which the conformal factor depends not only on the position of the curve but also on the direction of its tangent. We describe several properties for variational curve segmentation schemes that justify the construction of optimal conformal factors (i.e., learning) in strong connection with pattern matching. The determination of optimal curves (i.e., segmentation) can be performed using either the calculus of variations or dynamic programming. The technique is illustrated on a road detection problem for different signal to noise ratios.