Cooperative strategy for matching multi-level edge primitives

摘要

In this paper, we present a method for establishing the correspondence of edge primitives between two related image frames. We take into account different representations of edge primitives such as contour points, contour chains, closed curves and line segments. For the last three representations, a common term, metachain, is used. This term refers to any type of grouping of contour points. Since a contour point belongs to one metachain only after the segmentation of edge maps, a pair of matched contour points indicates a possible correspondence of the two metachains to which they belong. This observation leads us to develop a matching strategy which relies on the cooperation between the matching process of contour points and the matching process of metachains. This strategy involves three ordered processes. We first begin with the matching of contour points by using a similarity criterion. The obtained pairings of contour points will then be used to identify the correspondence of metachains by the use of a Hough-transform-like voting technique. Finally, knowing the matched metachains, we go back to rebuild the correspondence of contour points. Since a Hough transform technique is very noise resistant, our matching method does not require perfect results at the first step of matching. For this reason, we shall define a simple similarity function which does not use any local constraints for achieving this initial matching. The presented experimental results show that our method can be indifferently applied to solve both stereo and motion correspondence problem of edge primitives.