Curve Registration

Several algorithms adapted from computer vision have been proposed and used over time. In [13], Gueziec matches the crest lines using a combination of geometric hashing [23] and Hough transform (see for instance [22]). The basic idea was to index each point of each model crest line in a hash-table using its invariant characteristics. At recognition time, the invariants of each scene crest line point are used to recover, thanks to the hash table, the possible matches with model points (geometric hashing step). For each match (i.e., couple of model and scene points), a rigid transformation is computed by superimposing the Frenet frames and used to index the match in a new accumulator sampling the rigid transformation space (Hough transform step). Densely populated cells in this second accumulator are detected as rigid body transformations that are candidates to match a large number of crest points. For each such cell, a refined least-squares transformation is computed using the matches indexed in this cell.

0 0

Post a comment