Combining Line and Point Correspondences for Homography Estimation

By Elan Dubrofsky

We present a method to extend the normalized direct linear transform (DLT) algorithm for homography estimation. Previously, only point correspondences were used. Now, line correspondences can be included as well. This extension is point-centric in that the lines are normalized to fit with the normalization used for points. Therefore, this method will be most useful if there are more point correspondences than line correspondences. Our main contribution is the derivation of an equation to normalize lines compatible with point normalization. We also show that using lines directly produces more accurate results than using their point intersections. Testing is provided both using a simple chessboard example and a real world example from the UBC hockey tracking system.

This is joint work with Robert J. Woodham.

Visit the LCI Forum page