Iterative Matching of 3-D Profile Maps

Olli Jokinen
Institute of Photogrammetry and Remote Sensing
Helsinki University of Technology
P.O. Box 1200, FIN-02015 HUT, Finland
E-mail: Olli.Jokinen@@hut.fi

Date:

Keywords: light striping, registration, self-calibration, image algebra

MOS(AMS) Subject Classifications: 62J02, 62F35, 62H25, 65Y05

Abstract:

We present a method for the simultaneous matching of multiple sets of 3-D points when the sets can be represented as overlapping parametric surfaces such as profile maps acquired by light striping and when no exact corresponding points between the sets are known. The method is intended for solving the rigid body transformations between data sets measured from different viewpoints and for refining the calibration of the light striping system involving the projective transformation between the image and laser planes and the direction of scanning with respect to the laser plane.

The method iterates two steps, i.e., finding the corresponding points on the parametric domains of the surfaces and updating the unknown parameters so that the mean of the squares of weighted distances between the corresponding points is minimized. We apply the Levenberg-Marquardt algorithm and use an adaptive weighting to reject incompatible correspondences in regard to the direction of the surface normal and the distance between the corresponding points. The matching is performed only on smooth areas and the precision of the data is also included in the weighting.

The accuracy of the method is evaluated using synthetic data and the error caused by interpolation when determining the corresponding points is discussed. The precision of the solution is estimated applying first order error propagation techniques. The method is formulated using the notation of an image algebra and implemented using the Matlab software in a parallel way and proceeding from low to high resolution. An experimental testing shows that the method gives more accurate results in shorter computing time than if the corresponding points were searched as closest ones in 3-D.