Acceleration-limited egomotion estimation

J Mendelsohn and E P Simoncelli

GRASP Laboratory, University of Pennsylvania, Technical Report 406, Aug 1996.
Presented is an iterative method for the estimation of egomotion from corresponding pairs of points in two consecutive movie frames. The method is based on a linearization of the discrete motion model equation relating rigid-body motion to image velocities. In particular, it makes no assumptions regarding the magnitudes of the translational and rotational motion parameters, but relies on these changing slowly (small accelerations). A more efficient variant is also presented, in which a secondary linearization is performed with respect to changes in the rigid-body rotation vector.

Both methods are compared with a standard linearization of the differential motion constraint. Comparisons are made under two different rotation magnitudes, two levels of additive noise, both perspective and equidistant projection models, and a random selection of translation directions. These comparisons demonstrate the inability of the differential model to capture large rotations and the efficiency with which the other two linearizations can recover large motions.

  • Listing of all publications