Tomas Svoboda and Tomas Pajdla.

The whole
PostScript file compressed by `compress`.

*Abstract:*
Ego-motion detection and analysis provides useful tool for the
navigation of autonomous vehicle. The methods estimating the motion by
the analysis of epipolar geometry in image pairs often combine highly
robust estimation with nonlinear optimization to beat miscorrespondences
as well as Gaussian noise. However, these techniques are
computationally intensive and therefore not suitable for on line
ego-motion estimation on a moving vehicle.
We investigate shortcuts that lead to considerable reduction of
computation while attaining the robustness to miscorrespondences. The
main idea of this paper is to use {\em generate and test paradigm} in
treating miscorrespondences. We rank the correspondences according to
their the measure of their ``outlierness''. The measure suppresses those
correspondences which are not supported by others, hence likely wrong.
The subsets of the correspondences are drawn from the correspondences
with lowest outlierness. For each subset, the Essential Matrix, $E$, is
computed and the best subset is chosen to maximize the quality measure
of $E$ under restriction that the matrix is consistent with observed
data. The quality measure, $q_E$, is given by the difference of the two
largest singular values of $E$. Thanks to known calibration the
difference has to be zero for each ``true'' Essential Matrix.
In the experiments, presented in this paper, pure translations were
estimated using linear methods with error generally less than
\mbox{3 mm} from the scene in the distance about 1.8~m using 28
correspondences with 8 artificial miscorrespondences added i.e., with
almost 28\% contamination by outliers.