Fixation Simplifies 3D Motion Estimation

摘要

Fixation is defined as holding the gaze direction towards the same environmental point through time. It was proven in the past that fixation reduces the number of unknowns in passive visual navigation from five to four. In this paper, we show that fixation further simplifies 3D motion estimation because it decouples the motion parameter space. We project the spherical motion field in two latitudinal directions with respect to two different poles of the image sphere. The first projection enables the computation of the longitude of the translation direction and the torsion. The second projection gives the angle between the direction of fixation and translation, as well as the time to collision to the fixated scene point. Both computational steps are based on one-dimensional searches along meridians of the image sphere. The observer may move with all six degrees of freedom. We do not use the efference copy of the fixational rotation of the camera. Performance of the algorithm is tested on real world sequences with fixation accomplished either off-line or during the recording using an active camera. A comprehensive review of most theories on advantages of fixation clarifies the differences to our approach.