Robot vision
Performance of optical flow techniques
International Journal of Computer Vision
Applied Geometry for Computer Graphics
Applied Geometry for Computer Graphics
| Hi-index | 0.00 |
Computational vision models that attempt to account for perception of depth from motion usually compute the optical flow field first. From the optical flow the ego-motion parameter are then estimated, if they are not already known from a motor reference. Finally the depth can be determined. The better the ego-motion parameters are known by extra-retinal information to be restricted to certain values before the optical flow is estimated, the more reliable is a depth-from-motion algorithm. We show here, that optical flow induced by translational motion mixed with specific rotational components can be dynamically mapped onto a head-centric frame such that it is invariant under these rotations. As a result, the spatial optical flow dimension are reduced from two to one, like purely translational flow. An earlier introduced optical flow algorithm that operates in close approximation of existing brain functionality gains with this preprocessing a much wider range of applications in which the motion of the observer is not restricted to pure translations.