Recovering 3-D egomotion parameters from optic flow: from structural principles to analog architectures

Abstract

A computational architecture for direct estimation of heading direction from optic flow is proposed. By representing optic flow information as undulatory spatiotemporal patterns, global first-order differential analysis of the optic flow is reconducted to spatiotemporal filtering operations with local kernels characterized with respect to the fixation point(gaze). These operations can be mapped with high efficiency on analog architectures based on structured arrays of cells reacting collectively to spatiotemporal input stimuli. By observing the overall distribution of cells' activity, it is possible to gain information on the orientation of the heading in the retinal reference frame and on the angle between the heading and the gaze directions.