Active Egomotion Estimation: A Qualitative Approach

Abstract

Passive navigation refers to the ability of an organism or a robot that moves in its environment to determine its own motion precisely on the basis of some perceptual input, for the purposes of kinetic stabilization. The problem has been treated, for the most part, as a general recovery from dynamic imagery problem, and it has been formulated as the general 3-D motion estimation (or structure from motion) module. Consequently, if a robust solution to the passive navigation problem—as it has been formulated in the recovery paradigm—is achieved, we will immediately be able to solve many other important problems, as simple applications of the general principle. However, despite numerous theoretical results, no technique has found applications in systems that can perform well in the real world. In this paper, we outline some of the reasons behind this and we develop a robust solution to the passive navigation problem which is purposive, in the sense that it does not claim any generality. It just solves the kinetic stabilization problem and cannot be used as it is for other problems related to 3-D motion. qualitative, in the sense that the solution comes as the answer to a series of simple yes/no questions and not as the result of complicated numerical processing. active, in the sense that the activity of the observer (in this case “saccades”) is essential for the solution of the problem. The input to the perceptual process of kinetic stabilization that we have developed is the normal flow, i.e. the projection of the optic flow along the direction of the image gradient. Contributions of this work are the fact that translation can be estimated reliably from a normal flow field that also contains rotation.