Qualitative depth from monoscopic cues

Abstract

Depth perception is a vital feature for the survival of an animal. It is equally important for a mobile robot which has to navigate and avoid obstacles. Depth can be recovered from stereopsis, motion, focus or blur and other visual information. The paper presents an approach for obtaining relative depth from monocular sizes of objects in the retina (image plane) under size constancy. The method is shown quite robust since global features (length and area of objects) are used. The involved computation of depth is very simple. Another obvious merit of the proposed approach is that one needs not to know camera parameters and can estimate relative depth only in one image. Absolute depth can also be obtained from image sequence if egomotion is combined. >