| When the iris of a conventional camera is replaced by a mask with multiple apertures, a composite image is formed. Unlike binocular stereopsis, the views from each aperture are superimposed, so that conventional methods in stereo vision do not apply. Still, the local displacement between corresponding points in these views is related to their distance from the camera. This depth cue provides the basis for a new paradigm in passive range sensing--monocular stereopsis. This thesis presents a technique for computing a dense range image from one composite image acquired with a multiple aperture camera. The formation of the composite image is modelled as an echo process, where the depth of a point in the scene is directly related to the spatial delay of its visual echo. Cepstral analysis is the method used to detect this echo. A model of the composite image cepstrum allows measurement of monocular disparity to subpixel precision, as well as an estimate of its associated error distribution. This data, computed over a dense grid, is used to generate a piecewise planar representation of surfaces in the scene, based on a maximum likelihood criterion. Borrowing techniques from visual psychophysics, the spatial resolution of this result is evaluated in terms of an intelligent agent making decisions about its environment. This new range imaging technique is successfully applied to real-world scenes to demonstrate its potential for mobile robot navigation and obstacle avoidance. |
