g Sensors: What Can be Done With a Non- gid Vision Platform?
Amnon Shashua · 1998
We describe the principles of building a moving vision platform (a Rig) that once calibrated can thereon selfadjust to changes in its internal configuration and maintain an Euclidean representation of the 30 world using only projective measurements. Formally, we address the question of how to obtain an invariant 30 projective representation from a dynamic collection of cameras. such as visual recognition by alignment, image transfer in Photogrammetry and image-based rendering in Computer Graphics in which a projective representation is sufficient (for example, [ 17, 2, 121). However, in mensuration, navigation and robot guidance tasks an Euclidean representation is necessary - and to that end a growing body of literature is dedicated to the problem of obtaining an Euclidean representation [7,4, 11, 22, 11. We divide the various approaches into the following classes: We show that the maximal generality is reached when the rig consists of 5 cameras whose center of projection remain $xed relative to each other during the motion. In other words, the non-rigid component motion may consist of change of internal parameters and relative camera orientations. The configuration reduces to 3 views when the Rig is built using a single physical camera with half-mirrors (beam-splitters) for creating 3 distinct views. We also briefly discuss 2-view configurations using hawmirrors and the principle behind adapting the conjguration to allow for zoom lenses in the system. The new research paradigm on non-rigid rigs (we term “Omni-Rig ’7 is applicable to the design of hion-based