Omnidirectional Vision

University Of Pennsylvania, Philadelphia PA

Investigators

Abstract

Surround perception is crucial for an immersive sense of presence in communication and for efficient navigation and surveillance in robotics. To enable surround perception, new omnidirectional systems provide a new impetus for us to rethink the way images are acquired and analyzed. This project will be focused on two fundamental issues: the intrinsic geometric properties of catadioptric omnidirectional sensors and the space-variant signal analysis of omnidirectional images with non-uniform resolution. The PI has proven in the preliminary work that every catadioptric system can be shown to be equivalent to a generalized stereographic projection from a virtual sphere to the real image plane. He intends to develop a unifying framework for all catadioptric systems, where conventional perspective cameras will be just a special case, and exhaustively study invariants of the projections on reflective urfaces as well as efficient representations for reconstruction and motion estimation. Catadioptric designs need a new signal-theoretic treatment since shift-invariant processing does not apply in the non-uniform resolution of omnidirectional images. This research will make use of integral transforms that map the 2D signal to the spatial-frequency domain of a plane with virtually uniform resolution. The investigation will be extended from the simple edge detection to template matching and to the computation of optical flow in omnidirectional images.