Omnidirectional video cameras are becoming increasingly popular in computer vision. One family of these cameras uses a catadioptric system with a paraboloidal mirror and an orthographic lens to produce an omnidirectional image with a single center-of-projection. In this paper, we develop a novel calibration model that we combine with a beacon-based pose estimation algorithm. Our approach relaxes the assumption of an ideal paraboloidal catadioptric system and achieves an order of magnitude improvement in pose estimation accuracy compared to calibration with an ideal camera model. Our complete standalone system, placed on a radio-controlled motorized cart, moves in a room-size environment, capturing high-resolution frames to disk and recovering camera pose with an average error of 0.56% in a region 15 feet in diameter.