At the Computer Vision Laboratory at Columbia University, a complete complete class of omnidirectional sensors have been derived that use mirrors to enhance the field of view. Each sensing unit, called OMNICAMERA, is comprised of an off-the-shelf video camera, an off-the-shelf lens, and a single curved mirror. The curved mirror is chosen to guarantee a single effective center of projection for the complete sensing unit. This allows the construction of distortion-free images for any user selected portion of the acquired omnidirectional image.
Each OMNICAMERA sensing unit is capable of imaging a complete hemispherical field of view. The curved mirror used to accomplish this is hyperbolic when the imaging projection used is perspective, and parabolic when the imaging projection used is orthographic. The prototype sensing units we have developed use a telecentric lens to achieve orthographic projection and a parabolic mirror to achieve a hemispherical field of view. This configuration offers calibration advantages over one that uses a hyperbolic mirror.
Two sensing units can be placed back-to-back such that the foci of the two parabolic mirrors coincide. The result is a true omnidirectional sensor that has a single center of projection and a field of view that corresponds to the entire sphere! The single center of projection allows software creation of pure perspective images (viewing direction, image size, and focal length chosen by the user) from the two parabolic images.
The reasons for which other approaches to omnidirectional imaging (for instance, using a fish-eye lens) were avoided are described in other approaches.
Several prototypes of the OMNICAMERA sensor have been developed. The prototypes are geared towards a variety of applications, including, video surveillance, autonomous navigation, and teleconferencing. Each prototype is a true video rate (30 Hz) omnidirectional sensor that allows a remote user to interact with a dynamic scene. A software module, called OMNIVIDEO, allows the user to visually navigate through scene as it changes as well as generate multiple video-rate views of the scene.
Back to Omnicamera.