Many of the known visual systems in nature are characterized by a wide field of view allowing animals to keep the whole surrounding environment under control. In this sense, dragonflies are one of the best examples: their compound eyes are made up of thousands of separate light-sensing organs arranged to give nearly a 360° field of vision. However, animals with eyes on the sides of their head have high periscopy but low binocularity, that is their views overlap very little. Differently, raptors’ eyes have a central part that permits them to see far away details with an impressive resolution and their views overlap by about ninety degrees. Those characteristics allow for a globally wide field of view and for accurate stereoscopic vision at the same time, which in turn allows for determination of distance, leading to the ability to develop a sharp, three-dimensional image of a large portion of their view. In mobile robotics applications, autonomous robots are required to react to visual stimuli that may come from any direction at any moment of their activity. In surveillance applications, the opportunity to obtain a field of view as wide as possible is also a critical requirement. For these reasons, a growing interest in omnidirectional vision systems (Benosman 2001), which is still a particularly intriguing research field, has emerged. On the other hand, requirements to be able to carry out object/pattern recognition and classification tasks are opposite, high resolution and accuracy and low distortion being possibly the most important ones. Finally, three-dimensional information extraction can be usually achieved by vision systems that combine the use of at least two sensors at the same time. This article presents the class of hybrid dual camera vision systems. This kind of sensors, inspired by existing visual systems in nature, combines an omnidirectional sensor with a perspective moving camera. In this way it is possible to observe the whole surrounding scene at low resolution, while, at the same time, the perspective camera can be directed to focus on objects of interest with higher resolution.
There are essentially two ways to observe a very wide area. It is possible to use many cameras pointed on non-overlapping areas or, conversely, a single camera with a wide field of view. In the former case, the amount of data to be analyzed is much bigger than in the latter one. In addition, calibration and synchronization problems for the camera network have to be faced. On the other hand, in the second approach the system is cheaper, easy to calibrate, while the analysis of a single image is straightforward. In this case, however, the disadvantage is a loss of resolution at which objects details are seen, since a wider field of view is projected onto the same area of the video sensor and thus described with the same amount of pixel as for a normal one. This was clear since the mid 1990s with the earlier experiments with omnidirectional vision systems. Consequently a number of studies on omnidirectional sensors “enriched” with at least one second source of environmental data arose to achieve wide fields of view without loss of resolution. For example some work, oriented to robotics applications, has dealt with a catadioptric camera working in conjunction with a laser scanner as, to cite only few recent, in (Kobilarov 2006; Mei 2006). More surveillance application-oriented work has involved multi-camera systems, joining omnidirectional and traditional cameras, while other work dealt with geometric aspects of hybrid stereo/multi-view relations, as in (Sturm 2002; Chen 2003).
The natural choice to develop a cheap vision system with both omni-sight and high-detail resolution is to couple an omnidirectional camera with a moving traditional camera. In the sequel, we will focus on this kind of systems that are usually called “hybrid dual camera systems”.
Key Terms in this Chapter
Omnidirectional Camera: A camera able to see in all directions. There are essentially two different methods to obtain a very wide field of view: the older one involves the use of a special type of lens, usually referred to as fisheye lens, while the other one uses in conjunction rectilinear lenses and mirrors. Lenses obtained in the latter case are usually called catadioptric lenses and the camera-lens ensemble is referred to as catadioptric camera
PTZ Camera: A camera able to pan left and right, tilt up and down, and zoom. It is usually possible to freely control its orientation and zooming status at a distance through a computer or a dedicated control system
Catadioptric Camera: A camera that uses in conjunction catoptric, reflective, lenses (mirrors) and dioptric, refractive, lenses. Usually the purpose of these cameras is to achieve a wider field of view than the one obtained by classical lenses. Even if the field of view of a lens could be improved with any convex surface mirror, those of greater interest are conic, spherical, parabolic and hyperbolic-shaped ones
Stereo Vision: A visual perception process that exploits two different views to achieve depth perception. The difference between the two images, usually referred to as binocular disparity, is interpreted by the brain (or by an artificial intelligent system) as depth
Single Viewpoint Constraint: To obtain an image with a non-distorted metric content, it is essential that all incoming principal light rays of a lens intersect at a single point. In this case a fixed viewpoint is obtained and all the information contained in an image is seen from this point
Central Catadioptric Camera: A camera that combines lenses and mirrors to capture a wide field of view through a central projection (i.e. a single viewpoint). Most common examples use paraboloidal or hyperboloidal mirrors. In the former case a telecentric lens is needed to focalize parallel rays reflected by the mirror and there are no constraints for mirror to camera relative positioning: the internal focus of the parabola acts as the unique viewpoint in the latter case it is possible to use a normal lens, but mirror to camera positioning is critical for achieving a single viewpoint it is essential that the principal point of the lens coincides with the external focus of the hyperboloid to let the internal one be the unique viewpoint for the observed scene
Camera Calibration: A procedure used to obtain geometrical information about image formation in a specific camera. After calibration, it is possible to relate metric distances on the image to distances in the real word. In any case only one image is not enough to reconstruct the third dimension and some a priori information is needed to accomplish this capability