Retina-Like CCD Sensor for Active Vision

Abstract

The amount of information contained in images is often redundant for the execution of specific tasks. In order to limit the information to a quantity compatible with the computational capabilities of feasible processing systems, it is essential to avoid or discart at early stages (possibly at the sensor level) excessive detail. Research in the field of active vision and sensory fusion are developing new theories exploiting the advantages of sensory motor coordination.A fundamental problem for the reduction of visual data is to determine which part of image information is irrelevant before at least some pre-processing has been performed. In human vision, the structure of the visual sensor and the coordinated, task-driven, active control of eye movements have an essential role in the reduction of visual information. The distribution of the photosensitive elements in the human retina is such that the highest resolution is limited to the central part of the sensor (fovea) and decreases toward the periphery. The major advantage of this structure is the presence, on the same sensor, of a high-resolution sampling area and of a wide visual field. The exploitation of this structure in the context of different tasks is driven by the motor control strategies.The design of a space-variant vision system and the relation to dynamic vision is discussed. A prototype space-variant CCD visual sensor is the base of the system, coupled with dedicated driving hardware and a motor sub-system. The CCD sensor is characterized by a resolution decreasing from the center toward the periphery which resembles the sampling structure of the human retina. Several vision modules are presented which take advantage of the space-variant structure of the sensor and the dynamic capabilities of the vision system. The implications derived from the use of this space-variant sensor in relation to visual processing are discussed.