Multifoveal Imager for Stereo Applications

摘要

A multiresolution imager based on adaptive retinal structures, with data compressions above 85%, is presented in this article. The main goal of the imager platform is to speed up image processing with the use of selective data reductions to shorten the vision systems' tasks in stereo applications. Implemented on a field-programmable gate array, the platform can be configured as the front end of active vision systems, with use of adaptive foveal sensing on uniresolution images to cover wide fields of view, being also a development tool for multiresolution applications with different image formats and interfaces. The multifoveal imager provides the hierarchical data structures related to multiresolution levels, following instructions to control sensor parameters or to perform adaptive fovea fixations in real time, adapting its operation to the constraints of the active vision systems. It also uses intermediate resolution data to implement in hardware an efficient background extractor to cooperate with image processors in motion detection tasks and attention mechanisms. Some platform configurations are explained and experimental results are discussed in relation to the advantages of the adaptive retinal structures.