Cognitive Vision Systems: From Ideas to Specifications

Abstract

A computer vision system is expected to map an image (sequence) into an appropriate description of the world section in the field of view of the recording camera(s). The resulting description will be communicated usually to animate or inanimate recipients. Alternatively, it may be exploited to influence the depicted world section directly via actuators coupled to the computer vision system. Human expectations with respect to what constitutes an ‘appropriate description’ expand concurrently with the processing and storage capacities of computers. As a consequence, the algorithmic processes incorporated into a computer vision system become ever more complex and the diversity of computer vision systems expands continuously.

This contribution addresses engineering aspects of computer vision systems. A structure is suggested for the processing steps devoted to the extraction of spatiotemporally local signal characterisations, their aggregation into non-local image descriptors, and the construction of spatiotemporal geometric descriptions in the scene domain. It is postulated that the association of these signal and geometric descriptions with conceptual descriptions facilitates the design and implementation of versatile and encompassing computer vision systems. In view of the complexity encountered in the course of such an endeavor, a set of aspects is proposed whose consideration should help to clarify the design options and to enforce precise specifications as a precondition for a viable system design and evaluation.