Presenter

Ir. Thomas Geerinck - ETRO-VUB

Abstract

One of the monolithic goals of computer vision is to automatically interpret general digital images or videos of arbitrary scenes. However, the amount of visual information available to a vision system is enormous and in general it is computationally impossible to process all of this information bottom-up. To ensure that the process has tractable computational properties, visual attention plays a crucial role in terms of selection of visual information, allowing monitoring objects or regions of visual space and select information from them for report, recognition, etc.

This dissertation discusses one small but critical slice of a cognitive computer vision system, that of visual attention. In contrast to the attention mechanisms used in most previous machine vision systems, which drive attention based on the spatial location hypothesis, in this work we propose a novel model of object-based visual attention, in which the mechanisms which direct visual attention are object-driven. Considering the temporal dynamics associated with attention-dependent motion, an attention-based visual motion framework is also proposed. Finally, since a vision system will always have a set of tasks that defines the purpose of the visual process, a top-down approach is proposed to define the competition of the visual attention occurring not only within an object but also between objects, and illustrated in the framework of a surveillance system.