A novel hierarchical framework for object-based visual attention

Abstract

This paper proposes an artificial visual attention model which builds a saliency map associated to the sensed scene using a novel perception-based grouping process. This grouping mechanism is performed by a hierarchical irregular structure, and it takes into account colour contrast, edge and depth information. The resulting saliency map is composed by different parts or 'pre-attentive objects' which correspond to units of visual information that can be bound into a coherent and stable object. Besides, the ability to handle dynamic scenarios is included in the proposed model by introducing a tracking mechanism of moving objects, which is also performed using the same hierarchical structure. This allows to conduct the whole attention mechanism in the same structure, reducing the computational time. Experimental results show that the performance of the proposed model is compatible with the existing models of visual attention whereas the object-based nature of the proposed approach renders advantages of precise localization of the focus of attention and proper representation of the shapes of the attended 'pre-attentive objects'. © 2009 Springer Berlin Heidelberg.