This paper considers the problem of scene recognition for an attentive robot as is exploring its environment and thus generating a sequence of foveae. Previous work has shown that attentional sequences - spatio-temporally related sequence of observations obtained from the foveal sequence using a rich set of biologically motivated visual primitives - can be used for scene recognition tasks. For these applications, information contained within the attentional sequence must be integrated dynamically across individual visual primitive responses, time and scene models. We present two integration methods varying in the order of integration with respect to these three dimensions. A series of comparative experiments serve to demonstrate that these methods exhibit robust real-time performance in relatively complex scenes with simple models and few saccades.
展开▼
