Abstract-Multimodal attention is a key requirement for humanoid robots in order to navigate in complex environments and act as social, cognitive human partners. To this end, robots have to incorporate attention mechanisms that focus the processing on the potentially most relevant stimuli while controlling the sensor orientation to improve the perception of these stimuli. In this paper, we present our implementation of audio-visual saliency-based attention that we integrated in a system for knowledge-driven audio-visual scene analysis and object-based world modeling. For this purpose, we introduce a novel isophote-based method for proto-object segmentation of saliency maps, a surprise-based auditory saliency definition, and a parametric 3-D model for multimodal saliency fusion. The applicability of the proposed system is demonstrated in a series of experiments.Index Terms-audio-visual saliency, auditory surprise, isophote-based visual proto-objects, parametric 3-D saliency model, object-based inhibition of return, multimodal attention, scene exploration, hierarchical object analysis, overt attention, active perception

Diagnosis applications relying on Artificial Intelligence methods must deal with uncertain knowledge and fuzzy input data. Besides other methods, classical probability theory has been realized by many authors to be useful for such tasks, given that tools are available, making its application more handy. One such tool are Bayesian Networks (causal probabilistic networks, belief networks) [6].While the construction of the graph of a Bayesian Network along causal dependencies is often quite easy, it is usually difficult to specify the necessary probabilities, i.e. for every node the probabilities that its variable assumes a certain value given the predecessors' values. Instead, in practical applications, it is usually preferable to specify an unordered set of conditional probabilities, which does not necessarily match the set of probabilities in the network.Such a set of probabilities has the potential problem of being neither consistent nor complete with respect to the compound distribution of all random variables in the net. It is shown, that a test for consistency and completeness and answering queries about the compound probability can in general be done by solving a nonlinear equation system. However, due to the inherent complexity of solving such a system, this is in general not feasible. But some interesting special cases are presented for which an approximation yields useful results.