Bayesian visual feature integration with saccadic eye movements

Abstract

In order to allow humanoid robots to operate in unstructured environments, behaviors have to be implemented that support exploration of the environment. In active visual perception, saccadic eye movement is such a behavior that supports the exploration of salient locations within the current scene in a sequential manner. The proposed work deals with the integration of visual features extracted at different gazes during saccades executed on an active humanoid head. Using probabilistic methods to account for uncertainties during execution and perception, visual stimuli are integrated in an ego-centric representation. The resulting map stores the regarded stimuli in a consistent fashion. The approach is evaluated using three common types of feature extraction methods.