Pre-attentive filtering for robot perception of patterned activity

Why: Real-time machine perception benefits greatly from heuristics for quickly filtering out irrelevant stimuli and thus focusing computational effort where it is most likely to pay off. The robots built by the Humanoid Robotics Group all use one form or another of such heuristics for visual perception, such as biases towards skin-colored regions, moving objects, and bright stimuli [2]. More recently, we have investigated the utility of periodicity as a perceptual bias, demonstrating cross-modal priming where visually periodic motion influenced the perception of the sound of tools and toys [1]. But are such biases limited to low-level perceptual features? The goal of this work is to apply similar filtering to the detection of patterned activity, both to facilitate autonomous learning and as a building-block for intuitive human-robot communication (see Figure 1).