A fast and scalable system for visual attention, object based attention and object recognition for humanoid robots
Andreas Holzbach, Gordon Cheng · 2014
In this paper, we present a novel approach towards the integration of visual attention, object based attention and object recognition. Our system is scalable in regard to the required framerate or usage of computational power. Therefore, it is perfectly suited for robotic applications, where time is a crucial factor. We enhance and evaluate our previously presented visual attention system based on sampled template collation (STC) to fit into a humanoid robotic context by dynamically adjusting the required computational speed. We modify STC for object-based attention to segment the attended object from the surrounding background. Subsequently we combine it with a biologically-inspired object recognition system. We show that our approach significantly improves the recognition accuracy.