Abstract
For robots observing and learning human tasks, changes of objects' velocity and their access, or contingence have been proposed as events that segment human actions. In this paper, we propose a novel event for segmentation, superposition. Superposition can be calculated only with front surface's information of manipulated objects. Additionally, it can be useful in recognizing functions of daily objects. Second, a visual system to track multi-objects and to estimate their superposition is proposed. This system is an expanded template matching which uses an anticipated non shielded image to track an object, a non shielded image to estimate visual changes of an object. Finally, we tested out this system by an experiment that HRP-2VZ humanoid vision system with the estimating system observed a human using a tool in daily environment. This result showed the effectiveness of our approach.
