Contrôle de la perception orienté structure dans le domaine de la cuisine piloté par des méthodes d'attention computationnelle
Daniel Beßler · HAL (Le Centre pour la Communication Scientifique Directe) · 2014
In this thesis, knowledge-based (top-down) attention is investigated in the context of mobile robotics in the kitchen domain. Household robots have to accomplish everyday activities such as the preparation of breakfasts. The preparation of breakfasts requires that household robots find a set of items in the kitchen, carry items to the kitchen table as well it requires that household robots arrange items on the table. For example, there could be a particular household where a specific red cup is always used during breakfast. It is (usually) required to open drawers and doors of cabinets in order to search for kitchen items. In unknown kitchens, it is hard to find the location of the wanted item with only one try because the layout of kitchens differs significantly between different households and individual habits. Thus, knowledge about the household can reduce the problem space in this scenario by focusing on likely locations of the wanted item. For each investigated location, the robot has to estimate if a red cup is visible or not. Knowledge about the appearance of cups as well as the knowledge that the wanted cup is red can be used to focus the attention on visible regions, objects and features that might belong to the red cup.