Learning sequential visual attention control through dynamic state space discretization
ICRA'09 Proceedings of the 2009 IEEE international conference on Robotics and Automation
Online learning of task-driven object-based visual attention control
Image and Vision Computing
| Hi-index | 0.00 |
This work describes a framework for dealing with attention and categorization using a robot platform consisting of an articulated stereo-head with four degrees of freedom (pan, tilt, left verge, and right verge). As a practical result of this development, the system can select a region of interest, perform shifts of attention involving saccadic movements, perform an efficient feature extraction and identification/recognition, incrementally construct a world map, and keep this map consistent with a current perception of the world. Another important result for the attentional mechanism is that the system is capable to visit all regions of its restricted world, selecting one region at a time according to a salience map. For identification, the system starts without any knowledge of the environment and increases its knowledge base (associative memory) as necessary to deal with a current set of objects.