Vision attention learning model and its application in robot

摘要

This paper presents a computational model of neural network for both spatial and temporal weights, and a unified adaptation scheme based on two biologically plausible learning rules-Hebbian rule and lateral inhibition is proposed. This model is applied to color video environment to develop a set of complete spatiotemporal weights simulating receptive field of simple cell in primary visual cortex, which can extract features of this receptive field such as edges, color and motion simultaneously. For real time application in robot a simplified method in frequency domain is proposed to approximate this kind of spatiotemporal weights. Experimental results, for a robot vision with top-down guidance used our model, show that our model is efficient on attention selection and object tracking.