Real-Time Active Robotic Vision using Biologically Inspired Neural Models

Tyler Garaas
Visual Attention Lab
University of Massachusetts Boston

The discipline known as computer vision strives to extract meaning
from the patterns of light falling across a sensor. Many difficulties
hinder the advancement of this cause, ranging from ambiguity as to
what form best encapsulates the structure of the information to the
intractable computational bandwidth required to process sensor images
of sufficiently high resolution. In the present talk, I propose a 
system of neurons inspired by a subset of the visual areas of the 
primate brain in order to direct the “gaze” of a pan/tilt/zoom camera 
in real-time to areas of interest that are demarcated by one or more 
low-level visual features known to form the basis of visual abilities 
in primates. Results of the visual search performed by the neural
vision system developed here are compared directly with the results of
human subjects performing the same task. Finally, in order to overcome
the high computational bandwidth required to model millions of neurons
in real-time, I exploit the highly parallel nature of the neural model
– as it exists in biological vision systems – to develop a
computational software model where thousands of neurons can be modeled
simultaneously on recently available parallel computing architecture
known as the graphics processing unit (GPU).