The first part of the project will include investigation
of the architectures which make vision as we
know it possible. Clearly vision is a function of intelligence
as much as sensor input and image processing.
In order for a vision system to have reasonable
performance in a complex operating environment the
architecture will need to collect, integrate, and direct
multiple input sensors, and evaluate, maintain and update models of the current state of the operating
environment. In the past most computer vision systems
have concentrated on the processing of single
sensor input image data. Human visual systems, on
the other hand, use multiple sensor inputs in addition
to complex internal representation of the individual's
environment in order to make decisions. New computer
vision architectures will incorporate standard
video input with active feedback and range sensor
data to enhance the ability to collect important data.
In order to deal with low data rates and extremely
complex and varied tasks, humans use higher brain
functions to enhances their vision system. By maintaining
models of their environment, humans can perform
complex tasks using many parts of their brain in
parallel. Environmental modeling can also be used to
enhance the performance of computer vision systems.
In contrast to the traditional approaches of constructing
and updating kalman filter-like representations, ,
the new architecture will use the human model to
evaluate the system state as a function of deviations
from and movement with probabilistic models of task
environments.