Abstract
This paper presents a critique on previous work in
the field of vision aided navigation, particularly in the fusion of
visual and inertial sensors for navigation. Several improvements
and updates are proposed for the existent systems. GPS receivers
have allowed for accurate navigation for many vehicles and
robotic platforms. GPS based navigation can, however, prove to
be impractical in applications where there is no GPS reception
such as underground, indoors or in some urban areas. This pertains,
in particular, to many robotic applications where position
must be known in global coordinates or relative to a reference
point. An inertial navigation system (INS) can be used to calculate
one’s relative navigation state via dead-reckoning calculations.
The downfall of a low-cost INS is the errors associated with the
system. While these errors are initially small, integration causes
large drift errors over time. To combat this problem, cameras
can be used to estimate the errors present in the INS readings.
These results can then be used to correct the navigation state
output from the INS. While the motion estimations from the
cameras are not error-free, this method is made highly effective
because of the complementary nature of the errors from the
cameras and INS. Several improvements are proposed for this
method; algorithmically, in updates to its hardware, and with
the introduction of graphics processors to improve computational
performance. The overall system performance, individual steps,
algorithms, and results are compared to results from similar
works to those of the proposed improvements. It is shown that the
accuracy, responsiveness and overall performance of the system
can potentially be greatly improved.