Using a rat brain model and a single car-mounted webcam,
researchers at the University of Queensland have mapped a suburb of
Brisbane, Australia, in real time.
Called RatSLAM
(simultaneous location and mapping), and running on a PC, the
programme models cells known to exist in a rodent hippocampus.
The cells are all linked to visual observation: 'place' cells
fire when a rat is at a specific place in an environment, 'head
direction' cells fire when the rats head is pointing in a certain
direction, and 'grid' cells fire when observing similar-looking
locations.
Using 'continuous attractor networks', a type of neural network,
the researchers have selectively modelled the interaction of these
rat cells to produce behaviour suited to SLAM tasks.
The network was fed with data originally recorded from the
single camera of an Apple Macbook mounted on the roof of a car
driven on a 66km journey, threading its way at least once around
every road in the suburb.
From images captured every 100ms and processed on one core of a
2.4GHz dual core, the algorithms generated a viable road map of the
entire area, appropriately connecting roads together at many-way
junctions and making no significant errors. This involve closing 51
loops of road up to 5km long.
Following mapping, the algorithm could also located its position
within the environment in at most 6.5s of being shown a picture
from within it.
In other tests, with a camera mounted on a tiny vehicle inside
an unfamiliar building, the mapping algorithm sorted it self out
even when the vehicle was unexpectedly picked up and put down in
another part of the lab while mapping.