ITEE Ph.D confirmation seminar: David Prasser, 11.00AM, Thu 24 Apr 2003

Appearance Based Localisation for a Mobile Robot

Speaker: David Prasser, ITEE

When: 11.00AM, Thursday 24 Apr 2003

Venue: 78-420

Host: Dr Gordon Wyeth

Abstract:

  The problem of converting camera images to position information for
  mobile robots, known as localisation, has come under attention
  during the past decade. A related problem is the automatic
  construction or learning of a suitable map to be used for
  localisation. The combined task of learning an environment under
  uncertainty of motion and at the same time using the learnt
  environment to localise is known as Simultaneous Localisation and
  Mapping (SLAM).

  An appealing approach to localisation is to learn to relate an image
  I directly to world coordinates (x, y, theta) without considering a
  physical world model. Many learning techniques that attempt this
  operate on a two stage approach - a learning phase where the visual
  environment is learnt as a function of position and a testing or
  validation phase. However a visual learning method that functions in
  a SLAM system must, by definition, be able to function without first
  being exposed to the entire environment.

  A model that has recently been studied for visual recognition is the
  Extended Conjunction of Localised Features (ECLF) network. This
  network has the property that learning a set of input features can
  be incrementally accomplished with only one presentation of each
  input. This property of incremental learning makes the ECLF network
  attractive for the problem of SLAM, where a robot would be required
  to learn a visual mapping without knowing what input patterns it may
  encounter later during its further exploration of the environment.

  This network will be employed as part of the SLAM project that aims
  to use neural models to perform other aspects of the localisation
  and mapping process. This system will be based on a simplified model
  of the rat hippocampus and use a discrete output known as place
  cells. Each place cell corresponds to a small region of physical
  space and ideally the most activated cell indicates the robots
  actual location. The ECLF network should be able to be attached
  directly to the place fields, which will produce a wholly network
  based SLAM system.

Biography:

(biography unavailable)

Type:

Ph.D confirmation

Contact:

Dr Gordon Wyeth, seminar host (wyeth@itee.uq.edu.au)
or Guido Governatori (ITEE seminar co-ordinator) (guido@itee.uq.edu.au)

ITEE seminar web page: http://www.itee.uq.edu.au/~seminar

[All seminars]