Visual Pose Estimation under Time Varying Lighting Conditions
Speaker: Stephen Nuske, ITEE
When: 2006-08-22 14:00:00
Venue: 78-420
Host: Gordon Wyeth
Abstract:Pose estimation is the fundamental robotic task of estimating the
position of a robot with respect to the world. A robot estimates its
own pose by locating objects in its surroundings, which can be
performed visually, using images taken by a camera. Visual pose
estimation is a difficult task because the appearance of objects
change due to fluctuating lighting conditions, making object
locating unreliable. To counter this problem current approaches try
to factor out illumination at the image-level. However to properly
account for the effects of illumination, higher-level geometric and
photometric properties of the scene must be known. Many existing
visual pose estimation approaches ignore this requirement and only
use image-level appearance representations. Approaches that do use
higher-level 3D information are edge-based and require an edge
extraction process to match the models to the image. Edge-extraction
is known to suffer when the complexity of the recognition task
increases and is also not robust to such effects as shadows and
shading caused by non-uniform lighting. As yet there are no visual
pose estimation frameworks which can completely account for all
changes in illumination and viewing pose. In the domain of face
identification there are recent techniques which use representations
that avoid the limitations of appearance-based and edge-based
approaches. The representations are explicit 3D geometric and
photometric models that do not include illumination or viewing pose
specific information. The models can be rendered incorporating
illumination information into synthetic imagery, allowing
identification using a matching process between real and synthetic
imagery. This project will bring the technique of matching real and
synthetic imagery from the domain of object identification, and
apply it to the domain of visual pose estimation. The aim is to
perform robotic pose estimation with the incorporation of
illumination information within the process rather than attempting
to factor out illumination at the image-level.
Biography:Stephen graduated with a Bachelor of Software Engineering in 2005
from the University of Queensland. He is currently enrolled in a PhD
program jointly sponsored by ITEE and CSIRO's Autonomous Systems
Laboratory.
Type: Ph.D confirmation
Contact:Gordon Wyeth, seminar host (wyeth@itee.uq.edu.au)
or Guido Governatori (ITEE seminar co-ordinator)
(guido@itee.uq.edu.au)