Page 249 - Innovations in Intelligent Machines

242 J. Gaspar et al.
our purposes a good dewarping is simply the one that makes straight lines on
the ground plane appear straight in the Bird’s Eye View.
As long as the mirror, camera and support (mobile platform) remain fixed
to each other, the dewarpings for panoramic and bird’s eye views are time
invariant and can be programmed with 2D lookup tables. The dewarpings are
done efficiently in this way.
Doing fixed image dewarpings is actually a way to do (or help) Scene
Modelling. The image dewarpings make geometrical properties of the scene
clearly evident and as such simplify scene modelling to collecting a number of
features.

Geometric Scene Modelling and Model Tracking

Geometric models of the scene are collections of segments identified in Bird’s
5
Eye and Panoramic views . Ground segments are rigidly interconnected in the
Bird’s Eye views while vertical segments will vary their locations according
to the viewer location. Considering both types of segments, the models are
“wire-frames” whose links change according to the viewpoint.
Each scene model must have a minimal number of features (line segments)
in order to allow self-localisation. One line of the ground plane permits finding
only the orientation of the robot and gives a single constraint on its locali-
sation. Two concurrent ground lines, or one ground and one vertical, already
determine robot position and orientation. Given three lines either all vertical,
one on the ground, two on the ground (not parallel) or three on the ground
(not all parallel), always permit us to compute the pose and therefore form
6
valid models .
Figure 7 shows one example of modelling the scene using line segments
observed directly in the scene. The model is composed of three ground lines,
two of which are corridor guidelines, and eight vertical segments essentially
defined by the door frames. A single door frame (i.e. two vertical lines) and
one corridor guideline would suffice but it is beneficial to take more lines than
minimally required in order to improve the robustness of self-localisation.
7
In order to represent a certain scene area, and to meet visibility and qual-
ity criteria, a minimal number of segments are required. Models characterising
different world regions are related by rigid 2D transformations. These trans-
formations are firstly defined between every two neighbour models at locations
where both models are (partially but with enough relevance) visible. Naviga-
tion is therefore possible in the area composed as a union of individual areas,
provided by each individual model.
5
Despite the fact that localisation can be based on tracked image corners [74], more
robust and stable results are obtained with line segments as noted for example
by Spetsakis and Aloimonos in [78].
6
Assuming known the xy coordinates of the intersection of the vertical line(s) with
the ground plane.
7
see Talluri and Aggarwal in [83] for a geometrical definition of visibility regions