Page 260 - Innovations in Intelligent Machines

Toward Robot Perception through Omnidirectional Vision 253

Fig. 12. Three methods of topological localisation, (top, from left to right): local-
isation based on PCA, Chamfer distance and Hausdorff distance. The clear valleys
show that there is enough information to distinguish the robot locations. (Bottom)
localisation as found by each of the methods i.e. ordinates corresponding to mini-
mum values found at each time instant on the 3D plots

Fig. 13. Topological localisation experiments using the three methods over two
sequences acquired under different lighting conditions. From left to right: localisation
based on PCA, Distance transform and Hausdorff distance

As expected, the edges based methods are more suited to dealing with
very different illuminations. In our navigation experiments we use mainly the
PCA over brightness values, as most of our scenario is not subject to large
illumination changes, and using brightness values is more informative than
using only edges. For the parts of the scene where illumination can change
significantly we use the Hausdorff based method. The reason of its choice
when compared to the Distance transform, is that it is faster for the first
localisation at dropped-in-scene situations.

Integrated Navigation Experiments

The concluding experiment integrates global and local navigational tasks, by
combining the Topological Navigation and Visual Path Following paradigms.