Page 258 - Innovations in Intelligent Machines

Toward Robot Perception through Omnidirectional Vision 251

Fig. 11. (a) An omnidirectional image obtained at 11:00, (b) one obtained at 17:00;
(c) An edge-detected image and (d) its retrieved image

eigenspace using both dilated and un-dilated model views and pre-process the
run time edge images to dilate the edges. In our pre-processing we use low pass
filtering instead of edge dilation. The purpose here is to maintain the local
maxima of gradient magnitude at edge points while enlarging the matching
area. We found this to be a good tradeoff between matching robustness and
accuracy.
To test this view-based approximation we collected a sequence of images,
acquired at different times, 11am and 5pm, near a large window. Figure 11
shows the significant changes in illumination, especially near the large window
at the bottom left hand side of each omnidirectional image. Even so, the view
based approximation can correctly determine that the unknown image shown
in Fig. 11(a) was closest to the database image shown in Fig. 11(b), while PCA
based on brightness distributions would fail. For completeness, Fig. 11 (c) and
(d) shows a run-time edge image and its corresponding retrieved image using
the eigenspace approximation to the Hausdorff fraction.

Integrating Topological Navigation and Visual Path Following
When continuously operating, the mobile robot is usually performing topo-
logical navigation. At some points of the mission the navigation modality is
required to change to the visual path following. Thus, the robot needs to
retrieve the scene features (straight lines in our case) chosen at the time of
learning to specific this particular visual path following task.