Page 43 - Autonomous Mobile Robots

26 Autonomous Mobile Robots

1.3.5.2.2 Multi-baseline stereo vision
Two main challenges facing a stereo vision system are: mismatch (e.g., points
in the left image match the wrong points in the right image) and disparity
accuracy. To address these issues, multiple (more than two) cameras can be
used. Nakamura et al. [34] used an array of cameras to resolve occlusion by
introducing occlusion masks which represent occlusion patterns in a real scene.
Zitnick and Webb [35] introduced a system of four cameras that are horizontally
displaced and analyze potential 3D surfaces to resolve the feature matching
problem.
When more than two cameras or camera locations are employed, multiple
stereo pairs (e.g., cameras 1 and 2, cameras 1 and 3, cameras 2 and 3, etc.)
result in multiple, usually different baselines. In the parallel configuration,
each camera is a lateral displacement of the other. For a given depth, we then
calculate the respective expected disparities relative to a reference camera (say,
the leftmost camera) as well as the sum of match errors over all the cameras.
The depth associated with a given pixel in the reference camera is taken to be
the one with the lowest error. The multi-baseline approach has two distinctive
advantages over the classical stereo vision [36]:

• It can find a unique match even for a repeated pattern such as the
cosine function.
• It produces a statistically more accurate depth value.

1.3.5.2.3 General multiple views
During the 1990s significant research was carried out on multiple view geo-
metry and demonstrating that 3D reconstruction is possible using uncalibrated
cameras in general positions [14]. In visual guidance, we usually have the
advantage of having calibrated cameras mounted in rigid fixtures so there seems
little justification for not exploiting the simplicity and speed of the algorithms
described earlier. However, the fact that we can still implement 3D vision even
if calibration drifts or fixtures are damaged, adds robustness to the system
concept. Another advantage of more general algorithms is that they facilitate
mixing visual data from quite different camera types or from images taken from
arbitrary sequences in time.

1.3.5.3 Application examples
In this section we present some experimental results of real-time stereo-vision-
based obstacle detection for unstructured terrain. Two multi-baseline stereo
vision systems (Digiclops from Pointgrey Research, 6 mm lens) were mounted
at a height of 2.3 m in front and on top of the vehicle, spaced 20 cm apart.
The two stereo systems were calibrated so that their outputs were referred to

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 26 — #26