Graphics Reference
In-Depth Information
Chapter 9
Conclusion
This work has described the state of the art in three-dimensional computer vision
as well as achievements which have resulted from a variety of newly introduced
methods. In this final chapter we discuss the main contributions of this work and
possible future research directions.
The first part of this work has discussed three very general classes of three-
dimensional computer vision methods. As a first class of methods for three-
dimensional scene reconstruction, triangulation-based approaches have been re-
garded. We have seen that three-dimensional scene reconstruction based on point
correspondences between a set of images is a problem for which theoretically well-
founded mathematical frameworks are available in the contexts of classical bundle
adjustment and projective geometry. Similarly, intrinsic and extrinsic camera cali-
bration and self-calibration are solved problems for which many different solutions
exist. As a general rule, camera calibration based on linear methods in the frame-
work of projective geometry should be refined by a bundle adjustment stage. The
relative accuracies of linear projective geometry methods and bundle adjustment ap-
proaches for three-dimensional scene reconstruction and camera calibration strongly
depend on the utilised camera system and on the application at hand, such that gen-
eral rules are hard to obtain.
A major drawback even for recent camera calibration systems is the fact that the
calibration rig has to be identified more or less manually in the images. At this point
important progress has been achieved in this work by introducing a graph-based
method for the automatic detection of the calibration rig and its orientation. It has
been shown that for wide-angle lenses with strong distortions or non-pinhole op-
tical systems such as fisheye lenses or catadioptric omnidirectional cameras, it is
preferable to use chequerboard patterns instead of photogrammetric retro-reflective
markers as calibration rigs. For such optical systems, nontrivial corrections need to
be applied to the measured centres of circular markers while chequerboard corner
locations are point features and thus bias-free. A method for the localisation of che-
querboard corners at high accuracy based on a physical model of the point spread
function of the lens has been introduced; this method yields an accuracy compa-
rable to that of a circular marker detector under favourable illumination conditions
Search WWH ::
Custom Search