Next: Pose versus Correspondence Search
Up: Range Data Registration
Previous: Inexact Correspondences
When registering two (or more) data sets, there are usually
3D data points or other features that do not have any correspondence
between the data sets. The two main causes of this are: 1) regions of
the data where there is no overlap (a natural consequence of extending
the data description by incrementally fusing partially overlapping data)
and 2) noise outliers (79,51,10,18,62).
Hence, great effort has been made to identify outliers and partial
overlap based on techniques such as high dimensional distance
measurement (24,63,37), orientation consistency
(79), interpoint distance (20), boundary points
removal (57,72), threshold (62), and motion
properties (43,56,42). It has been shown that the
identification of outliers and partial overlap are essential steps to
the estimation of motion parameters. Moreover, these two steps
are often interweaved and affect each other especially when no
exact information is available about the distribution of points,
occlusion, appearance and disappearance of points. This implies
that assuming image data as a black or grey box, the development
of techniques to register such data is still a challenging task
and prone to local minimum convergence as pointed out above.
Future research may focus on the use of structural data information,
motion properties, special information on objects, or special imaging
configurations for a more accurate estimation of image correspondences.
Next: Pose versus Correspondence Search
Up: Range Data Registration
Previous: Inexact Correspondences
Bob Fisher
2003-07-13