James McGreger
Computer Science
M.S. 1998
Registration of Laser-Range Images
Research Objectives:
Being able to accurately model a 3D scene is important to many applications. Three major steps are
involved in the process: (1) data acquisition from multiple viewpoints,
|
 |
(2) view registration, and (3) view integration or data fusion. Our focus is on view
registration, which is the problem of determining the rigid transformation that
describes the motion of the camera platform between two or more views. While a
completely data-driven method is the goal, semi-automatic methods requiring
minimal user interaction are also being investigated.
Methodology and Results:
Several point-based methods have been proposed for fine-tuning a given motion estimate. One exemple is the
so-called Iterative Closest Point (ICP) algorithm, which is widely used and
constitutes a generic enough framework to describe other related algorithms.
Starting from the given motion estimate, point correspondences are established
and fed to a least-squares solver that produces a new motion estimate. The
process continues until convergence or for a predetermined number of iterations.
Different versions of the ICP algorithm are studied to determine their
robustness and possible application to real data. Sometimes an approximate
motion estimate is not available, e.g. when the camera platform is moved around
manually. Under such circumstances, it may be impossible to establish good point
correspondences, even via user interaction. To overcome this problem, we take a
plane-based approach. The underlying philosophy is that planar surface patches
are typically plentiful in man-made scenes, e.g. buildings; they can be detected
and modeled robustly via appropriate segmentation and least-squares fitting, and
a user can then easily identify correspondences. The one stipulation is that the
set of planes selected in each view must span a 3D space. The approach can
easily be extended to use other data models such as generalized cylinders.
This work was conducted by James McGreger
while at IRIS lab under the supervision of J. Gregor. This work was
supported by DOE's University Research Program in Robotics under grant
DOE-DE-FG02-86NE37968.