| l, Rue de l'Université, F-88100
ABSTRACT
be used to quickly provide an initial match when registering a
1 INTRODUCTION
The theme of our research group at St-Dié deals with task-
oriented 3-D modelling of existing environments (Even,
2004). Compared to other teams working in the computer
vision domain, our goal focuses on the acquisition of 3-D
models that are well adapted to given tasks. Typical re-
quirements include acquisition time, geometrical reliabil-
ity or visual realism, but also relevant structuration and se-
mantical contents. The operator's knowledge of the model
destination and his ability to select appropriate details in
the sensed data are necessary and manual interventions are
thus required within the modelling process. Therefore we
put emphasis on interactive methods with efficient com-
puter assistances that are based on geometrical reasoning,
image processing or expertise integration.
We chose a top-down approach (from 3D to 2D) where
solid primitives are selected and assembled together di-
rectly on the images in order to ensure the conformity to
reality (Even, 2001). It provides a more natural way to
build the 3-D model than a classical bottom-up approach
(from 2D to 3D) where homologous image features are first
extracted and then processed to basically build a surface
modcl. Its high flexibility to cope with hard environment
constraints and erroneous hypotheses was recently demon-
strated in the scope of a radioactive cell modelling work
(Leroux et al., 2002).
This interactive modelling method includes many objects
registration tasks to make their projection match relevant
image features. They are mainly intended to put objects at
their correct place in 3-D space, assuming that the back-
ground image is oriented. But inversely, they are also used
to orient new images using already modelled objects. The
object degrees of freedom (3 translations + 3 rotations) are
successively controlled with the mouse. Most often this
work proves uneasy. mainly because it is difficult to de-
termine the rotation axis. In structured environment, some
204
A GENERIC PROCEDURE FOR INTERACTIVE 3D MODEL REGISTRATION ON IMAGES
Philippe EVEN
LORIA / IUT
de St-Dié,
Saint Dié des Vosges, France,
Philippe.Even@loria.fr
KEY WORDS: Three-dimensional, Model, Image, Registration, User, Interface.
This paper presents a generic and user-friendly procedure for interactive 3-D model registration on images. Such tasks
are numerous in a CAD-based model acquisition process through interactive means. Moreover object alignment can also
whole model on actual data. This task could be performed
by setting a minimal set of 2D-3D point matches. Our goal is to set up a more user-friendly solution, which minimizes
the amount of manual work. We have implemented a two-step procedure relying on the selection in the image of some
structured object features. First a visual feature interpretation method is used. It provides a solution that fixes at least
two rotation parameters. Remaining degrees-of-freedom are then easily handled through interactive modes. Several
interpretation methods are proposed to cope with various contexts according to the available features. Achieved accuracy
is a posteriori estimated to help the operator decide on the registration reliability.
geometrical constraints can often be used to back-project
image features such as points, lines, or elliptical contours.
They deal with parallelism, orthogonality, axial symmetry,
etc. Each onc may require a dedicated interface, and so
globally hamper the system appropriation by the user.
Therefore we implemented a generic two-steps procedure.
Based on the available constraint, a direct solution is first
computed in order to fix the rotation axis. The remaining
degrees of freedom are then interactively determined using
the mouse. An estimation of the solution consistency is
finally displayed to the user so that he could decide of the
registration reliability.
2 ANALYTICAL SOLUTIONS TO THE REGIST-
RATION PROBLEM
Our top-down approach aims at limiting the number of vi-
sual feature extraction tasks. Many authors already ad-
dressed the problem of registration from a restricted set
of points or lines. In the RANSAC paradigm (Fischler and
Bolles. 1981), a small set of n control points is used to
provide a first solution, further enhanced with consistent
additional points. It was shown that a unique solution can
be computed from four coplanar points. Analytical solu-
tions were later proposed to solve this perspective-4-points
problem with different angular configurations between the
points (Horaud et al., 1989).
The pose determination problem can also be solved from
lines. It was shown that the projection of three lines with
known relative orientations leads to an eighth-degree equa-
tion (Dhome et al., 1989). Actually two vertical lines and a
horizontal line are sufficient to directly determine the rota-
tion matrix (Shum et al., 1998). Moreover the case of lines
belonging to planes with known configurations was also
studied (Chen, 1991); for coplanar configurations where
the lines belong to the same plane, orthogonal configura-
tions where the planes are mutually perpendicular, and pa-
Internatio
rallel cot
closed-fc
cones in
Recent v
its projec
and Liu,
lutions. .
reconstri
of views
3 THE
31 Pri
The mai
the deter
plemente
course be
tasks. Fit
the scene
to fix at
remainin
tively wil
A set of
enough t
countere(
user inter
the rigid
The two i
used to d
rations ai
- Stereo |
- Shape «
of a shap
- Coplan
in two vi
- Parallei
lines in a
- Orthoga
thogonal
The oper
= Prelin
to be reg
that obje
relevant «
a fixed in
to movet
termine t]
= Firsts
lution R.
arranged
P, is defi:
to the vie
It was pr
these sol
features t
lution to
