Ly
IS
j-
l-
/
>. UM 02 C 025 0
— LA
ISPRS Commission III, Vol.34, Part 3A , Photogrammetric Computer Vision“, Graz, 2002
1. The improvement in the objective function (8), after
some iteration step, is below a chosen value.
2. The number of iterations exceeds a chosen constant.
5 SIMULTANEOUS REGISTRATION WITHOUT
CORRESPONDENCES
Given are N clouds of 3D data points (systems
¥1,..., XN) that partially overlap, but now we have nei-
ther correspondences between point pairs, nor confidence
values, as in Sec. 4. But we still assume that a good initial
position of these N 3D point sets in a global coordinate
system is known. Furthermore we know which of the pairs
of systems X;, X4 actually overlap.
In Sec. 4 each data point x; € Xj that corresponds to y; €
X, contributes to the functional F' in Eq. (8) with the term
Q2(x;,y;) which is quadratic in the unknowns vj; (x;) =
(€; — En) - (ej — ex) x xi. Now we also want to set
up such a quadratic functional for each data point x; in
an overlapping region. We present a strategy based on a
quadratic approximation of the squared distance function,
cf. (Pottmann Hofer, 2002).
If two systems 9;, X, overlap we triangulate the point
cloud in X, and refer to the triangulated point cloud as
T (X). First we determine those points x; € X; that ac-
tually lie in the overlapping region, i.e., their distance to
T (X4) is below a certain threshold. For literature on these
thresholding techniques, see e.g. (Blais Levine, 1995, Eg-
gert et al., 1998, Zhang, 1994).
Now, the goal is to bring the points x; closer to the geo-
metric shape 7'(3,), i.e., to move the points x; to lower
levels of the squared distance function to the triangulated
surface T' (X). In (Pottmann Hofer, 2002) it is described
how to compute for any point x; € IR? a local quadratic
approximant Fg x, =: F: to a smooth surface. For our ap-
plications this has to be a nonnegative quadratic function,
Fi(x) > 0,Vx € R?. Here we do not have a smooth sur-
face but a triangulated point cloud T (X4). We can apply
the results of (Pottmann Hofer, 2002), if we are able to
locally approximate the triangulated surface by a smooth
surface. For this we can use local quadric fits to T (X4),
see e.g. (Yang Lee, 1999). In case that the triangulation is
too coarse, we may first apply mesh refinement techniques
(e.g. interpolatory subdivision) to get a sufficiently dense
triangulation (Dyn et al., 1990, Zorin, 1997).
Now the registration of two such overlapping point clouds
is found by iteratively minimizing the functional
Y Fi(xi * vj (xi), (13)
which is quadratic in the unknowns ¢;, €;, Ck, C. If we
do not consider points x; from one overlapping region
(3, X.) only, but points from all overlapping regions si-
multaneously, then we have to minimize the functional
(13) for the unknowns e1, €, . , €N, EN. In order to fix
certain systems Xj, one simply has to set the vectors ¢;, €i
equal to zero.
As a simple example for a quadratic approximant F 7 of
the squared distance function of T (X4) at the point x;, we
would like to present a squared point-tangent plane dis-
tance: Let y; denote the closest point of T (Y) to xi. We
estimate the unit normal vector n; to 7 (X) in y;, e.g. by
computing a regression plane using the points in a certain
neighborhood of y;. In the following we refer to the plane
with normal vector n; through y; as the tangent plane of
T (X4) in y;. Let d; denote the oriented distance of x; to
this plane. If x; is already close to T (X) it is better to first
refine the triangulation and then compute the nearest point
and proceed as mentioned above.
We want to minimize the squared distance of the displaced
point x; = x; + vji (x) to the tangent plane at y;. The
distance is given by d; 4- n;vj,(x;) and thus we get the
functional
Qs(xi) = Fi(x}) 7 (di vjx(xi) : Mi)” (14)
= (d; + (€; Em Cr) | n de det(c; — Cgk,Xi, n;))?,
which is quadratic in the unknowns cj, €;, Cx, Ck. In
this way we can compute a quadratic term )3(x;) for all
points x; that have been found to lie in an overlapping re-
gion. Hence, to simultaneously register all N point clouds
(where at least one is fixed) we have to minimize the fol-
lowing functional
F=Y Qax:)- (15)
As mentioned in Sec. 4 one has to observe the fact that the
map x; + X; + Vjo(X;) is no Euclidean rigid body motion.
See Sec. 4.2 for the computation of the rigid motion which
brings x; close to tips of the vectors x; + Vjo(Xi).
The squared distance function to the tangent plane has al-
ready been used in several variants to the ICP algorithm,
cf. (Bergevin et al., 1996, Chen Medioni, 1992). Our ap-
proach of a local quadratic approximant of the squared dis-
tance function is more general and includes the squared
distance function to the tangent plane as a special case.
Fig. 3 gives an example for our algorithm in the case
N = 2, namely the registration of a point cloud to a CAD
model. In Fig. 4 the mean squared error of the data points
to the CAD surface after each iteration is given, both for
our algorithm and for the standard ICP algorithm. Our al-
gorithm converges much faster than ICP, and with respect
to the number of iterations it is comparable to Chen and
Medioni's method. However, we solve just a linear sys-
tem in each iteration step, whereas Chen and Medioni in
each step run a numerical optimization algorithm. We are
currently working on an efficient organization of the local
quadratic approximants in a spatial data structure in order
to speed up the algorithm. Then, industrial inspection tasks
are expected to come close to real time performance, since
the computations on the approximation of the squared dis-
tance function to the CAD model can be done in a pre-
processing step. Further ideas for acceleration and stabi-
lization of the algorithm can be taken from (Rusinkiewicz
Levoy, 2001).
A - 269