100830-
any-
surement
formation
1ing is an
inimizing
, we also
ze typical
show co-
ethod has
stment of
3D can be
1eeds real
in many
different
or LS3D.
omorphic
S. Three
adjacent
each rule
Also the
shown in
nner data
ding the
surface
) ... The
) , which
t squares
sters is as
(I)
conjugate
is used:
Q)
where (x, y,7) € F ,(x',y',z') € G , R(P,®,K) is
: ; T. ;
the orthogonal rotation matrix, [# a" af. is the translation
vector, and m is the scale factor.
In order to perform least squares estimation, the true error
vector Vx. y.z) is introduced to describe the discrepancy
between the conjugate surfaces:
V(x,yz)yeG(x, v2) PF v.z) (3)
As continuous 3D surfaces have to be sampled with discrete
coordinates, a 3D surface matching is automatically translated
into a registration of point clouds. So the true error vector of
equation (3) can be approximately expressed with the
Euclidean distance of conjugate points, and the aim of least
squares estimation is defined as follows:
DI dd || = min (4)
If the square of the distance is set to D=d 2 the new
mathematical model of 3D surface matching can be simply
defined as follows:
D -(x-xYy *(y-yy *(z-zy (5)
Where (x, V,Z ) is the point coordinate of the template
surface, and the X', y ,Z')is the point coordinate of the
>» p
searching surface.
Since equation (5) is nonlinear, it must be linearized by the
Taylor expansion, ignoring 2nd and higher order terms.
D+V=D, De +20, +204
ot, ot, ot,
oD oD oD oD
+— do +— do +— dk + — dm
op ow OK om
where V can be considered as residual of the Taylor
expansion, and the Euclidean distance of conjugate points will
be set to zero at the end of the LS3D routine. Hence, the
observation error equation can be simplified as:
oD oD D
V=D +—dt +—dt 36D y +
dele LOL
(7)
oD D D
ED laua
Op ow OK om
The matrix form of equation (7) is as follows:
V=A4AX+1, P (8)
where A is the design matrix,
X I , 1, 10,0. K, ml is the parameter vector, and
l= D, is the constant vector that consists of the Euclidean
distances between the template and searching surface elements,
P is the weight matrix of the error observation vector.
The distribution of the random variable is
V- N(0, 0) , with the statistical expectation
E(V) «0, E{VV"} = 00, = oll’ . Hence, system
(8) is a typical Gauss-Markoff estimation model. In order to
control the estimation quality, an additional error observation
vector of the unknown parameters could be imported
(Robert,2004; Gruen and Akca, 2005).
V=IX+L, P (9)
Where / is the identity matrix, and L , is the constant vector
of the error equation, P is a priori weight matrix of unknown
parameters. If zero weight (2 ); — 0) is set, the i-th
parameter is assigned as free variable, and if an infinite weight
value (D )i — OO) is set, the i-th parameter is assigned as
constant Combining equation (8) and equation (9), the
maximum likelihood solution of unknown parameters can be
estimated as follows:
Nec PACPY'GÉPLCPL) — Qo
25:2
a, = PL +V. PV) x (11)
Vou AX 41 (12)
V =IX+L, (13)
where À is the final estimation value of least squares
— 2
routine, O, is the mean square error of the weighted units of
the observations, M is the number of error observation
equations, and 4 is the number of unknown parameters,
Ky =N—U are the components of abundant observation.
2.2 Conjugate Points Rules
Since it is rather difficult to locate feature points in a local
window on 3D surfaces, how to establish the conjugate points
between 3D overlapping regions, is the core strategy in the 3D
surface matching procedure. In our method, the conjugate
points rule is unlimited. We could define some new rules for
specified applications, because the mathematical model of the
adjustment, defined in 2.1, is generic for available rules. And,
it is the major advantages of our proposed method compared
with existing methods.
In this section, we list three strategies for establishing conjug-
ate points on 3D surfaces. The first definition called LND rule
is the same as the Least Normal Distance method (Robert, 2004;
Gruen and Akca, 2005), using pedal point of triangle in normal
direction. The second definition called LZD rule is the same as
the Least Z-Difference method (Rosenholm, 1988), using the
intersection point of a triangle in vertical direction. The last
definition called ICP rule is the same as the Iterative Closest
Point method (Besl and McKay, 1992), using two, the closest
points in the entire point sets as conjugate points. With the 3D
surface representation of triangulated irregular network
>
structure, the conjugate point rules can be listed as follows:
x E
T # 7 x
(a) (b) (c)
Figure 1. Conjugate points definitions for surface matching
with TIN structure: (a) LND rule, (b) LZD rule, (c) ICP rule.
where A, JB and C are the 3 vertexes of the candidate
conjugate region on the searching surface, and 7? is the normal
vector of the located triangle, V is the vertical vector of the
located triangle, $, . .$,. S, are behalf of the Euclidean
distances from the interpolation point to the 3 vertexes, and the