3D COORDINATE TRANSFORMATIONS

R E DEAKIN
Department of Land Information
RMIT University
GPO Box 2476V MELBOURNE VIC 3001
AUSTRALIA

Phone: +61 3 9925 2213

Fax: +61 3 9663 2517
e-mail: deakin@rmit.edu.au

(published in Surveying and Land Information Systems, Vol.58, No.4, Dec.1998, pp.223-34)

ABSTRACT and control systems used in large manufacturing

A three-dimensional (3D) conformal coordinate
transformation, combining axes rotations, scale
change and origin shifts is a practical mathematical
model of the relationships between different 3D
coordinate systems. Applications in geodesy and
photogrammetry often use simplified
transformation models under the assumption of
small or negligible rotations, but in other areas of
interest rotations may be large. In such other cases,
approximate values of rotations are required to
perform initial transformations before the simplified
models are employed. This paper uses an example
applicable to the construction industry to
demonstrate methods of calculating approximate
rotations and performing initial transformations
prior to computing transformation parameters. A
rigorous development and proof of the 3D
conformal transformation is given as well as the
necessary assumptions for the simplified model. In
addition, this paper also explains how least squares
may be used in determining transformation
parameters.
INTRODUCTION
Coordinate transformations, conformal and
otherwise, are widely used in surveying and related
professions. For instance, in geodesy, 3D
transformations are used to convert coordinates
related to the Australian Geodetic Datum to the
new Geocentric Datum of Australia (Featherstone
1996), in engineering they form part of monitoring
projects such as the construction of the ANZAC
frigates for the Australian and New Zealand Navies
(Bellman & Anderson 1995) and in photo-
grammetry they are used in the orientation
(interior and exterior) of aerial photographs. In
two-dimensional (2D) form, transformations are
used in cadastral survey re-establishments (Bebb
1981, Sprott 1983 and Bird 1984), matching
digitized cadastral maps (Shmutter and Doytsher
1991) and "sewing together" the edges of strips of
digital images (Bellman, Deakin and Rollings
1992).
In general, the effect of a transformation on a 2D
or 3D object will vary from a simple change of
location and orientation (with no change in shape
or size) to a uniform change in scale (no change in
shape) and finally to changes of shape and size of
different degrees of nonlinearity (Mikhail 1976).
The most common transformations in surveying
applications, and the only type dealt with in this
paper, are conformal, i.e. transformations that
preserve angles and thus the shape of objects.
Theory and applications of other coordinate
transformations, such as affine, polynomial,
projective etc. can be found in Mikhail (1976) and
Moffitt and Mikhail (1980).

1
3D conformal coordinate transformations − also B
known as similarity transformations (Blais 1972 C
A
and Bervoets 1992) and in the 2D case, sometimes F
called Helmert transformations after the German
geodesist F.R. Helmert (1843-1917) − are often G
U E
given in the form
N D
⎡E ⎤ ⎡X ⎤ ⎡ TE ⎤ The
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ OBJECT
⎢N ⎥ = λ R κ φ ω ⎢ Y ⎥ + ⎢ TN ⎥ (1) H
⎢⎣U ⎥⎦ ⎢⎣ Z ⎥⎦ ⎢⎣ TU ⎥⎦

E, N and U (East, North, Up) and X, Y and Z are Design E

3D "design" and "survey" coordinates respectively, origin
λ is a scale factor, R κ φ ω is a rotation matrix (the
product of rotations ω, φ and κ about the X, Y Z U′

and Z axes in turn) and TE , TN and TU are N′

translations between the origins of the two
coordinate systems. In the context of this paper,
(1) – often called a seven-parameter
transformation, three rotations three translations
and one scale factor – represents the mathematical
relationship between a constructed object in the
XYZ survey system and its transformed position in
the ENU design system. Its practical use in the
construction industry, where components of
structures are manufactured "off-site" and brought
together "on-site" can be explained in the following
way.
At the off-site location the object is measured in
situ and coordinated in an arbitrary XYZ survey
system. To confirm that it will fit in its designed
on-site location its XYZ coordinates are scaled and
rotated until the coordinate axes are parallel with
the ENU design axes. The origins of the two
coordinate systems are then brought together by
adding the translations. This has the effect of
superimposing the object over its design location
and its "fit" can be checked by comparing
coordinate differences.
κ
φ Y
Survey ω E′
origin
X
Figure 1. An object with coordinates in two 3D
rectangular systems
The seven parameters in (1) can be determined by
solving a system of equations derived from
"common" points whose coordinate values are
known in both the survey and design systems.
Each common point (or control point) generates
three equations, thus a minimum of three such
points is required to solve for the parameters, but it
is usual (and also good practice) to include extra
control points in surveys of this type; the additional
points leading to a redundant system of equations
(more equations than unknowns). Least squares
can then be employed to determine the best
estimates of the parameters which minimize the
sum of the squares of the residuals (small
corrections to the transformed coordinates) at the
control points. The residuals, three for each point
in the direction of the coordinate axes, can then be
used as a measure of how well the constructed
object fits its design values.

2
Translation parameters may be eliminated from the iterative, requiring an initial transformation to
solution by using "centroidal" coordinates approximately align the two coordinate axes, a
E , N and U and X , Y and Z having a common simple method of determining the parameters of
the initial transformation is given and a worked
origin at the centroid of the control points. The
example of a problem is provided.
transformation then combines scale and rotation
only
2D CONFORMAL TRANSFORMATIONS
⎡E ⎤ ⎡X⎤
⎢ ⎥ ⎢ ⎥
⎢N ⎥ = λ R κ φ ω ⎢ Y ⎥ (2a) C.F. Gauss showed that the necessary and
⎢U ⎥ ⎢Z ⎥ sufficient condition for a conformal transformation
⎣ ⎦ ⎣ ⎦
from the ellipsoid to the map plane is given by the
complex expression (Lauf 1983)
[Translations can also be eliminated by moving the
origins to a common control point, both techniques
y + i x = f (χ + i ω) (3)
having the desirable effect of reducing the size of
the numbers involved in the computations.]
where the function f (χ + i ω) is analytic,
Due to the nature of the matrix R κ φ ω it is not containing isometric parameters χ (isometric
possible to solve directly for the individual latitude) and ω (longitude). i is the imaginary
rotations ω, φ and κ . But if the coordinate axes number ( i 2 = −1 ) and the xy rectangular
are approximately parallel, a matrix RS of coordinates have the y-axis "up the page" and the
independent small rotations δω, δφ and δκ can x-axis "across the page". [It should be noted here
that isometric means: of equal measure, and on the
be developed, which together with an approximate
surface of the ellipsoid (or sphere) latitude and
value of the scale factor λ ′ with its small unknown
longitude are not equal measures of length. This is
correction δλ ( λ = λ ′ + δλ ) enables a least obvious if we consider a point near the pole where
squares solution of δω , δφ , δκ and δλ using similar distances along a meridian and a parallel of
latitude will correspond to vastly different angular
⎡E ⎤ ⎡X ⎤ values of latitude and longitude. Hence in
⎢ ⎥ ⎢ ⎥ conformal map projections, isometric latitude is
⎢N ⎥ = ( λ ′ + δλ ) RS ⎢ Y ⎥ (2b) determined to ensure that angular changes
⎢U ⎥ ⎢Z ⎥ correspond to linear changes.]
⎣ ⎦ ⎣ ⎦

This solution technique requires iteration, each

iteration preceded by a transformation using (2b) A necessary condition for an analytic function is
with the process being terminated when δω, δφ , that it must satisfy the Cauchy-Riemann equations
δκ and δλ reach some predetermined negligible
values. ∂y ∂x ∂y ∂x
= and =− (4)
∂χ ∂ω ∂ω ∂χ
This paper shows how the general principles of
conformal transformation, originally developed by
Using this theorem, a conformal transformation
C.F. Gauss (1777-1855), are used to derive a
from the XY rectangular coordinate system
transformation between two plane rectangular
(isometric parameters) to the EN (East, North)
coordinate systems which is equivalent to 2D
rectangular system (also isometric parameters) is
rotation, scaling and translation. This concept is
given by the complex expression
extended to 3D systems and the derivation of (1)
and (2) is set out together with a proof showing
N + i E = f (Y + i X) (5)
that angles are preserved in the transformation.

The application of the least squares principle to the A function f ( Y + i X ) which satisfies the Cauchy-
solution of the transformation parameters is Riemann equations is a complex polynomial, hence
explained and the necessary system of equations is (5) can be given as
developed in matrix form. Since the solution is

3
 n N′ Y
∑ (a + i bk )( Y + i X )
k
N + iE = (6)
k
Y sin α X sin α
k =0
α
Y cos α
Expanding (6) to the first power (k = 1) and N P
equating real and imaginary parts gives
α X cos α
N = a0 + a1Y − b1X TN E′
(7)
E = b0 + a1X + b1Y
X
which are essentially the same equations as in E
Jordan/Eggert/Kneissal (1963, pp. 70-73) in the TE
section headed "Das Helmertsche Verfahren
(Helmertsche Transformation)" although as noted Figure 2. 2D Conformal Transformation
(Rotation and Translation)
by Bervoets (1992) in his bibliography, there is no
reference to the original source. It is probable that
Thus it is seen that the 2D conformal
Helmert developed this conformal transformation
transformation is equivalent to a translation and
in his masterpiece on geodesy, Die mathematischen
rotation of rectangular axes with a scale factor
und physikalischen Theorem der höheren Geodäsie,
between the EN and XY coordinates.
(The mathematics and physical theorems of higher
geodesy) on which he worked from 1877 and
published in two parts: vol. 1, Die mathematischen
THE 3D ROTATION MATRIX
Theorem (1880) and vol. 2, Die physikalischen
Theorem (1884) [DSB 1972].
The 2D transformation can be extended to three
dimensions by firstly considering a sequence of
In (7), a0 and b0 are translations between the rotations of ω, φ and κ about the X, Y and Z-axes
coordinate axes and the coefficients a1 and b1 can
in turn.
be considered as functions of scale λ and rotation
α between the coordinate axes Z′
Z′′ ( Z′′′ ) Z

a1 = λ cos α
(8)
b1 = λ sin α
κ
Substituting (8) into (7), re-arranging and using
matrix notation gives the familiar 2D conformal
transformation Y′′′

⎡E ⎤ ⎡ cos α sin α ⎤ ⎡ X ⎤ ⎡ TE ⎤ Y′ ( Y′′ )

⎢N ⎥ = λ ⎢− sin α cos α ⎥ ⎢ Y ⎥ + ⎢ T ⎥ (9) ω
φ
⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ N⎦ X ( X′ )

where the coefficient matrix on the right-hand-side Y

is the rotation matrix Rα
X′′ X′′′

⎡ cos α sin α ⎤
Rα = ⎢ ⎥ (10)
⎣ − sin α cos α ⎦ Figure 3. 3D rotations ω, φ and κ

The 2D conformal transformation can be Rotations are considered as positive anti-clockwise

represented by the familiar diagram
when looking along the axis towards the origin; the
positive sense of rotations being determined by the
right-hand-grip rule where an imaginary right hand
grips the axis with the thumb pointing in the
positive direction of the axis and the natural curl of
the fingers indicating positive direction of rotation.

4
The three rotations in order are:
(i) Rotation of ω about the X-axis. This rotates
the Y and Z axis to Y ′ and Z ′ with the X
and X ′ axes coincident.
Coordinates in the new system will be given
by the matrix equation
⎡X ′⎤ ⎡1 0 0⎤ ⎡ X ⎤
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ Y ′ ⎥ = ⎢0 cos ω sin ω ⎥ ⎢ Y ⎥
⎢⎣ Z ′ ⎥⎦ ⎢⎣0 − sin ω cos ω ⎥⎦ ⎢⎣ Z ⎥⎦
(Rω )
⎡ c φc κ c ω sκ + sω sφc κ sω sκ − c ω sφc κ ⎤
⎢ ⎥
Rκ φ ω = ⎢ −c φsκ c ω c κ − sω sφsκ sωc κ + c ω sφsκ ⎥ (15)
⎢ sφ −sω c φ cωc φ ⎥
⎣ ⎦
where, for instance, c κ sφsω = cos κ sin φ sin ω .
Rotation matrices, e.g. Rα , R κ , R φ , Rω and
R κ φ ω are orthogonal, i.e. the sum of squares of the
elements of any row or column is equal to unity.
(11) They have the unique property that their inverse is
equal to their transpose, i.e. R −1 = R T which will
be used in later developments.

(ii) Rotation of φ about the new Y ′ axis. This

rotates the X ′ and Z ′ to X ′′ an d Z ′′ with THE 3D CONFORMAL TRANSFORMATION

the Y ′ and Y ′′ axes coincident. The 3D conformal coordinate transformation is an

extension of the 2D case; (14) and (15) combined
Coordinates in the new system will be given with a scale factor and translations to give (as
by the matrix equation previously stated)

⎡ X ′′ ⎤ ⎡cos φ 0 − sin φ⎤ ⎡ X ′ ⎤ ⎡E ⎤ ⎡X ⎤ ⎡ TE ⎤
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ Y ′′ ⎥ = ⎢ 0 1 0 ⎥ ⎢ Y′ ⎥ (12)
⎢N ⎥ = λ R κ φ ω ⎢ Y ⎥ + ⎢ TN ⎥ (1)
⎢⎣ Z ′′ ⎥⎦ ⎢⎣ sin φ 0 cos φ ⎥⎦ ⎢⎣ Z ′ ⎥⎦ ⎢⎣U ⎥⎦ ⎢⎣ Z ⎥⎦ ⎢⎣ TU ⎥⎦
(R φ )
To prove that this transformation is indeed
(iii) Rotation of κ about the new Z ′′ axis. This conformal, i.e. angles between points in the XYZ
rotates the X ′′ and Y ′′ to X ′′′ and Y ′′′ with survey system are preserved when transformed into
the Z ′′ and Z ′′′ axes coincident. the ENU design system, consider the following.

Coordinates in the new system will be given (i) Let three points a, b and c in the survey
by the matrix equation system be transformed into A, B and C in
the design system; points in both systems
⎡ X ′′′ ⎤ ⎡ cos κ sin κ 0⎤ ⎡ X ′′ ⎤ located by vectors a, b, c and A, B, C.
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ Y ′′′ ⎥ = ⎢− sin κ cos κ 0⎥ ⎢ Y ′′ ⎥ (13)
Using (1) with R as the rotation matrix and t
⎢⎣ Z ′′′ ⎥⎦ ⎢⎣ 0 0 1⎥⎦ ⎢⎣ Z ′′ ⎥⎦ as the vector of translations, we may write
( κ)
R the transformation of a to A as

The coefficient matrices R κ , R φ , Rω above are A = λRa + t (16)

3D rotation matrices which can be multiplied
together (in that order) to give another rotation and similarly for the other vectors.
matrix R κ φ ω (Mikhail and Moffitt 1980).

⎡ X ′′′ ⎤ ⎡X ⎤ ⎡X ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ Y ′′′ ⎥ = R κ R φ Rω ⎢ Y ⎥ = R κ φ ω ⎢ Y ⎥ (14)
⎢⎣ Z ′′′ ⎥⎦ ⎢⎣ Z ⎥⎦ ⎢⎣ Z ⎥⎦

with

5
 U C 2
• (a) u = uT u

B V and remembering that for orthogonal

• N rotation matrices R −1 = R T , hence
θ′
Z
U •A R T R = I where I is the identity matrix

Ru = ( Ru) Ru = u T R T Ru = u T u
c E
• (b)
2 T
b
•
θ v
u Equating (a) and (b) gives Ru = u which
•a
can be substituted into (18) to give
X Y
Ru u
U= =R = Ru (19)
u u
Figure 4. Conformal transformation
of two vectors.
and using similar reasoning
(ii) Let θ be the angle at a in the survey system
between the vectors u = b − a and V = Rv (20)
v = c − a ; and θ′ the complementary angle
at A in the design system between vectors (v) In the design system the angle θ′ between U
U = B − A and V = C − A . and V is given by the dot product
cos θ′ = U • V = U T V and using (19)
(iii) In the survey system, the angle between the and (20) we may write
vectors is found from the vector dot product
of unit vectors u and v
cos θ′ = ( Ru) Rv = u T R T Rv = u T v
T
(21)
cos θ = u • v = u 1v1 + u 2 v 2 + u 3 v3
Comparing (17) and (21) shows that θ = θ′ and
where u 1 , u 2 , u 3 and v1 , v 2 , v3 are the constitutes a proof that angles are preserved by the
transformation (1). [Baetslé (1966) has a similar
components of the two unit vectors, noting proof but in a slightly different form.]
u
that a unit vector is defined as u =
u Conformal mapping has a long mathematical
where u is the magnitude. history. Vlcek (1966) in a discussion paper on the
topic notes that the famous French mathematician
The dot product is equivalent to a matrix
Liouville (1847) determined all the conformal
multiplication, hence
transformations in an analytical way and in a
theorem bearing his name showed that the only
cos θ = u • v = u T v (17) ways of conformal representation of space on itself
are:
(iv) In the design system we may use (16) to
write 1. by translation and rotation, accompanied
by constant magnification,
U = B − A = λ R b + t − (λ R a + t) = λ R u 2. by inversion with respect to the sphere,
which, at the very least, demonstrates that
and the unit vector U conformal 3D transformations have been with us
for a long time.
λRu Ru
U = = (18)
λRu Ru

Now using matrix algebra we may write two

results (a) and (b)

6
THE 3D ROTATION MATRIX FOR SMALL ANGLES CENTROIDAL COORDINATES

For small angles δω , δφ , δκ the rotation matrix In a solution for the transformation parameters, the
R κ φ ω may be simplified by the approximations three translations TE , TN and TU can be
eliminated by adopting a system of "centroidal"
cos δω ≈ 1 coordinates, i.e. a system of coordinates whose
origin is at the centroid of the n control points.
sin δφ ≈ δφ (radians)
sin δκ sin δω ≈ 0 Denoting the coordinates of the centroid (in both
systems) with a subscript g
and (15) becomes the anti-symmetric (or skew-
symmetric) matrix (Harvey 1986). X1 + X 2 + + X n
Xg =
n
⎡ 1 δκ −δφ⎤ Y1 + Y2 + + Yn
⎢ ⎥ Yg = (24)
RS = ⎢− δκ 1 δω ⎥ (22) n
⎢⎣ δφ −δω 1 ⎥⎦ Z1 + Z 2 + + Z n
Zg =
n
It should be noted that RS is no longer orthogonal
and similarly for E g , N g and U g . Centroidal
but its inverse will, nevertheless, be given by its
transpose ( RS−1 = RST ), since it is the approximate coordinates (denoted with an over-bar) are

form of the orthogonal matrix R κT φ ω (Hotine 1969, X i = X i − Xg

p. 263).
Yi = Yi − Yg (25)
In the least squares development to follow it is Zi = Zi − Zg
useful to split RS (the rotation matrix for small
angles) into two parts with similar expressions for E , N and U .

⎡1 0 0⎤ ⎡ 0 δκ −δφ⎤ Thus, using centroidal coordinates [see (2), where

⎢ ⎥ ⎢ ⎥
RS = I + δR = ⎢0 1 0⎥ + ⎢ −δκ 0 δω ⎥ (23)
⎢⎣0 0 1⎥⎦ ⎢⎣ δφ − δω 0 ⎥⎦
Assuming small angles is a convenient and
practical technique of simplifying R κ φ ω (whose
non-independent elements are functions of the
rotation angles ω , φ and κ ) to RS (whose
independent elements are δω , δφ and δκ ), thus
enabling the solution of δω, δφ and δκ via a
system of linear equations (see the following
section on least squares solution). Solutions, based
on this assumption, will not be theoretically correct
but will be "practically correct" if an iterative
process is adopted where each n th iteration is
preceded by a transformation using previously
derived values, and the process terminated when
δω n , δφ n and δκ n converge to negligible values.
A practical test on the assumption of small angles
in the initial transformation will be revealed by
noting whether the solution converges.
the origins of both systems are common] the
transformation is reduced to a combination of scale
and rotation only and the size of the numbers
involved in the computations is reduced. After
solving for the scale factor λ and the elements of
the rotation matrix R κ φ ω , the translations can be
found by substituting the coordinates of the
centroid into (1) and re-arranging
⎡ TE ⎤ ⎡ Eg ⎤ ⎡X g ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ TN ⎥ = ⎢N g ⎥ − λ R κ φ ω ⎢ Yg ⎥ (26)
⎢⎣ TU ⎥⎦ ⎢U g ⎥ ⎢Zg ⎥
⎣ ⎦ ⎣ ⎦
LEAST SQUARES SOLUTION OF TRANSFORMATION
PARAMETERS
To develop the least squares solution of the
parameters, consider (2b) which pre-supposes that
the rotation angles are small which may be due to
(i) an initial transformation which approximately
aligns the coordinate axes, or (ii) a knowledge that

7
the rotations are small. Bearing this in mind ⎡δω ⎤
(noting that a method of determining the ⎢ δφ ⎥ ⎡Ei − λ′ Xi ⎤
⎢ ⎥
parameters of an initial transformation is discussed x = ⎢ ⎥, fi = ⎢N i − λ ′ Yi ⎥
⎢ δκ ⎥
later) we may use (23) to write for each of the n ⎢ ⎥ ⎢U i − λ ′ Z i ⎥
⎣ ⎦
control points an equation of the form ⎣ δλ ⎦

⎡E ⎤ ⎡ X ⎤ ⎡ vE ⎤ Equation (29a) is a "standard" least squares form

⎢ ⎥ ⎢ ⎥ ⎢ ⎥ (Mikhail 1976)
⎢N ⎥ = ( λ ′ + δλ ) ( I + δR) ⎢ Y ⎥ + ⎢ v N ⎥ (27)
⎢U ⎥ ⎢ Z ⎥ ⎢⎣ vU ⎥⎦
⎣ ⎦ ⎣ ⎦ v + Bx = f (29b)

where vE , v N and vU are residuals at the control with the solution for the parameters as
points. Now, since the rotation angles and δλ are
( )
−1
small, their products will be negligible x = BT WB B T Wf (30a)
( δλ δR ≈ 0 ), thus (27) can be expanded and re-
arranged as where W is the weight matrix associated with the
observations, which in this case are the triplets of
⎡δω ⎤ centroidal coordinates X i , Yi , Z i and E i , N i , U i .
⎡ vE ⎤ ⎡ 0 −Z Y X ⎤ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ δφ
⎢vN ⎥ + λ ′ ⎢ Z 0 −X Y ⎥ ⎢ ⎥ Mikhail (1976 pp. 64-66) defines weight matrices
⎢ δκ ⎥ W, cofactor matrices Q and variance-covariance
⎢⎣ vU ⎥⎦ ⎢− Y X 0 Z ⎥⎢ ⎥
⎣ ⎦ δλ matrices Σ as follows
⎣ ⎦ (28)
⎡E ⎤ ⎡X⎤
⎢ ⎥ ⎢ ⎥ W = Q −1 (30b)
= ⎢N ⎥ − λ′ ⎢ Y ⎥
⎢U ⎥ ⎢Z ⎥
⎣ ⎦ ⎣ ⎦
Σ = σ 02 Q (30c)
An equation of the form of (28) can be written for
each control point and represented symbolically in where the elements of Q are estimates of the
partitioned matrix form as variances and covariances of the observations and
σ 02 is the reference variance whose posteriori
⎡ v1 ⎤ ⎡ B1 ⎤ ⎡ f1 ⎤ estimate is (Mikhail 1976, p. 288)
⎢v ⎥ ⎢B ⎥ ⎢f ⎥
⎢ 2⎥ + ⎢ 2⎥x = ⎢ 2⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
(29a)
σ 02 =
v T Wv f Wf − x B Wf
=
T T T
( ) (30d)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ n−u n−u
⎣ v n ⎦ ⎣ Bn ⎦ ⎣fn ⎦
n and u are the number of observations and
where each component v i ( i = 1, 2, … , n ) of the
unknown parameters respectively.
vector of residuals v and Bi of the matrix of
coefficients B is The partitioned form of W in (30a) is

⎡ vE i ⎤ ⎡ 0 −Zi Yi X i ⎤ ⎡W1 0 0 ⎤
⎢ ⎥
⎢ ⎥
v i = ⎢ v N i ⎥ , Bi = λ ′ ⎢ Z i 0 − X i Yi ⎥
⎢0 W 0 ⎥
W=⎢
2 ⎥ (30e)
⎢⎣ vU i ⎥⎦ ⎢− Yi X i 0 Z i ⎥⎦ ⎢ ⎥
⎣ ⎢ ⎥
⎣0 0 Wn ⎦
and the vector of parameters x and the
components fi of the vector of numeric terms f where each diagonal element Wi contains weights
are associated with the coordinate triplets of each i th
control point, and the elements of Wi must be
obtained from a pre-analysis of the survey
techniques used to determine their values. The
off-diagonal elements of W are null matrices,

8
indicating that the observations (coordinate
triplets) are considered to be independent of each (ii) In the survey system, the angle θ between
other. the vectors is found from the dot product

In this paper, and the example following, it is cos θ = a • b = a1b1 + a2 b 2 + a3 b3 (33)

assumed that the coordinates (derived from survey
measurements) are independent of each other and
all having equal precision. This allows W to be where a1 , a2 , a3 and b1 , b 2 , b3 are the
replaced by the identity matrix I in (30a) to give components of the two unit vectors.

( ) A unit vector p , perpendicular to the plane

−1
x = BT B BT f (31)
containing a and b, can be obtained from
the vector cross product

CALCULATION OF APPROXIMATE ROTATIONS b×a

p = = p1 i + p 2 j + p 3 k (34)
sin θ
The solution above requires that rotations be small,
i.e. RS approximates R κ φ ω . To ensure this a where
preliminary (or initial) transformation is made
using centroidal coordinates and approximations b 2 a3 − b 3 a 2
p1 =
for the rotation matrix R A and scale factor λ A sin θ
b3 a1 − b1 a3
⎡X ⎤ ⎡X ⎤ p2 =
⎢ ⎥ ⎢ ⎥ sin θ
⎢Y⎥ = λ A RA ⎢ Y ⎥ (32)
⎢Z ⎥ ⎢Z ⎥ b1 a2 − b 2 a1
⎣ ⎦ ⎣ ⎦ p3 =
IN ITIAL O RIGIN AL sin θ

An approximate vale of λ A can be obtained by A second cross product (a × p) gives a third

ratios of distances in the survey and design systems,
unit vector q perpendicular to both a and
but for most practical applications will be very close
to unity. The approximate rotation matrix R A can p
be derived in the following manner.
q = a × p = q1 i + q 2 j + q 3 k (35)
(i) Shifting both the survey and design systems
to the centroid and choosing two control where
points A and B gives the vector pairs a and
b in the survey system, and A and B in the q 1 = a 2 p 3 − a3 p 2
design system whose components are the q 2 = a3 p1 − a1 p 3
centroidal coordinates of the respective q 3 = a1 p 2 − a2 p1
points. Each vector can be reduced to a unit
vector using the following: if Thus q , a and p are the unit vectors of
v = v1 i + v 2 j + v3 k is a vector whose
another orthogonal centroidal coordinate
components are v1 , v 2 , v3 ( i , j , k being unit system ξ( Xi) , η( Et a) and ζ( Zeta) .
vectors in the directions of the X, Y and Z- ζ
axes respectively) the unit vector
v A
v = = v1 i + v 2 j + v3 k has •
v η
v1 v v ξ
^
components v1 = , v 2 = 2 , v3 = 3 p
r r r
^
a ^
where v =r = v12 + v 22 + v32 is the • q

magnitude. Centroid

9
 Figure 5. ξ ηζ coordinate system where the left-hand-side of (39) can be
considered as an initial transformation of the
survey coordinates. Thus with an
(iii) The ξ ηζ and X Y Z centroidal systems,
approximate rotation matrix R A = R 2T R1
are rotated with respect to each other, the
ξ -axis making angles α1 , β1 , γ 1 with the and the approximate scale factor λ A we
have (32) as given above where
X , Y and Z -axes respectively. Similarly,
the η and ζ -axes make angles α 2 , β 2 , γ 2
and α 3 , β3 , γ 3 . The elements of q are the ⎡R 11 R 12 R 13 ⎤
⎢ ⎥
direction cosines cosα1 , cosβ1 and cos γ 1 RA = R 2T R1 = ⎢R 21 R 22 R 23 ⎥ (40)
(Wolf 1974) and ⎢⎣R 31 R 32 R 33 ⎥⎦

ξ = X cos α1 + Y cos β1 + Z cos λ1 (36a) R A has the same form as R κ φ ω (15) and values of
the approximate rotations ω A , φ A and κ A can be
or calculated from

ξ = X q1 + Y q 2 + Z q 3 (36b) −R 21
tan κ A = (41)
R 11
Similarly for the elements of a and p we
may write
R 11
cos φ A = (42)
η = X a1 + Y a 2 + Z a 3 cos κ ′
(36c)
ζ = X p1 + Y p2 + Z p3

Combining (36b) and (36c) in the form of a R 33 cos κ ′

cos ω A = (43)
rotation matrix R1 gives R 11

An example of the computation of approximate

⎡ξ ⎤ ⎡q1 q 2 q3 ⎤ ⎡ X ⎤ ⎡X⎤
⎢ ⎥ = ⎢a a a ⎥ ⎢ ⎥ = R ⎢ ⎥ rotations is contained in the Appendix.
⎢ η⎥ ⎢ 1 2 3⎥ ⎢ Y ⎥ 1⎢Y ⎥ (37)
⎢⎣ ζ ⎥⎦ ⎢⎣p1 p2 p3 ⎥⎦ ⎢⎣ Z ⎥⎦ ⎢Z ⎥
⎣ ⎦
CONCLUSION
(iv) Replacing survey system vectors a and b
with design system vectors A and B in steps This paper has presented the necessary equations
and computation technique for the practical use of
(ii) and (iii) yields unit vectors Q , A and P ,
3D conformal transformations in survey
and a second rotation matrix R 2 measurement. These transformations, combined
with least squares, can be usefully employed on
⎡ξ ⎤ ⎡Q1 Q 2 Q3 ⎤ ⎡ E ⎤ ⎡E ⎤ large construction projects where components,
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥ often manufactured "off-site", must be brought
⎢ η⎥ = ⎢ A1 A 2 A3 ⎥ ⎢ N ⎥ = R 2 ⎢ N ⎥ (38)
together "on-site" to fit within a design coordinate
⎢⎣ ζ ⎥⎦ ⎢P P P ⎥⎢ U ⎥ ⎢U ⎥
⎣ 1 2 3 ⎦⎣ ⎦ ⎣ ⎦ system. The least squares transformation method
can be used to compare the off-site component,
(v) Re-arranging (38) and substituting into (37), measured in situ in XYZ survey coordinates, with
using the orthogonal property of rotation its proposed location in the on-site ENU design
( )
matrices R 2−1 = R 2T , gives coordinates; comparison being via the coordinate
residuals at common points. This pre-analysis
"tool" can be used to prevent costly (and often
⎡E ⎤ ⎡X⎤ embarrassing) misalignment of components in a
⎢ ⎥ ⎢ ⎥ large engineering structure.
⎢N ⎥ = R 2T R1 ⎢ Y ⎥ (39)
⎢U ⎥ ⎢Z ⎥
⎣ ⎦ ⎣ ⎦

10
An example computation of the transformation
parameters and residuals of a simple plane figure Harvey, B.R. 1986, 'Transformation of 3D co-
ABC with measured XYZ coordinates and ENU ordinates', The Australian Surveyor, Vol. 33, No.
design coordinates is provided in the Appendix. 2, June 1986, pp. 105-125.

REFERENCES Helmert, F.R. 1880, Die mathematischen und

Baetslé, P.L. 1966, 'Conformal transformations in
three dimensions', Photogrammetric Engineering,
Vol. 32, No. 5, September 1966, pp. 816-824.
Bebb, G. 1981, 'The applications of
transformations to cadastral surveying',
Information-Innovation-Integration: Proceedings of
the 23rd Australian Survey Congress, Sydney,
March 28 – April 3, 1981, The Institution of
Surveyors Australia, pp. 105-117.
Bellman, C. and Anderson, L. 1995, 'Close range
photogrammetry for dimensional control in
shipbuilding', SAMS'95: Proceedings of the 3rd
Symposium on Surveillance and Monitoring
Surveys, Melbourne, Australia, November 1-2,
1995, ed, M.R. Shortis and C.L. Ogleby,
Department of Geomatics, University of
Melbourne, pp. 1-8.
Bellman, C. Deakin, R. and Rollings, N. 1992,
'Colour photomosaics from digitized aerial
photographs', Looking North: Proceedings of the
34th Australian Surveyors Congress, Cairns,
Queensland, May 23-29, 1992, The Institute of
Surveyors Australia, pp. 481-495.
Bervoets, S.G. 1992, 'Shifting and rotating a
figure', Survey Review, Vol. 31, No. 246,
October 1992, pp. 454-464.
Bird, D, 1984, 'Letters to the Editors: Least squares
reinstatement', The Australian Surveyor, Vol. 32,
No. 1, March 1984, pp. 65-66.
physikalischen Theorem der höheren Geodäsie,
Vol. 1, Die mathematischen Theorem, Leipzig.
Helmert, F.R. 1884, Die mathematischen und
physikalischen Theorem der höheren Geodäsie,
Vol. 2, Die physikalischen Theorem, Leipzig.
Hotine, M. 1969, Mathematical Geodesy, ESSA
Monograph 2, United States Department of
Commerce, Washington, D.C.
Jordan/Eggert/Kneissl, 1963, Handbuch der
Vermessungskunde (Band II), Metzlersche
Verlagsbuchhandlung, Stuttgart.
Lauf, G.B. 1983, Geodesy and Map Projections,
TAFE Publications, Collingwood.
Liouville, J. 1847, 'Note au sujet de l'article
precedent', Journal de Math. pures et appl., Vol.
12.
Mikhail, E.M. 1976, Observations and Least
Squares, IEP−A Dun-Donelley, New York.
Moffitt, F.H. and Mikhail, E.M. 1980,
Photogrammetry, 3rd ed, Harper & Row, New
York
Shmutter, B and Doytsher, Y. 1991, 'A new
method for matching digitized maps', Technical
Papers 1991 ACSM-ASPRS Annual Convention,
Baltimore, USA, Vol. 1, Surveying, pp. 241-246.

Blais, J.A.R. 1972, 'Three-dimensional

similarity', The Canadian Surveyor, Vol. 26, No.
1, March 1972, pp. 71-76.

DSB, 1972, Dictionary of Scientific Biography, Vol.

VI, C.C. Coulston Editor in Chief, Charles
Scribner's Sons, New York.

Featherstone, W.E. 1996, 'An updated explanation

of the Geocentric Datum of Australia and its
effects upon future mapping', The Australian
Surveyor, Vol. 41, No. 2, June 1996, pp. 121-
130.

11
Sprott, J. S. 1983, 'Least squares reinstatement', matrix Rκ φ ω with values κ = 10 ,
The Australian Surveyor, Vol. 31, No. 8,
φ = 94 , ω = 310 then (iii) rounding the
December 1983, pp. 543-556.
transformed centroidal coordinates to the
Vlcek, J. 1966, 'Discussion paper: Simultaneous nearest 0.1m and "adding back"
three-dimensional transformation', E g , N g and U g .
Photogrammetric Engineering, Vol. 32, No. 2,
March 1966, pp. 178-180. The solution for the transformation parameters is
accomplished in the following steps.
Wolf, P. R. 1974, Elements of Photogrammetry,
McGraw–Hill, New York. Step 1: Shift the origin of the survey system to
the centroid by subtracting X g , Yg and
APPENDIX
Z g from the coordinates of A B and C
Calculation of transformation parameters between then calculate the unit vectors a and b .
survey and design locations of the figure ABC.
A B C
X -240.000 380.000 -140.000
Survey system X, Y, Z
Y 20.000 -240.000 220.000
1620.0 X
B 740.0 Y Z 120.000 -540.000 420.000
340.0 Z
Table 1. Survey system centroidal coordinates
A
1000.0 X • 1240.0 X g
1000.0 Y Centroid 980.0 Y g a = −240 i + 20 j + 120 k
1000.0 Z 880.0 Z g
a = −0.891953 i + 0.074329 j + 0.445976 k
C
1100.0 X
b = 380 i − 240 j − 540 k
1200.0 Y b = 0.540874 i − 0.341605 j − 0.768610 k
1300.0 Z

Step 2: Use the vector dot product (33) to

calculate angle θ between a and b
Design system E, N, U
2540.6 E a • b = cosθ = a1 b1 + a2 b 2 + a3 b3
B 1668.1 N = − 0.850607457
1216.1 U
θ = 148.277801360
A
1911.9 E • 2000.0 E g
1435.2 N Centroid 1500.0 Ng
554.1 U 800.0 U g Step 3: Use the vector cross product (34) to
C calculate the unit vector p perpendicular
1547.5 E to the plane containing A, B and the
1396.7 N
629.8 U centroid

b×a
p = = p1 i + p 2 j + p 3 k
Figure A sin θ
= − 0.181090 i + 0.845086 j − 0.503027 k
Note: In Figure A the ENU design system
coordinates have been generated from the
Step 4: Use the vector cross product (35) to
XYZ survey system coordinates by (i)
calculate the unit vector q perpendicular
shifting to the centroid of ABC,
(ii) transforming to E , N and U using (2) to a and p
with a scale factor λ = 1 and rotation

12
 q = a × p = q1 i + q 2 j + q 3 k ⎡Q1 Q 2 Q3 ⎤
= − 0.414278 i − 0.529438 j − 0.740317 k ⎢ ⎥
R 2 = ⎢ A1 A 2 A3 ⎥
⎢P P P ⎥
⎣ 1 2 3⎦

Step 5: Form the rotation matrix R1 in (37) from ⎡ 0.933776 0.065506 −0.351812⎤
⎢
the elements of the unit vectors q , a and = ⎢ −0.327359 −0.240782 −0.913707 ⎥⎥
p ⎢⎣ −0.144563 0.968366 −0.203392⎥⎦

⎡ q1 q 2 q 3 ⎤ Step 9: Compute the numeric values in the

⎢ ⎥ approximate rotation matrix R A using
R1 = ⎢ a1 a 2 a 3 ⎥
⎢⎣p1 p2 p3 ⎥⎦ (40).

⎡ −0.414278 −0.529438 −0.740317 ⎤ R A = R 2T R1

⎢ ⎥
= ⎢ −0.891953 0.074329 0.445976⎥
⎢⎣ −0.181090 0.845086 −0.503027 ⎥⎦
= ⎢ 0.012267 0.765774 −0.642993⎥
Step 6: Shift the origin of the design system to
the centroid by subtracting E g , N g and
U g from the coordinates of A B and C
then calculate the unit vectors A and
B.
⎡−0.068675 −0.640878 −0.764565⎤
⎢ ⎥
⎢⎣ 0.997564 −0.053536 −0.044728⎥⎦
R A is of similar form to R κ φ ω and, if
required, values of the approximate
rotations ω A , φ A and κ A can be
calculated from it by using (41), (42) and
(43).

A B C −R 21 −0.012267
-88.100 540.600 -452.500 tan κ A = =
E R 11 −0.068675
N -64.800 168.100 -103.300
κ A = 190o 07 ′ 39′′
U -245.900 416.100 -170.200

Table 2. Design system centroidal coordinates

R 11
cos φ A =
A = −88.1 i − 64.8 j − 245.9 k cos κ ′

A = −0.327359 i − 0.240782 j − 0.913707 k φ A = 85o 59′ 59′′

B = 540.6 i + 168.1 j + 416.1 k
B = 0.769429 i + 0.239255 j + 0.592230 k R 33 cos κ ′
cos ω A =
R 11
ω A = 129o 52′ 40′′
Step 7: Use the unit vectors A and B in steps 2
and 3 to calculate the angle between A
and B, then the unit vector P followed Step 10: Use (32) with R A from Step 9 and
by the unit vector Q
λ A = 1 to transform the original X Y Z
centroidal coordinates (Table 1) to an
P = −0.144563 i + 0.968366 j − 0.203392 k
initial set of approximately transformed
Q = 0.933776 i + 0.065506 j − 0.351812 k centroidal coordinates to be used in the
least squares solution.
A B C
Step 8: Form the rotation matrix R 2 in (38) X -88.0833 540.5791 -452.4958
from the elements of the unit vectors Y -64.7877 168.0919 -103.3042
Q , A and P Z -245.8534 416.0761 -170.2227

13
 Table 3. Initial Survey system centroidal ⎡ vE A ⎤ ⎡ 0.011 ⎤
coordinates (Iteration 1) ⎢v ⎥ ⎢ 0.006 ⎥
⎢ EB ⎥ ⎢ ⎥
⎢ vE C ⎥ ⎢−0.016 ⎥
The least squares solution of δλ (correction to the ⎢ ⎥ ⎢ ⎥
approximate scale factor) δω , δφ and δκ (small ⎢vN A ⎥ ⎢ 0.009 ⎥
rotation angles) follows; noting that the ⎢ v N ⎥ = ⎢−0.002 ⎥ metres
⎢ B⎥ ⎢ ⎥
approximate scale factor is assumed to be unity ⎢vN C ⎥ ⎢−0.008 ⎥
(λ ′ = 1) . ⎢v ⎥ ⎢ 0.037 ⎥
⎢ UA ⎥ ⎢ ⎥
⎢ vU B ⎥ ⎢−0.011 ⎥
Step 11: With a re-arrangement of the order of ⎢v ⎥ ⎢−0.026 ⎥
residuals in (30) the coefficient matrix B ⎣ UC ⎦ ⎣ ⎦
becomes
and the transformed centroidal
⎡ 0 245.8534 −64.7877 −88.0833⎤ coordinates of A, B and C can be found
from (2b) with the elements of RS given
⎢ 0 −416.0761 168.0919 540.5791⎥
⎢ ⎥ above ( δω = 0.28′′ , δφ = −171
. ′′ ,
⎢ 0 170.2227 −103.3042 −452.4958⎥
⎢ ⎥ δκ = 0.40′′ ) and
⎢ −245.8534 0 88.0833 −64.7877 ⎥
λ ′ + δλ = 1.000 041 870 .
⎢ 416.0761 0 −540.5791 168.0919⎥
⎢ ⎥
⎢ −170.2227 0 452.4958 −103.3042⎥ A B C
⎢ 64.7877 −88.0833 0 −245.8534 ⎥ X -88.0892 540.6055 -452.5164
⎢ ⎥
⎢ −168.0919 540.5791 0 416.0761⎥ Y -64.7906 168.0985 -103.3079
⎢ 103.3042 −452.4958
⎣ 0 −170.2227 ⎥⎦ Z -245.8635 416.0888 -170.2259

and the vector of numeric terms is Table 4. Survey system centroidal coordinates
(Iteration 2)
⎡−0.0167 ⎤
⎢ 0.0209⎥ Step 14: A second iteration using the centroidal
⎢ ⎥ coordinates of Table 4 gives solutions as
⎢ −0.0042⎥
⎢ ⎥
⎢ −0.0123⎥ ⎡δω ⎤ ⎡ 3.27 × 10−8 ⎤ radians ⎡0.01 ⎤ sec
f = ⎢ 0.0081⎥ ⎢ δφ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 3.22 × 10−9 ⎥ radians ⎢0.01 ⎥ sec
⎢ ⎥ x= = =
⎢ 0.0042⎥ ⎢ δκ ⎥ ⎢ 176. × 10−8 ⎥ radians ⎢0.00 ⎥ sec
⎢ −0.0466⎥ ⎢ ⎥ ⎢ −8 ⎥ ⎢ ⎥
⎢ ⎥ ⎣ δλ ⎦ ⎢⎣ −3.13 × 10 ⎥⎦ ⎣0.031⎦ ppm
⎢ 0.0239⎥
⎢ 0.0227 ⎥ These solutions, which are at least an order of
⎣ ⎦
magnitude less than the corrections of the first
iteration (demonstrating that the solution is
Step 12: Forming the matrix products B T B converging) are probably due to rounding errors in
(normal coefficient matrix) and B T f the calculations and the first iteration results can
(vector of numeric terms) and solving be assumed as "exact".
(31) gives the solution as To transform any additional points in the XYZ
survey system to the ENU design system the
⎡δω ⎤ ⎡ 1.38 × 10−6 ⎤ radians ⎡ 0.28 ⎤ sec following process should be adopted:
⎢ δφ ⎥ ⎢ ⎥ ⎢ . ⎥ sec
−8.30 × 10−6 ⎥ radians ⎢−171
x=⎢ ⎥=⎢ = ⎥
(a) convert to centroidal coordinates then
⎢ δκ ⎥ ⎢ 1.93 × 10−6 ⎥ radians ⎢ 0.40 ⎥ sec
⎢ ⎥ ⎢ −5 ⎥ ⎢ ⎥ perform an initial transformation as per Step
⎣ δλ ⎦ ⎢⎣4.1870 × 10 ⎥⎦ . ⎦ ppm
⎣ 41870 10 using (32) with the numerical values of
R A as determined in Step 9,

Step 13: Substituting the vector x into (30) gives

the residuals at A, B and C

14
 ⎡X ⎤ ⎡X ⎤
⎢ ⎥ ⎢ ⎥
⎢Y⎥ = λ A RA ⎢ Y ⎥ (32)
⎢Z ⎥ ⎢Z ⎥
⎣ ⎦ IN ITIAL ⎣ ⎦ O RIGIN AL

(b) then transform to the ENU system using

(2b) modified by setting the approximate
scale factor to unity ( λ ′ = 1) and adding
back the coordinates of the centroid in the
design system.

⎡E ⎤ ⎡ X ⎤ ⎡ Eg ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢N ⎥ = (1 + δλ ) RS ⎢ Y ⎥ + ⎢N g ⎥
⎢⎣U ⎥⎦ ⎢ Z ⎥ ⎢U g ⎥
⎣ ⎦ ⎣ ⎦

with the values for δω , δφ , δκ and δλ

from the least squares solution as per Step 12

15