Вы находитесь на странице: 1из 19

Quaternions and Rotations

(Com S 477/577 Notes)

Yan-Bin Jia
Nov 24, 2014

Introduction

Up until now we have learned that a rotation in R3 about an axis through the origin can be represented by a 3 3 orthogonal matrix with determinant 1. However, the matrix representation seems
redundant because only four of its nine elements are independent. Also the geometric interpretation
of such a matrix is not clear until we carry out several steps of calculation to extract the rotation
axis and angle. Furthermore, to compose two rotations, we need to compute the product of the
two corresponding matrices, which requires twenty-seven multiplications and eighteen additions.
Quaternions are very efficient for analyzing situations where rotations in R3 are involved. A
quaternion is a 4-tuple, which is a more concise representation than a rotation matrix. Its geometric meaning is also more obvious as the rotation axis and angle can be trivially recovered. The
quaternion algebra to be introduced will also allow us to easily compose rotations. This is because
quaternion composition takes merely sixteen multiplications and twelve additions.
The development of quaternions is attributed to W. R. Hamilton [5] in 1843. Legend has it that
Hamilton was walking with his wife Helen at the Royal Irish Academy when he was suddenly struck
by the idea of adding a fourth dimension in order to multiply triples. Excited by this breakthrough,
as the couple passed the Broome Bridge of the Royal Canal, he carved the newfound quaternion
equations
i2 = j 2 = k2 = ijk = 1
into the stone of the bridge. This event is marked by a plaque at the exact location today. Hamilton
spent the rest of his life working on quaternions, which became the first non-commutative algebra
to be studied.

Quaternion Algebra

The set of quaternions, together with the two operations of addition and multiplication, form a
non-commutative ring.1 The standard orthonormal basis for R3 is given by three unit vectors
Sections 2.1, 2.2, 3, and 4 are based on Chapters 36 of the book [9] by J. B. Kuipers, Sections 1 and 6 are
partially based on the essay by S. Oldenburger [10] who took the course, and Appendix B is based on [6].
1
For the purpose of this course, you dont really need to know what a ring is although it can be found in a standard
algebra text such as the one by Hungerford [7] or Jacobson [8].

i = (1, 0, 0), j = (0, 1, 0), k = (0, 0, 1). A quaternion q is defined as the sum of a scalar q0 and a
vector q = (q1 , q2 , q3 ); namely,
q = q0 + q = q0 + q1 i + q2 j + q3 k.

2.1

Addition and Multiplication

Addition of two quaternions acts component-wise. More specifically, consider the quaternion q
above and another quaternion
p = p0 + p1 i + p2 j + p3 k.
Then we have
p + q = (p0 + q0 ) + (p1 + q1 )i + (p2 + q2 )j + (p3 + q3 )k.
Every quaternion q has a negative q with components qi , i = 0, 1, 2, 3.
The product of two quaternions satisfies these fundamental rules introduced by Hamilton:
i2 = j 2 = k2 = ijk = 1,

ij = k = ji,

jk = i = kj,
ki = j = ik.

Now we can give the product of two quaternions p and q:


pq = (p0 + p1 i + p2 j + p3 k)(q0 + q1 i + q2 j + q3 k)
= p0 q0 (p1 q1 + p2 q2 + p3 q3 ) + p0 (q1 i + q2 j + q3 k) + q0 (p1 i + p2 j + p3 k)
+(p2 q3 p3 q2 )i + (p3 q1 p1 q3 )j + (p1 q2 p2 q1 )k.

(1)

Whew! It is too long to remember or even to understand what is going on. Fortunately, we can
utilize the inner product and cross product of two vectors in R3 to write the above quaternion
product in a more concise form:
pq = p0 q0 p q + p0 q + q0 p + p q.

(2)

In the above, p = (p1 , p2 , p3 ) and q = (q1 , q2 , q3 ) are the vector parts of p and q, respectively.
Example 1. Suppose the two vectors are given as follows:
p
q

= 3 + i 2j + k,

= 2 i + 2j + 3k.

We single out their vector parts p = (1, 2, 1) and q = (1, 2, 3) and calculate their inner and cross products:
pq

pq

=
=

2,


i
j k

1 2 1


1 2 3
8i 4j.

By (2) the quaternion product is then


pq

= 6 (2) + 3(i + 2j + 3k) + 2(i 2j + k) + (8i 4j)


= 8 9i 2j + 11k.

We see that the product of two quaternions is still a quaternion with scalar part p0 q0 p q and
vector part p0 q + q0 p + p q. The set of quaternions is closed under multiplication and addition.
It is not difficult to verify that multiplication of quaternions is distributive over addition. The
identity quaternion has real part 1 and vector part 0.

2.2

Complex Conjugate, Norm, and Inverse

Let q = q0 + q = q0 + q1 i + q2 j + q3 k be a quaternion. The complex conjugate of q, denoted q , is


defined as
q = q0 q = q0 q1 i q2 j q3 k.
From the definition we immediately have
(q ) = q0 (q) = q,

q + q = 2q0 ,

q q = (q0 q)(q0 + q)

= q0 q0 (q) q + q0 q + (q)q0 + (q) q

q02
q02

by (2)

+qq

+ q12 + q22 + q32

= qq .
Given two quaternions p and q, we can easily verify that
(pq) = q p .

(3)

The norm of a quaternion q, denoted by |q|, is the scalar |q| = q q. A quaternion is called a
unit quaternion if its norm is 1. The norm of the product of two quaternions p and q is the product
of the individual norms, for we have
|pq|2 = (pq)(pq)
= pqq p

= p|q|2 p
= pp |q|2

= |p|2 |q|2 .
The inverse of a quaternion q is defined as
q 1 =

q
.
|q|2

We can easily verify that q 1 q = qq 1 = 1. In the case q is a unit quaternion, the inverse is its
conjugate q .
3

2.3

Power, Exponential, and Logarithm

Let us first look at a unit quaternion q = q0 + q. That q02 + kqk2 = 1 implies that there exists a
unique [0, ] such that cos = q0 and sin = kqk. The quaternion can thus be rewritten in
= q/kqk:
terms of and the unit vector u
sin .
q = cos + u
A general quaternion q = q0 + q can be represented as a unit quaternion scaled by the norm |q|:
sin ),
q = |q|(cos + u

(4)

= q/|q| and = arccos(q0 /|q|). Eulers identity for a complex number


where u
p
a + bi = a2 + b2 ei ,

where i2 = 1 and = atan2(b, a), generalizes to the quaternion q in a way that (4) can be
rewritten as
q = |q|eu .
This allows us to define the power of q as
 
sin()),
q = |q| eu = |q| (cos() + u

R.

(5)

Intuitively, the power is taken over the norm of the quaternion while a scaling is performed on its
phase angle.
An exponential of q makes use of the Taylor expansion that treats q just as an ordinary variable:
eq =

i
X
q
i=0

i!

The sum on the right hand side has a closed form that transforms the above into
kqk)
eq = exp(q0 + u

(6)

q0

sin kqk) .
= e (cos kqk + u

(7)

The logarithm of q is accordingly defined as


arccos
ln q = ln |q| + u

q0
|q|

The two operations are inverses of each other as we can verify


eln q = eln |q|+u arccos(q0 /|q|)
= |q|eu arccos(q0 /|q|)


kqk
q0

+u
= |q|
|q|
|q|
= q.

(8)

Pure Quaternions
R3
v =0+v

R4

v
Quaternions

Figure 1: R3 is viewed as the space of pure quaternions.

Quaternion Rotation Operator

How can a quaternion, which lives in R4 , operate on a vector, which lives in R3 ? First, we note
that a vector v R3 is a pure quaternion whose real part is zero.
Using the unit quaternion q we define an operator on vectors v R3 :
Lq (v) = qvq
= (q02 kqk2 )v + 2(q v)q + 2q0 (q v).

(9)

Here we make two observations. First, the quaternion operator (9) does not change the length of
the vector v for
kLq (v)k = kqvq k

= |q| kvk |q |

= kvk.

Second, the direction of v, if along q, is left unchanged by the operator Lq . To verify this, we let
v = kq and have
qvq = q(kq)q
= (q02 kqk2 )(kq) + 2(q kq)q + 2q0 (q kq)
= k(q02 + kqk2 )q
= kq.

Essentially, any vector along q is thus not changed under Lq . This makes us guess that the operator
Lq acts like a rotation about q, which will be made precise by the next theorem.
Before proceeding with the theorem, we remark that the operator Lq is linear over R3 . For any
two vectors v1 , v 2 R3 and any a1 , a2 R we can show that
Lq (a1 v 1 + a2 v 2 ) = a1 Lq (v 1 ) + a2 Lq (v 2 ).
Theorem 1 For any unit quaternion
q = q0 + q = cos
5

sin ,
+u
2
2

(10)

and for any vector v R3 the action of the operator


Lq (v) = qvq
as the axis of rotation.
on v is equivalent to a rotation of the vector through an angle about u
Proof Given a vector v R3 , we decompose it as v = a + n, where a is the component along
the vector q and n is the component normal to q. Then we show that under the operator Lq , a is
invariant, while n is rotated about q through an angle . Since the operator is linear, this shows
that the image qvq is indeed interpreted as a rotation of v about q through an angle .
We know from an early reasoning that a is invariant under Lq . So let us focus on the effect of
Lq on the orthogonal component n. We have
Lq (n) = (q02 kqk2 )n + 2(q n)q + 2q0 (q n)
= (q02 kqk2 )n + 2q0 (q n)

u n),
= (q02 kqk2 )n + 2q0 kqk(

= q/kqk. Denote n = u
n. So the last equation
where in the last step above we introduced u
becomes
Lq (n) = (q02 kqk2 )n + 2q0 kqkn .
(11)
Note that n and n have the same length:
k = knk k
kn k = kn u
uk sin

= knk.
2

Finally, we rewrite (11) into the form







2
2
sin
n + 2 cos sin
n
Lq (n) =
cos
2
2
2
2
= cos n + sin n .
Namely, the resulting vector is a rotation of n through an angle in the plane defined by n and
n . See the figure below. This vector is clearly orthogonal to the rotation axis.

Note that the quaternion negation q, when applied to v, will result in the same vector:
Lq = (q)v(q) = qvq . It describes the rotation about an axis in the opposite direction
u
through the negative amount .
We substitute the unit quaternion form (10) into (9) to obtain the resulting vector from rotating
through :
a vector v about the axis u







2
2
sin + 2 cos
sin v u
sin v
u
sin
v+2 u
Lp (v) =
cos
2
2
2
2
2
2
= cos v + (1 cos )(
u v)
u + sin (
u v).
(12)
Example 2. Consider a rotation about an axis defined by (1, 1, 1) through an angle of 2/3. About this
axis, the basis vectors i, j, and k generate the same cone when rotated through 2. We define a unit vector
1
= (1, 1, 1).
u
3

n
Lq (n)

Figure 2: Quaternion acts as rotation.


Let the rotation angle = 2/3. The quaternion q defining the rotation is then given as
q

=
=

sin
+u
2
2
1
1
1 1
+ i + j + k.
2 2
2
2

cos

Let us compute the effect of rotation on the basis vector i = (1, 0, 0). We obtain the resulting vector
using (12):



1
1
1
1
1
1
1
1
3 1
1
1 0
0

+

+ 1+
v =
2
2
2
3
3
3
1
1
0
0

1
1
0
2
2
1 1

+ 2 + 2
=
0
1
0
21
2
= j.

The rotation of v under the operator Lq can also be interpreted from the perspective of an
observer attached to the vector v. What he sees happening is that the coordinate frame rotates
through the angle about the same axis defined by the quaternion.
Theorem 2 For any unit quaternion
q = q0 + q = cos

sin ,
+u
2
2

and for any vector v R3 the action of the operator


Lq (v) = q v(q ) = q vq
through an angle while v is not rotated.
is a rotation of the coordinate frame about the axis u
Equivalently, the operator Lq rotates the vector v with respect to the coordinate frame through
an angle about q.
7

The quaternion operator Lq (v) = qvq may be interpreted as a point or vector rotation with
respect to the (fixed) coordinate frame. The quaternion operator Lq (v) = q vq may be interpreted
as a coordinate frame rotation with respect to the (fixed) space of points.
We end this section with a result that the inner product of two vectors is not changed if these
vectors are left multiplied with a unit quaternion to become 4-tuples. Here, the inner product of
two quaternions p = p0 + p and q = q0 + q is defined to be
p q = p0 q0 + p q.
sin 2 be a unit quaternion. Then
Proposition 3 Let a, b R3 be two vectors, and q = cos 2 + u
a b = (qa) (qb) = (aq) (bq).
Proof Since a and b are pure quaternions 0+a and 0+b, we left multiply q with them respectively
according to (1):

a sin ,
+ a cos + u
2
2
2

b sin .
qb =
u b sin + b cos + u
2
2
2

qa =
u a sin

Based on the above, we calculate their inner product as 4-tuples:

+ (a b) cos2 + a (
u b) cos sin
2
2
2
2

u a) (
u b) sin2
+ b (
u a) cos sin + (
2
2
2

2
2
+ (a b) cos
+ (
u a) (
u b) sin2
= (
u a)(
u b) sin
2
2
2

= (
u a)(
u b) sin2 + (a b) cos2 + (a b) sin2 (
u a)(
u b) sin2
2
2
2
2
= a b.

(qa) (qb) = (
u a)(
u b) sin2

Similary, we can show that a b = (aq) (bq).

Quaternion Operator Sequences

Let p and q be two unit quaternions. We first apply the operator Lp to the vector u and obtain
the vector v. To v we then apply the operator Lq and obtain the vector w. Equivalently, we apply
the composition Lq Lp of the two operators:
w = Lq (v)
= qvq
= q(pup )q
= (qp)u(qp)
= Lqp (u).
8

Because p and q are unit quaternions, so is the product qp. Hence the above equation describes a
rotation operator whose defining quaternion is the product of the two quaternions p and q. The
axis and angle of the composite rotation is given by the product qp.
Similarly, consider the quaternion operators Lp (u) = p up and Lq (v) = q vq which carry
out rotations of the coordinate system determined by quaternions p and q, respectively. Then the
quaternion product pq defines an operator L(pq) , which represents a sequence of operators Lp
followed by Lq . The axis and angle of rotation of L(pq) are those represented by the quaternion
product pq.
Example 3. We now use the quaternion method to find the axis and angle of the composite rotation in
the Satellite tracking example from the notes titled Space Rotations. Recall that the tracking application
takes a rotation about the z-axis through a bearing angle followed by a rotation about the new y-axis
through an elevation angle . After these two rotations, the new x-axis points toward the satellite. The two
rotations are respectively described by the two quaternions below:
p
q

+ sin k,
2
2

= cos + sin j.
2
2
= cos

Since we are rotating the coordinate frame, the two operators Lp and Lq are applied sequentially. The
composite rotation operator is L(pq) , which transforms coordinates in the station frame to those in the
tracking frame. And the quaternion describing the composition rotation is the product pq which is as
follows.




pq =
cos + sin k
cos + sin j
2
2
2
2

= cos cos + cos sin j + sin cos k + sin sin (k j)


2
2
2
2
2
2
2
2

= cos cos sin sin i + cos sin j + sin cos k.


2
2
2
2
2
2
2
2
The axis of the composite rotation is defined by the vector



v = sin sin , cos sin , sin cos


.
2
2
2
2
2
2
And the angle of rotation satisfies

sin
2

cos

cos ,
2
2

cos

kvk.

The cosine is same as obtained in Section 3 of the handouts titled Rotation in the Space for we have
cos

1
2

= 2 cos2 cos2 1
2
2
cos + 1 cos + 1
= 2

1
2
2
cos cos + cos + cos 1
.
=
2
= 2 cos2

(13)

Note that the rotation axis and angle in that section transforms coordinates in the tracking frame to those
in the station frame. This explains why the axis v in (13) is opposite to the one obtained in that section
while the angle is the same.

Unit Quaternion Differentiation and Integration

The unit quaternion function q(t) describes the varying rotation of a frame (e.g., attached to some
object) with time t in reference to a fixed world frame. It determines the angular velocity (t) of
the frame with respect to the world frame. How to characterize the changing rate of q(t), that is,
its derivative q(t)?

Theorem 4 Let q(t) be a unit quaternion function, and (t) the angular velocity determined by
q(t). The derivative of q(t) is
1
(14)
q = q.
2
Proof At t + t, the rotation is described by q(t + t). This is after some extra rotation during
t performed on the frame that has already undergone a rotation described by q(t). This extra
= /kk through the angle = kkt. It can be
rotation is about the instantaneous axis
described by a quaternion:

sin
+
2
2
kkt
kkt
sin
= cos
+
.
2
2

q = cos

(15)

The rotation at t + t is thus described by the quaternion sequence q(t), q, implying


q(t + t) = q q(t).
We are now ready to derive q(t).

First, let us obtain the difference




kkt
kkt
sin
q(t + t) q(t) =
cos
+
q(t) q(t)
2
2
kkt
kkt
sin
q(t) +
q(t).
= 2 sin2
4
2

(16)

(by (15) and (16))

The first term in the last equation above is of higher order than t, thus its ratio to t goes to
zero as the latter does. Hence
q(t)

=
=
=
=
=

q(t + t) q(t)
t
sin(kkt/2)
lim

q(t)
t0
 t
kkt
d

sin

q(t)
dt
2
t=0
kk

q(t)

2
1
(t)q(t).
2
lim

t0

10

(17)

(18)

In mechanics, the angular velocity is often in terms of the rotated frame (which is the body
frame of a moving object). Denote this angular velocity by . Thus, = q q. Substituting the
expression = q q into (14), we obtain
1
(19)
q = q
2
If q is known , we can recover the angular velocity from (14) by right multiplying its both sides
with q :
= 2qq
.
(20)
The second derivative of the quaternion follows from differentiating (14):
1
+ q)
(q

2
1
1
+ q
q
=
2
4


1
1
2
=
kk + q.
4
2

q =

(21)
(by (14))

We can also recover the angular acceleration if the first and second derivatives of q are both known.
This is done by right multiplying (21) with q :
= 2
q q qq

= 2
q q 2qq
qq

(by (20))

= 2(
q q (qq
) ).

be the angular velocity of the rotating frame with respect to a fixed frame that instantaLet
neously coincides with it. It is obtained from by rotating the world frame to coincide with the
rotating frame according to q. This establishes
= q q.

Combining (14) with the above, we have another expression of the quaternion derivative:
1

q = q .
2
The differential equation (14) can be solved via numerical integration. Let h be the time step
size, and denote by qk and k the quaternion and angular velocity at the time kh. Eulers method
approximates the quaternion at the next time step by
1
qk+1 = qk + h k qk .
2
Due to the increment, the length of qk+1 will be different from unity. It is then normalized:
qk+1
qk+1
.
kqk+1 k

Eulers method is of first order and known to be inaccurate due to the truncation error, which
will propagate to the subsequent normalization. Standard integration methods of higher order
such as Adams-Bashforth and Runge-Kutta [11] can be employed. Special integration methods for
quaternions have also been developed, and shown to be more effective. We refer to [14] for a survey
of these methods with performance comparisons.
11

Applications

In physics, quaternions are correlated to the nature of the universe at the level of quantum mechanics. They lead to elegant expressions of the Lorentz transformations, which form the basis of
the modern theory of relativity. In signal processing, Quaternion Fourier Transform (QFT) is a
powerful tool. The QFT restores the lost commutative property at the cost of no longer being a
division algebra. It can be used, for instance, to embed a watermark in a color image. Other applications of QFT include face recognition (jointly with Quaternion Wavelet Transform) and voice
recognition [10].
Appendix B describes an application of quaternions in optimally matching a set of data points
against a shape model.
Homogeneous coordinates are introduced to make translation multiplicative, along with scaling
and rotation. They are convenient in representing points, lines, and planes, and fundamental for
studying projections. Like quaternions, homogeneous coordinates are 4-tuples. This suggests that
there might be a way of doing scaling and translation using some sort of quaternion operator. As
of now, no such way has been found as quaternions and their rotation operators are algebraically
incompatible with homogeneous coordinates.
In 1873, quaternions were extended to dual quaternions by Clifford [2] to represent both rotations and translations. Dual quaternions have found applications in kinematics, robotics, motion
estimation, and computer graphics.

Quaternion Interpolation

In computer graphics and animation, there is often a need to interpolate between an objects initial
orientation (i.e., a rotation of the body frame with respect to the world frame) and final orientation
to generate a smooth rotating motion. Let the the two rotations be represented respectively by the
following two unit quaternions:
1
1
1 sin ,
+u
2
2
2
2
2 sin ,
= cos
+u
2
2

r1 = cos
r2

i is a unit vector represents the axis of the ith rotation, and i the corresponding
where for i = 1, 2, u
rotation angle. For interpolation to be meaningful, r1 6= r2 must hold.

A.1

Constant Change Rates in Rotation Axis and Angle

It is easy to interpolate the rotation angle linearly as


( ) = (1 )1 + 2 ,

(22)

1 and u
2 would yield the
where [0, 1]. However, linear interpolation between the unit vectors u
2 that is not unit. If we simply normalize it as w = v/kvk, the resulting
vector v = (1 )
u1 + u
curve w( ) is not constant speed in terms of . This is often not desired or visually appealing as
1 to u
2.
the object may seem to be rotating unstably from u

12

1 and u
2 lie on the unit sphere, it is natural to interpolate them using their shortest path
Since u
1 and u
2 , as illustrated in
on the sphere. This is the shorter one of the two great arcs connecting u
the figure below.

u2

u( )
u1

( ) moving at constant speed on this great arc from u


1 to u
2 as increases
Picture a point u
( ) is a constant speed parametrization of the arc over [0, 1]. To derive
from 0 to 1. Essentially, u
it, we first construct the normal to the plane:
n=

1 u
2
u
.
2k
k
u1 u

(23)

1 =
1 =
In the case u
u2 , we may simply pick the vertical plane containing them. Denoting u
(ux , uy , uz ), the vertical plane has the normal
(uy , ux , 0)
n= q
u2x + u2y

if u2x + u2y 6= 0, and otherwise n = (1, 0, 0) by choice.


1 to u
2 about n is in (0, ]. It can be easily
Due to the choice of n, the rotation angle from u
obtained s
2 ).
= arccos(
u1 u
(24)
( ) is determined from a rotation of u
1 about n through the angle , as a
Then the vector u
quaternion product:

 


( ) = cos
1 cos
u
u
+ n sin
n sin
.
2
2
2
2
has a simpler form not involving quaternion multiplications:
In fact, u
( ) =
u

sin((1 ))
sin( )
1 +
2.
u
u
sin
sin

(25)

1 and u
2 lie
The correctness of the above expression can be first established for the case that u
1 = (cos 1 , sin 1 ). Then u
2 = (cos(1 +
in the xy-plane, and n is along the z-direction. Let u
13

), sin(1 + )). We have


sin( )
sin((1 ))
1 +
2
u
u
sin
sin


sin((1 )) cos 1 + sin( ) cos(1 + ) sin((1 )) sin 1 + sin( ) sin(1 + )
=
,
sin
sin


sin (cos( ) cos 1 sin( ) sin 1 ) sin (cos( ) sin 1 + sin( ) cos 1 )
=
,
sin
sin
= (cos(1 + ), sin(1 + ))
( ).
= u
1 and u
2 do not lie in the xy-plane, we rotate n to coincide with the z-axis. Let R be the
If u
corresponding rotation matrix. Then
( ) = R1 (R
u
u)


sin((1
))
sin( )
1
= R
(R
u1 ) +
(R
u2 )
sin
sin
sin((1 ))
sin( )
1 +
2.
=
u
u
sin
sin
Finally, we can interpolate between r1 and r2 over [0, 1]:
( )
( )
( ) sin
+u
2
2

sin((1 ))
sin( )
(1 )1 + 2
(1 )1 + 2
1 +
2 sin
+
,
u
u
= cos
2
sin
sin
2

r( ) = cos

(26)

by (22) and (25), where is given in (24). The interpolation has constant change rates in both the
rotation angle and the axis.

A.2

Spherical Linear Interpolation

In computer graphics, the widely used algorithm Slerp (spherical linear interpolation) [13] takes
the following form
r( ) = r1 (r1 r2 ) ,
[0, 1],
(27)
where the power of a unit quaternion is given by (5).
It has been shown [3] that r( ) parametrizes the shortest path connecting r1 and r2 on the 3D
unit quaternion sphere in the 4D space. A major appeal is that interpolation is carried out as a
rotation about a fixed axis at constant angular velocity.

Application: 3-D Shape Registration

An important problem in model-based recognition is to find the transformation of a set of data


points that yields the best match of these points against a shape model. The process is often
referred to as data registration. The data points are typically measured on a real object by range
sensors, touch sensors, etc., and given in Cartesian coordinates. The quality of a match is often
14

q1
q5

rotation
p5

p1

q7

translation

q3

q4

p3

q2
q6

p7
p4

p2
p6

Model

Data
Figure 3: Matching two point sets pi and q j .

described as the total squared distance from the data points to the model. When multiple shape
models are possible, the one that results in the least total distance is then recognized as the shape
of the object.
Quaternions are very effective in solving the above least-squares-based registration problem. Let
us begin with a formulation of the problem in 3D. Let {p1 , p2 , . . . , pn } be a set of data points. We
assume that p1 , . . . , pn are to be matched against the points q 1 , . . . , q n on a shape model. Namely,
the correspondences between the data points and those on the model have been predetermined.
Then the problem is to find a rotation, represented by an orthogonal matrix R with det(R) = 1,
and a translation b as the solution to the following minimization:
min
R,b

n
X
i=1

kRpi + b q i k2 .

We begin by computing the centroids of the two sets of points:


n

=
p

1X
pi ;
n
i=1

=
q

n
1X
qi.
n
i=1

The relative coordinates of all the points to their centroids are obtained as, for 1 i n,
;
pi = pi p
qi = q i q .

15

(28)

Clearly, we have
n
X

i=1
n
X

pi =
q i =

n
X

i=1
n
X
i=1

i=1

n
X

pi n
p =

i=1
n
X

qi n
q =

i=1

i=1

kRpi + b q i k2 =
=
=

n
X

i=1
n
X

i=1
n
X
i=1

n
X
i=1

n
X
i=1

1
n

qi n

Let us rewrite the objective function in (28) in terms of


n
X

1X
pi = 0;
n

pi n

i=1
n
X

q i = 0.

(29)
(30)

i=1

p, q, pi , q i :

kRpi q i + R
p q + bk2
+ b)
(Rpi q i + R
p q + b) (Rpi q i + R
pq
kRpi

q i k2

n
X

(Rpi

i=1

kRpi q i k2 + 2 R

n
X
i=1

pi

q i )

(R
p q + b) + nkR
p q + bk2

n
X

i=1

q i

kRpi q i k2 + nkR
p q + bk2 ,

+ bk2
(R
p q + b) + nkR
pq
by (29) and (30).

The minimizing translation b should make the second term in the last equation above zero, yielding:
R
b=q
p.

(31)

Thus we have decomposed the problem of data registration into two phases: the first of which
determines its optimal translation, as given by equation (31), and the second of which determines
) + q
the optimal rotation of the set {pi }. Note that every point pi is transformed into R(pi p
before matching against q i . Equivalently, to find the best match of the two point sets {pi } and
{q i }, we first translate {pi } to let their centroid coincide with that of {q i }, and then rotate about
the common centroid.
By the reasoning so far, the optimal rotation can be solved from the formulation below:
min
R

n
X
i=1

kRpi q i k2 .

(32)

Here we present an exact solution to (32) as described in [6] using quaternions. An equivalent
quaternion-based solution is given in [4]. The version of matching two curves (or surfaces), also
assuming pointwise correspondences, is solved exactly in [12] in a somewhat similar manner without
the use of quaternions.
First, we rewrite the summation in (32) as follows:
n
X
i=1

kRpi

q i k2

n
n
n
X
X
X

q i q i
(Rpi q i ) +
(Rpi Rpi ) 2
=

i=1

i=1

i=1

i=1

n
X


kpi k2 + kq i k2 2
16

n
X
i=1

Rpi q i .

The first summand in the last equation above does not depend on the rotation, so we need only
minimize the second summand. Equivalently, this can be done through a maximization:
max
R

n
X
i=1

Rpi q i .

(33)

The rotation matrix R has nine entries, only four of which are independent due to the orthogonality and unit determinant of R. Instead, we represent rotations using unit quaternions.
Essentially, we find the unit quaternion q that maximizes
n
X
(qpi q ) q i .

(34)

i=1

R4 .

Here we view quaternions as vectors in


Let q = (q0 , q1 , q2 , q3 )T and q = (q0 , q1 , q2 , q3 )T .
Also, the points p1 , . . . , pn and q 1 , . . . , q n are viewed as 4-tuples with pi = (0, pi1 , pi2 , pi3 )T and
, q , q )T by a slight abuse of notation.
q i = (0, qi1
i2 i3
The summand in (34) is the dot product of q i with the vector that results from applying the
rotation q to pi , which is equal to the dot product of pi with the vector that results applying the
inverse rotation q to qi . Namely,
(qpi q ) q i = pi (q q i q).

(35)

We also have
(qpi ) (q i q) = (qpi ) (qq q i q)

= (qpi ) (q(q q i q))


= pi (q q i q)

(36)

by Proposition 3. Comparion of (35) with (36) gives us


(qpi q ) q i = (qpi ) (q i q).

(37)

The above allows us to rewrite (qpi )(q i q) as a matrix product. For this purpose, we define matrices

0 qi1
qi2
qi3
0 pi1 pi2 pi3

p
0
qi3
qi2
0
pi3 pi2
i1
and Qi = qi1
,
Pi =

q
p
qi3
0
qi1
0
pi1
i2
i2 pi3

pi3 pi2 pi1


0
qi3
qi2
qi1
0

for 1 i n. Then the quaternion products qpi and q i q are equivalent to the matrix products
Pi q and Qi q, and (qpi ) (q i q) = q T PiT Qi q.
We thus have
n
n
X
X
(qpi ) (q i q)
(by (37))
(qpi q ) qi =
i=1

i=1

n
X

q T PiT Qi q

i=1

= qT

n
X
i=1

17

PiT Qi

q.

It is easy to verify that each matrix PiT Qi is symmetric, so is the 4 4 matrix


M=

n
X

PiT Qi .

i=1

Thus M has real eigenvalues only, say, 1 , 2 , 3 , 4 with 1 2 3 4 .2 Let v1 , v 2 , v 3 , v 4


be the corresponding orthogonal unit eigenvectors. Eigenvectors corresponding to different eigenvalues must be orthogonal to each other. Multiple eigenvectors corresponding to the same eigenvalue
are chosen to be orthogonal to each other. The quaternion q is a linear combination of these
eigenvectors:
q = 1 v 1 + 2 v 2 + 3 v 3 + 4 v 4 .
Therefore we have
q T M q = (1 v 1 + 2 v 2 + 3 v 3 + 4 v 4 )T M (1 v 1 + 2 v 2 + 3 v 3 + 4 v 4 )
= (1 v 1 + 2 v 2 + 3 v 3 + 4 v 4 ) (1 1 v 1 + 2 2 v2 + 3 3 v 3 + 4 4 v4 )
= 1 21 + 2 22 + 3 23 + 4 24 .

The product q T M q achieves its maximum when 1 = 1 and 2 = 3 = 4 = 0. Therefore, the


unit quaternion q that maximizes (34) is the eigenvector that corresponds to the largest eigenvalue
of the matrix M . It describes the optimal rotation for (32), i.e, for data registration.
When the corresponding points q1 , . . . , q n are unknown, a well-known method called the Iterative Closest Point (ICP) [1] solves the registration problem. Given a set of data points {p1 , . . . , pn },
(0)
(0)
the ICP algorithm finds the initial corresponding points q 1 , . . . , q n as the closest points on the
(0)
(0)
surface model to p1 = p1 , . . . , pn = pn , respectively. Then it applies the introduced quaternion(0)
(0)
based method to determine the rotation and translation that best match {pi } with {q i }. The
(0)
(1)
second iteration applies the just found transformation to every pi , obtaining pi , and then de(1)
(1)
termines its new corresponding point q i on the model as the closest point to pi . Recompute the
best rotation and translation using quaternions, and so on. The algorithm stops when the change
in the new transformation becomes small enough.

References
[1] P. J. Besl and N. D. McKay. A method for registration of 3-D shapes. IEEE Transactions on
pattern analysis and machine intelligence, 14(2):239256, 1992.
[2] W. K. Clifford. Preliminary sketch of bi-quaternions. Proceedings of the London Mathematical
Society, s14(1):381395, 1873.
[3] E. B. Dam, M. Koch, and M. Lillholm.
Quaternions, interpolation, and animation.
http://web.mit.edu/2.998/www/QuaternionReport1.pdf.
[4] O. D. Faugeras and M. Hebert. The representation, recognition, and locating of 3-D objects.
International Journal of Robotics Research, 5(3):2752, 1986.
2

Multiplicities of the eigenvalues are counted.

18

[5] W. R. Hamilton. On quaternions; or on a new system of imagniaries in algebra. London,


Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 25(3):489495, 1844.
[6] B. K. P. Horn. Closed-form solution of absolute orientation using unit quaternions. Journal of
Optical Society of America A, 4(4):629642, 1987.
[7] T. W. Hungerford. Algebra. Springer-Verlag, 1974.
[8] N. Jacobson. Basic Algebra. W. H. Freeman & Co.,1985.
[9] J. B. Kuipers. Quaternions and Rotation Sequences. Princeton University Press, 1999.
[10] S. Oldenburger. Applications of Quaternions. Written project of the course Problem Solving
Techniques in Applied Computer Science (Com S 477/577), Department of Computer Science,
Iowa State University, 2005.
[11] W. H. Press, S. A. Teukolsky, W. T.Vetterling, and B. P. Flannery. Numerical Recipies in C,
2nd edition. Cambridge University Press, Inc., 2002.
[12] J. T. Schwartz and M. Sharir. Identification of partially obscured objects in two and three
dimensions by matching noisy characteristic curves. International Journal of Robotics Research,
6(2):2944, 1987.
[13] K. Shoemake. Animating rotation with quaternion curves. Computer Graphics, 19(3):245254,
1985.
[14] F. Zhao and B. G. M. van Wachem. A novel quaternion integration approach for describing
the behaviour of non-spherical particles. Acta Mechanica, 224:30913109, 2013.

19