Indirect Kalman Filter for 3D Attitude Estimation

Nikolas Trawny and Stergios I. Roumeliotis

Department of Computer Science & Engineering
University of Minnesota

Multiple Autonomous
Robotic Systems Laboratory
Technical Report
Number 2005-002
March 2005

Dept. of Computer Science & Engineering

University of Minnesota
4-192 EE/CS Building
200 Union St. S.E.
Minneapolis, MN 55455
Tel: (612) 625-2217
Fax: (612) 625-0572
URL: http://www.cs.umn.edu/˜trawny
 Indirect Kalman Filter for 3D Attitude Estimation
Nikolas Trawny and Stergios I. Roumeliotis
Department of Computer Science & Engineering
University of Minnesota
Multiple Autonomous Robotic Systems Laboratory, TR-2005-002
March 2005

1 Elements of Quaternion Algebra

1.1 Quaternion Definitions
The quaternion is generally defined as

q̄ = q4 + q1 i + q2 j + q3 k (1)
where i, j, and k are hyperimaginary numbers satisfying

i2 = −1 , j2 = −1 , k2 = −1 ,
−ij = ji = k , − jk = kj = i , − ki = ik = j (2)

Note that this does not correspond to the Hamilton notation. It rather is a convention resulting in multiplications
of quaternions in “natural order” (see also section 1.4 and [5, p. 473]). This is in accordance with the JPL Proposed
Standard Conventions [2].
The quantity q4 is the real or scalar part of the quaternion, and q1 i + q2 j + q3 k is the imaginary or vector part. The
quaternion can therefore also be written in a four-dimensional column matrix q̄, given by
· ¸
q £ ¤T
q̄ = = q1 q2 q3 q4 (3)
q4

If the quantities q and q4 fulfill

 
kx sin(θ/2)
q = ky sin(θ/2) = k̂ sin(θ/2), q4 = cos(θ/2) (4)
kz sin(θ/2)

the elements q1 , . . . , q4 are called “quaternion of rotation” or “Euler symmetric parameters” [5]. In this notation, k̂
describes the unit vector along the axis and θ the angle of rotation. The quaternion of rotation is a unit quaternion,
satisfying q
p
|q̄| = q̄ T q̄ = |q|2 + q42 = 1 (5)
Henceforth, we will use the term “quaternion” to refer to a quaternion of rotation.
The quaternion q̄ and the quaternion −q̄ describe a rotation to the same final coordinate system position, i. e. the
angle–axis representation is not unique [5, p. 463]. The only difference is the direction of rotation to get to the target
configuration, with the quaternion with positive scalar element q4 describing the shortest rotation [2].

2
1.2 Quaternion Multiplication
The quaternion multiplication is defined as
q̄ ⊗ p̄ = (q4 + q1 i + q2 j + q3 k) (p4 + p1 i + p2 j + p3 k)
= q4 p4 − q1 p1 − q2 p2 − q3 p3 + (q4 p1 + q1 p4 − q2 p3 + q3 p2 ) i
+ (q4 p2 + q2 p4 − q3 p1 + q1 p3 ) j + (q4 p3 + q3 p4 − q1 p2 + q2 p1 ) k
 
q4 p1 + q3 p2 − q2 p3 + q1 p4
 −q3 p1 + q4 p2 + q1 p3 + q2 p4 
=  
 q2 p1 − q1 p2 + q4 p3 + q3 p4 
−q1 p1 − q2 p2 − q3 p3 + q4 p4
where we have used the relations defined in Eq. (2).
The quaternion multiplication can alternatively be written in matrix form. For this, we first introduce the matrix-
notation for the cross-product using the skew-symmetric matrix operator bq ×c, defined as
 
0 −q3 q2
bq ×c =  q3 0 −q1  (6)
−q2 q1 0
The cross product can then be written as
¯ ¯     
¯i j k ¯¯ q2 p3 − q3 p2 0 −q3 q2 p1
¯
q × p = ¯¯q1 q2 q3 ¯¯ = q3 p1 − q1 p3  =  q3 0 −q1  p2  = bq ×cp (7)
¯p1 p2 p3 ¯ q1 p2 − q2 p1 −q2 q1 0 p3
The quaternion multiplication can now be rewritten in matrix form as
q̄ ⊗ p̄ = L(q̄)p̄
  
q4 q3 −q2 q1 p1
 −q3 q4 q q  
2   p2

q̄ ⊗ p̄ =  1  (8)
 q2 −q1 q4 q3   p3 
−q1 −q2 −q3 q4 p4
· ¸· ¸
q4 I3×3 − bq ×c q p
q̄ ⊗ p̄ = (9)
−qT q4 p4

· ¸
q4 p + p 4 q − q × p
=
q4 p4 − qT p
· ¸
p4 q + pq4 + bp ×cq
=
p4 q4 − pT q

· ¸· ¸
p4 I3×3 + bp ×c p q
q̄ ⊗ p̄ =
−pT p4 q4
 
p4 −p3 p2 p1 q1
 p3 p4 −p1 p2   
q̄ ⊗ p̄ =    q2  (10)
 −p2 p1 p4 p 3   q3 
−p1 −p2 −p3 p4 q4
q̄ ⊗ p̄ = R(p̄)q̄

Properties of L and R
L(q̄ −1 ) = LT (q̄) (11)
−1 T
R(p̄ ) = R (p̄) (12)
(13)

TR-2005-002 3
 £ ¤
L = Ψ(q̄) q̄ (14)
£ ¤
R = Ξ(p̄) p̄ (15)
(16)

with matrices Ψ and Ξ defined as

· ¸
q4 I3×3 − bq ×c
Ψ= (17)
−qT
· ¸
p4 I3×3 + bp ×c
Ξ= (18)
−pT

If two rotations, CA and CB are related, such that CA CX = CX CB (hand-eye calibration), then the related
matrix L(q̄A ) − R(q̄B ) is skew-symmetric, and of rank 2. In particular, q̄A4 = q̄B4 and ||qA || = ||qB ||.
Quaternions also have an neutral element with respect to multiplication, which is defined as
£ ¤T
q̄0 = 0 0 0 1 (19)

q̄ ⊗ q̄0 = q̄0 ⊗ q̄ = q̄ (20)

The inverse rotation is described by the inverse or complex conjugate quaternion, denoted as
· ¸ · ¸ · ¸
−q −k̂ sin(θ/2) k̂ sin(−θ/2)
q̄ −1 = = = (21)
q4 cos(θ/2) cos(−θ/2)

q̄ ⊗ q̄ −1 = q̄ −1 ⊗ q̄ = q̄0 (22)

(q̄ ⊗ p̄)−1 = p̄−1 ⊗ q̄ −1 (23)

1.3 Useful Identities

1.3.1 Properties of the cross product skew-symmetric matrix bω ×c
Anti-Commutativity
bω ×c = −bω ×cT (24)

ba ×cb = −bb ×ca (25)

T T
⇔ a bb ×c = −b ba ×c (26)

Distributivity over Addition

ba ×c + bb ×c = ba + b ×c (27)

Scalar Multiplication
c · bω ×c = bc ω ×c (28)

Cross Product of Parallel Vectors

¡ ¢T
ω × (c · ω) = c · bω ×cω = −c · ω T bω ×c = 03×1 (29)

TR-2005-002 4
Lagrange’s Formula
¡ ¢
ba ×cbb ×c = baT − aT b I3×3 (30)
T T
⇔ a × (b × c) = b(a c) − c(a b) (31)

ba ×cbb ×c + abT = bb ×cba ×c + baT (32)

b(a × b) ×c = baT − abT (= (a × b) × c) (33)

Jacobi Identity

ba ×cbb ×cc + bb ×cbc ×ca + bc ×cba ×cb = 0 (34)

Rotations

bCa ×c = Cba ×cCT (35)

C(a × b) = (Ca) × (Cb) (36)

Cross product of vectors in quaternion notation If we define the quaternions

· ¸ · ¸ · ¸ · ¸
a b c a×b
ā = , b̄ = , c̄ = = (37)
0 0 0 0

we can show that

1¡ ¢
c̄ = b̄ ⊗ ā + ā ⊗ b̄−1 (38)
2
1¡ ¢
= (L(b̄) + RT (b))ā (39)
2 µ· ¸ · ¸¶
1 −bb ×c − bb ×c b−b a
= (40)
2 −bT + bT 0 0

TR-2005-002 5
Powers of bω ×c

bω ×c2 = ωω T − |ω|2 · I3×3 (41)

¡ ¢
bω ×c3 = ωω T − |ω|2 · I3×3 bω ×c
= ωω T bω ×c − |ω|2 · bω ×c
T
= ω (−bω ×cω) − |ω|2 · bω ×c
T
= −ω (ω × ω) − |ω|2 · bω ×c
= −|ω|2 · bω ×c (42)

bω ×c4 = bω ×c3 · bω ×c
= −|ω|2 · bω ×c2 (43)

bω ×c5 = bω ×c3 · bω ×c2

¡ ¢
= −|ω|2 · bω ×c ωω T − |ω|2 · I3×3
= +|ω|4 · bω ×c (44)

bω ×c6 = bω ×c5 · bω ×c
= +|ω|4 · bω ×c2 (45)

bω ×c7 = bω ×c5 · bω ×c2

= −|ω|6 · bω ×c (46)

and so on.

TR-2005-002 6
1.3.2 Properties of the matrix Ω
The matrix Ω appears in the product of a vector and a quaternion, and is used for example in the quaternion derivative.
It has the following properties:
 
0 ωz −ωy ωx
−ωz 0 ωx ωy 
Ω(ω) =   ωy −ωx
 (47)
0 ωz 
−ωx −ωy −ωz 0
· ¸
−bω ×c ω
= (48)
−ω T 0

· ¸
2 bω ×c2 − ωω T −bω ×cω
Ω(ω) =
ω T bω ×c −ω T ω
· ¸
−|ω|2 · I3×3 03×1
=
01×3 −|ω|2
= −|ω|2 · I4×4 (49)

Ω(ω)3 = −|ω|2 · Ω(ω) (50)

Ω(ω)4 = |ω|4 · I4×4 (51)

Ω(ω)5 = |ω|4 · Ω(ω) (52)

Ω(ω)6 = −|ω|6 · I4×4 (53)

and so on.

1.3.3 Properties of the matrix Ξ

The matrix Ξ(q̄) appears in the multiplication of a vector with a quaternion. The relationship between Ξ(q̄) and Ω(a)
is equivalent to that between the multiplication matrices L(q̄) and R(p̄) (cf. section 1.2). It is defined as
 
q4 −q3 q2  
· ¸ q4 q3 −q2 −q1
 q3 q4 −q1   q ·I + bq ×c
Ξ(q̄) =  −q2 q1 = 4 3x3 T , ΞT (q̄) = −q3 q4 q1 −q2  (54)
q4  −q
q2 −q1 q4 −q3
−q1 −q2 −q3
and it can be shown that
ΞT (q̄)Ξ(q̄) = I3×3 (55)
T T
Ξ(q̄)Ξ (q̄) = I4×4 − q̄ q̄ (56)
ΞT (q̄)q̄ = 03×1 (57)
The relationship between Ξ and Ω is given by [3, eq. (60)]
Ω(a)q̄ = Ξ(q̄)a (58)

1.4 Relationship between Quaternion and Rotational Matrix

Given a vector p we define the corresponding quaternion as
· ¸
p
p̄ = (59)
0

TR-2005-002 7
We will be using the following two relations between vectors expressed in different coordinate frames
L
p=L G
G C(q̄) p (60)

where q̄ = L L
G q̄ and G C(q̄) is the (3 × 3) rotational matrix that expresses the (global) frame {G} with respect to the
(local) frame {L}.
A vector can also be transformed from one coordinate frame to another by pre- and postmultiplying its quaternion
by the rotation quaternion and its inverse, respectively.

L L
p̄ = ⊗ G p̄ ⊗ L
G q̄ G q̄
−1
(61)
· ¸· ¸
q4 I3×3 − bq ×c q p
= ⊗GL q̄
−1
−qT q4 0
· ¸ · ¸
q4 p − q × p −q
= ⊗
−qT p q4
· 2 T
¸
q4 p − q4 q × p + qq p + q4 p × q − (q × p) × q
=
+q4 qT p − q4 qT p − qT (q × p)
· 2 ¡ ¢ ¸
q4 p − 2q4 q × p + qqT p − (1 − q42 I3×3 ) − qqT p
=
−qT bq ×cp
· ¡ 2 ¢ ¸· G ¸
2q4 − 1 I3×3 − 2q4 bq ×c + 2qqT p
=
0 0

This gives us the relationship between a quaternion and its corresponding rotational matrix.
L
¡ 2 ¢ T
G C(q̄) = 2q4 − 1 I3×3 − 2q4 bq ×c + 2qq (62)

which can also be written as

L
G C(q̄) = ΞT (q̄)Ψ(q̄) (63)

using the definitions of Ξ and Ψ from Section 1.2.

A similar form can be derived for the triple product of quaternions, although without an obvious physical interpre-
tation.

q̄ ⊗ p̄ ⊗ q̄ −1 (64)
T
=L(q̄)R (q̄)p̄ (65)
· ¸· ¸
C(q̄) 0 p
= (66)
0 1 p4
· ¸
C(q̄)p
= (67)
p4

In case of only a very small rotation δ q̄, we can use the small angle approximation to simplify the above expression.
We can write the quaternion describing a small rotation as
· ¸
δq
δ q̄ = (68)
δq4
· ¸
k̂ sin(δθ/2)
= (69)
cos(δθ/2)
·1 ¸
δθ
≈ 2 (70)
1

leading to the following expression for the corresponding rotational matrix

L
G C(δ q̄) ≈ I3×3 − bδθ ×c (71)

TR-2005-002 8
 The result from eq. (62) could have also been found by rewriting Euler’s formula which relates the rotational
matrix and the angle–axis representation [2]
L
GC = cos(θ) · I3×3 − sin(θ)bk̂ ×c + (1 − cos(θ)) k̂k̂T (72)
¡ ¢
= 2 cos2 (θ/2) − 1 · I3×3 − 2 cos(θ/2) sin(θ/2)bk̂ ×c + 2 sin2 (θ/2)k̂k̂T (73)

Substituting the appropriate quaternion components according to eq. (4) readily yields equation (62). The latter
can also be expressed in terms of quaternion components as
 2 
q1 − q22 − q32 + q42 2 (q1 q2 + q3 q4 ) 2 (q1 q3 − q2 q4 )
L
G C(q̄) =
 2 (q1 q2 − q3 q4 ) −q12 + q22 − q32 + q42 2 (q2 q3 + q1 q4 )  (74)
2 (q1 q3 + q2 q4 ) 2 (q2 q3 − q1 q4 ) −q12 − q22 + q32 + q42
 2 
2q1 + 2q42 − 1 2 (q1 q2 + q3 q4 ) 2 (q1 q3 − q2 q4 )
= 2 (q1 q2 − q3 q4 ) 2q22 + 2q42 − 1 2 (q2 q3 + q1 q4 ) (75)
2 (q1 q3 + q2 q4 ) 2 (q2 q3 − q1 q4 ) 2q32 + 2q42 − 1
 
1 − 2q22 − 2q32 2 (q1 q2 + q3 q4 ) 2 (q1 q3 − q2 q4 )
= 2 (q1 q2 − q3 q4 ) 1 − 2q12 − 2q32 2 (q2 q3 + q1 q4 ) (76)
2 (q1 q3 + q2 q4 ) 2 (q2 q3 − q1 q4 ) 1 − 2q12 − 2q22

Another alternative form of eq. (62) arises after exploiting the relationship between qqT and bq ×c2 (cf. eq. (41))
L
G C(q̄) = I3×3 − 2q4 bq ×c + 2bq ×c2 (77)

Note that, due to the convention chosen for the quaternion multiplication (cf. eq. (2)), the product of two rotational
matrices will correspond to the product of two quaternions in the same order [2, 5]. Thus,
µ ¶
L1 L1 L2 L2 L1 L1 L2
L2 C(L2 q̄) · G C(G q̄) = G C L2 q̄ ⊗ G q̄ (78)

· ¸
L1 k̂ sin(θ/2)
Finally, we can interpret L2 q̄
= as the quaternion describing the rotation of coordinate frame {L2 }
cos(θ/2)
to coordinate frame {L1 }, with the axis of rotation k̂ expressed in {L1 } (cf. figure 1). This can also be expressed as
the matrix exponential ³ ´
L
G C(q̄) = exp −bk̂ ×cθ (79)

1.5 Quaternion Time Derivative

When the local coordinate frame {L} is moving with respect to the global reference frame {G}, we can compute the
rate of change or the derivative of the corresponding quaternion describing their relationship. We do that by computing
the limit of the difference quotient

L(t) 1 ³L(t+∆t) L(t)

´
˙
G q̄(t) = lim G q̄ − G q̄ (80)
∆t→0 ∆t

L(t+∆t)
The quaternion G q̄ can be expressed as the product of two quaternions
L(t+∆t) L(t+∆t) L(t)
G q̄ = L(t) q̄ ⊗ G q̄ (81)

where · ¸
L(t+∆t) k̂ sin(θ/2)
L(t) q̄ = (82)
cos(θ/2)
L(t+∆t)
Note, that the quaternion L(t) q̄ describes the rotation of reference frame {L(t)} to reference frame {L(t+∆t)}
in terms of the rotation angle θ and the axis of rotation k̂ (the latter being expressed in {L(t + ∆t)}).

TR-2005-002 9
 Figure 1: Relationship between global and local frame

In the limit, as ∆t → 0, the angle of rotation will become very small, so that we can approximate the sin and
cos-functions by their first-order Taylor expansion
· ¸ · ¸ ·1 ¸
L(t+∆t) k̂ sin(θ/2) k̂ · θ/2 2 · δθ
q̄ = ≈ = (83)
L(t) cos(θ/2) 1 1
The vector δθ has the direction of the axis of rotation and the magnitude of the angle of the rotation. Dividing this
vector by ∆t will yield, in the limit, the rotational velocity
δθ
ω = lim (84)
∆t→0 ∆t

We are now ready to derive the quaternion derivative as

L(t) 1 ³L(t+∆t) L(t)
´
G
˙
q̄(t) = lim G q̄ − G q̄
∆t→0 ∆t
1 ³L(t+∆t) L(t) L(t)
´
= lim q̄ ⊗ G q̄ − q̄0 ⊗ G q̄
∆t→0 ∆t L(t)
µ· 1 ¸ · ¸¶
1 2 · δθ − 0 L(t)
≈ lim ⊗ G q̄
∆t→0 ∆t 1 1
· ¸
1 ω L(t)
= ⊗ G q̄ (85)
2 0
· ¸
1 −bω ×c ω L(t)
= q̄
2 −ω T 0 G
1 L(t)
= Ω(ω) G q̄ (86)
2
1 L(t)
= Ξ( q̄)ω (87)
2 G
with  
0 ωz −ωy ωx
−ωz 0 ωx ωy 
Ω(ω) = 
 ωy
 (88)
−ωx 0 ωz 
−ωx −ωy −ωz 0

TR-2005-002 10
and (cf. sections 1.3.2 and 1.3.3)  
q4 −q3 q2
L(t)  q3 q4 −q1 
Ξ(G q̄) = 
−q2
 (89)
q1 q4 
−q1 −q2 −q3
Note that ω = L(t) ω, i. e. the turn rate is expressed in the local, not the inertial coordinate frame [5, p. 482].

1.6 Quaternion Integration

Integrating a quaternion is equivalent to solving the first order differential equation (cf. eq. (86))

L ˙ 1
G q̄(t) = Ω(ω) L
G q̄(t) (90)
2
where we have dropped the time index from the leading superscript and instead denoted the time variability of the
quaternion by writing q̄ = q̄(t) for greater notational clarity. It should be clear that the leading superscript L refers to
the local frame {L} at time instant t.
The solution to this differential equation has the general form [4, p. 40]
L
G q̄(t) = Θ(t, tk ) L
G q̄(tk ) (91)
Differentiating and reordering the terms yields the governing equation for Θ(t, tk ) as
L ˙ L −1
Θ̇(t, tk ) = G q̄(t) G q̄ (tk ) (92)
1
= Ω(ω(t)) L L −1
G q̄(t) G q̄ (tk ) (93)
2
1
= Ω(ω(t))Θ(t, tk ) (94)
2
with the initial condition
Θ(tk , tk ) = I4×4 (95)
Under certain assumptions we can obtain closed form solutions for this equation. The simplest assumption is that
ω is constant over the integration period ∆t = tk+1 − tk , thus making the differential equation linear time invariant.
This assumption leads to the zeroth order quaternion integrator. The next more accurate approximation is to assume a
linear evolution of ω during ∆t. We will term the resulting formula the first order quaternion integrator.

1.6.1 Zeroth Order Quaternion Integrator

If ω(t) = ω is constant over the integration period ∆t = tk+1 − tk , the matrix Ω does not depend on time, and
Θ(tk+1 , tk ) can be expressed as [4, eq. (2-58a)]
µ ¶
1
Θ(tk+1 , tk ) = Θ(∆t) = exp Ω(ω)∆t (96)
2
We can rewrite this matrix exponential using its Taylor series expansion
µ ¶2 µ ¶3
1 1 1 1 1
Θ(∆t) = I4×4 + Ω(ω)∆t + Ω(ω)∆t + Ω(ω)∆t + . . . (97)
2 2! 2 3! 2
Using the properties of the matrix Ω described in section 1.3.2, this transforms into
µ ¶2
1 1 1
Θ(∆t) = I4×4 + ∆t Ω(ω) − ∆t |ω|2 · I4×4
2 2! 2
µ ¶3 µ ¶4
1 1 2 1 1
− ∆t |ω| · Ω(ω) + ∆t |ω|4 · I4×4
3! 2 4! 2
µ ¶5 µ ¶6
1 1 1 1
+ ∆t |ω|4 · Ω(ω) − ∆t |ω|6 · I4×4 − . . . (98)
5! 2 6! 2

TR-2005-002 11
 Reordering and expanding with |ω| yields
à µ ¶2 µ ¶4 !
1 1 1 1
Θ(∆t) = 1− ∆t |ω|2 + ∆t |ω|4 − . . . I4×4
2! 2 4! 2
à µ ¶3 µ ¶5 !
1 1 1 1 1 1
+ |ω|∆t − |ω|∆t + |ω|∆t − . . . Ω(ω) (99)
|ω| 2 3! 2 5! 2

After close inspection, we recognize the Taylor series expansions of the sin and cos functions
µ ¶ µ ¶
|ω| 1 |ω|
Θ(∆t) = cos ∆t · I4×4 + sin ∆t · Ω(ω) (100)
2 |ω| 2

We can finally write the zero-th order quaternion integrator as

L
G q̄(tk+1 ) = Θ(tk+1 , tk ) L
G q̄(tk )
µ µ ¶ µ ¶ ¶
|ω| 1 |ω| L
= cos ∆t · I4×4 + sin ∆t · Ω(ω) G q̄(tk ) (101)
2 |ω| 2

A close look reveals that Θ(tk+1 , tk ) is nothing but the multiplication matrix associated with a specific quaternion,
so that we can rewrite the zero-th order quaternion integration as a quaternion product
 ³ ´
ω |ω|
· sin ∆t
L
G q̄(tk+1 ) =
 |ω| ³ 2 ´  ⊗ L G q̄(tk ) (102)
cos |ω|2 ∆t

This quaternion product corresponds to rotating the original frame around the rotation axis defined by ω through the
angle |ω|∆t, which corresponds exactly to the assumption of a constant ω.
The above expression will cause numerical instability for very small ω, due to |ω| appearing in the denominator.
We will therefore compute the limit of the above equation as |ω| goes towards zero, making multiple use of L’Hôpital’s
rule.
µ µ ¶ µ ¶ ¶
|ω| 1 |ω|
lim Θ(∆t) = lim cos ∆t I4×4 + sin ∆t Ω(ω)
|ω|→0 |ω|→0 2 |ω| 2
µ µ ¶ ¶
1 |ω|
= I4×4 + lim sin ∆t Ω(ω)
|ω|→0 |ω| 2
∆t
= I4×4 + Ω(ω) (103)
2

1.6.2 First Order Quaternion Integrator

The first order quaternion integrator makes the assumption of a linear evolution of ω during the integration interval
∆t. In this case, we have to modify the matrix Θ(tk+1 , tk ) from eq. (96). For that purpose, we introduce the average
turn rate ω̄, defined as
ω(tk+1 ) + ω(tk )
ω̄ = (104)
2
We can also define the derivative of the turn rate ω̇ and the associated matrix Ω(ω̇), which, in the linear case, is
constant µ ¶
ω(tk+1 ) − ω(tk )
Ω(ω̇) = Ω (105)
∆t
Note that higher order derivatives of Ω(ω) are zero.
Following Wertz [8, p. 565], in order to compute the quaternion at time instant tk+1 , we write its Taylor series
expansion around time instant tk

L 1
G q̄(tk+1 ) =L L ˙
G q̄(tk ) +G q̄(tk )∆t +
L¨
G q̄ (tk )∆t
2
+ ... (106)
2

TR-2005-002 12
 Repeatedly applying the definition of the quaternion time derivative (eq. (86)) yields
à µ ¶2 µ ¶3 !
L 1 1 1 1 1
G q̄(tk+1 ) = I4×4 + Ω(ω(tk ))∆t + Ω(ω(tk ))∆t + Ω(ω(tk ))∆t + . . . L G q̄(tk )
2 2! 2 3! 2
µ ¶
1 1 1
+ ∆t2 Ω(ω̇(tk ))LG q̄(t k ) + Ω(ω̇(tk ))Ω(ω(tk )) + Ω(ω(t k ))Ω( ω̇(t k )) ∆t3L
G q̄(tk ) + . . . (107)
4 12 24

If we write the average Ω(ω̄) as

tZk+1
1 1
Ω(ω̄) = Ω(ω(τ )) dτ = Ω(ω(tk )) + Ω(ω̇(tk ))∆t (108)
∆t 2
tk

we can reorder the terms in eq. (107) to form

à µ ¶2 µ ¶3
L 1 1 1 1 1
G q̄(t k+1 ) = I 4×4 + Ω( ω̄)∆t + Ω( ω̄)∆t + Ω( ω̄)∆t + ...
2 2! 2 3! 2
!
1³ ´
3 L
+ Ω(ω̇(tk ))Ω(ω(tk )) − Ω(ω(tk ))Ω(ω̇(tk )) ∆t G q̄(tk ) (109)
48

Recognizing the first term as the Taylor series expansion of the matrix exponential, and after replacing Ω(ω̇) with
its definition (eq. (105)), we obtain the final formula
µ µ ¶ ¶
L 1 1³ ¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢´ 2 L
G q̄(tk+1 ) = exp Ω( ω̄)∆t + Ω ω(t k+1 ) Ω ω(t k ) − Ω ω(t k ) Ω ω(tk+1 ) ∆t G q̄(tk ) (110)
2 48

TR-2005-002 13
2 Attitude Propagation
The Kalman Filter consists of two stages to determine an estimate of the current attitude. In the first stage, the
propagation, the filter produces a prediction of the attitude based on the last estimate and some proprioceptive mea-
surements. This estimate is then corrected in the update stage, where new absolute orientation measurements are taken
into account.
One way to predict the system state is to feed the control commands given to the system into a system model and
thus to predict the behavior of the system. An alternative way to estimate position and orientation is to use data from an
inertial measurement unit (IMU) as dynamic model replacement. The IMU provides measurements of the translational
accelerations and rotational velocities acting on the system. In this paper, we will pursue the latter approach.
Before discussing the state equation and the error propagation, we will describe the model for the gyros that will
provide measurements of the rotational velocity.

2.1 Gyro Noise Model

As part of the IMU, a three-axis gyro provides measurements of the rotational velocity. Gyros are known to be subject
to different error terms, such as a rate noise error and a bias. In accordance with the literature [3, 1], we use a simple
model that relates the measured turn rate ω m to the real angular velocity ω as
ω m = ω + b + nr (111)
In this equation, b denotes the gyro bias and nr the rate noise, assumed to be Gaussian white noise with characteristics
E [nr ] = 03×1 (112)
£ ¤
E nr (t + τ )nr T (t) = Nr δ(τ ) (113)
The gyro bias is non-static and simulated as a random walk process
ḃ = nw (114)
with characteristics
E [nw ] = 03×1 (115)
£ ¤
E nw (t + τ )nw T (t) = Nw δ(τ ) (116)
The bias is therefore a random quantity and needs to be estimated along with the quaternion.
For simplification we will assume that the noise is equal in all three spatial directions, i. e.
Nr = σr2c · I3×3 (117)
2
Nw = σw c
· I3×3 (118)
The subscript c indicates, that these are the noise covariances for the continuous-time system, distinguishing them
from the noise covariances in the discretized system used later.
In order to determine the units of the covariances [1] we consider the scalar case for illustration purposes, which
extend directly to the vector case. For compatibility, nr has the same units as ω, therefore
h i £ ¤ rad2
E [nr (t + τ )nr (t)] = σr2c δ(τ ) = (119)
sec2
But the unit of δ(τ ) is defined as [4, p. 221]
1
[δ(τ )] = (120)
sec
Therefore
£ 2¤ rad2
σrc = · sec (121)
sec2
µ ¶2
rad 1
= · 1 (122)
sec sec
µ ¶2
rad 1
= · (123)
sec Hz

TR-2005-002 14
and µ ¶
rad 1 rad
[σrc ] = ·√ =√ (124)
sec Hz sec
Analogously, we obtain
µ ¶2
£ 2 ¤ rad
σwc = /sec · sec (125)
sec
rad2
= (126)
sec3
µ ¶2
rad
= · Hz (127)
sec

and µ ¶
rad rad√
[σwc ] = Hz = √ · (128)
sec sec3
In the discrete case, nrd and nwd will have to have the same units as the turn rate (cf. also the state equation in
section 2.2), namely rad
sec . In order to preserve equivalence of the noise strength, we have to incorporate the sampling
frequency when discretizing, yielding
σ
σrd = √ rc (129)
∆t
√
σwd = σwc · ∆t (130)
1
where ∆t is the inverse of the sampling frequency, or fsample = ∆t . The discrete variances will be used in the
simulation of noisy real-world data.
Further explanation for conversion between continuous and discrete noise variance is provided by Simon [6,
pp. 230-233]. To make the analogy to the derivation in the book, the bias driving noise, nw , corresponds to the
process noise, whereas the rate noise, nr , corresponds to the measurement noise.

2.2 The State Equation

In direct consequence of the previous analysis, we define a seven-element state vector consisting of the quaternion and
the gyro-bias · ¸
q̄(t)
x(t) = (131)
b(t)
Using the definition of the quaternion derivative (eq. (86)) and the error model (eqs. (111) and (114)), we find the
following system of differential equations governing the state

L ˙ 1
G q̄(t) = Ω(ω m − b − nr ) L
G q̄(t) (132)
2
ḃ = nw (133)

Taking the expectation of the above yields the prediction equations (cf. [3, p. 422]) for the state within the EKF-
framework

Lˆ ˙ 1
G q̄ (t) = Ω(ω̂) L ˆ
G q̄ (t) (134)
2
˙
b̂ = 03×1 (135)

with
ω̂ = ω m − b̂ (136)
Since the bias is constant over the integration interval, we may integrate the quaternion using the zeroth order (cf.
section 1.6.1) or first order (cf. section 1.6.2) integrator, using ω̂ instead of ω.

TR-2005-002 15
2.3 Error and Covariance Representation
Usually, the error vector and its covariance is expressed in terms of the arithmetic difference between the state vector
and its estimate. In the problem at hand, however, this representation is problematic, due to the presence of constraints
in the system. The fact that the quaternion is enforced to be of unit length (cf. eq. (5)) makes the corresponding
covariance matrix singular [3, p. 423], which is difficult to maintain numerically. For stability reasons, we will
therefore use a different, six-dimensional representation of the error vector.
Instead of using the arithmetic difference between quaternion and quaternion estimate to define the error, we will
employ the error quaternion δ q̂; a small rotation between the estimated and the true orientation of the local frame of
reference. Instead of a difference, this error is defined as a multiplication.
L
G q̄ =L
L̂
δ q̄ ⊗ L̂ ˆ
G q̄ (137)
L L L̂ ˆ−1
L̂
δ q̄ = G q̄ ⊗ G q̄ (138)

Since the rotation associated with the error quaternion δ q̄ can be assumed to be very small, we can employ the
small angle approximation (as seen in section 1.4) and define the attitude error angle vector δθ as follows
· ¸
δq
δ q̄ = (139)
δq4
· ¸
k̂ sin(δθ/2)
= (140)
cos(δθ/2)
·1 ¸
δθ
≈ 2 (141)
1

This error angle vector δθ is of dimension 3 × 1 and will be used together with the bias error in the error state vector.
The bias error is defined as
∆b = b − b̂ (142)
We can now define the error vector as · ¸
δθ
x̃ = (143)
∆b
In the next section we will develop the continuous time first order state equation for the error vector.

2.4 Continuous Time Error State Equations

In order to derive the continuous time linear state equations for the error vector, we will start with the definition of the
error quaternion (eq. (137))
d
q̄ = δ q̄ ⊗ q̄ˆ | (144)
dt
q̄˙ = δ˙q̄ ⊗ q̄ˆ + δ q̄ ⊗ q̄ˆ˙ (145)

Substituting the definitions for q̄˙ (eq. (85)) and q̄ˆ˙ (eq. (134)) leads to
· ¸ · ¸ · ¸
1 ω ˙ 1 ω̂ 1 ω̂
⊗ q̄ = δ q̄ ⊗ q̄ˆ + δ q̄ ⊗ ( ⊗ q̄ˆ) | − (δ q̄ ⊗ ⊗ q̄ˆ) (146)
2 0 2 0 2 0
µ· ¸ · ¸ ¶
˙ 1 ω ω̂
δ q̄ ⊗ q̄ˆ = ⊗ q̄ − δ q̄ ⊗ ⊗ q̄ˆ | ⊗ q̄ˆ−1 , eq. (138) (147)
2 0 0
µ· ¸ · ¸¶
1 ω ω̂
δ˙q̄ = ⊗ δ q̄ − δ q̄ ⊗ (148)
2 0 0

Combining the gyro model for ω (cf. eq. (111)) and the definition of ω̂ (cf. eq. (136)) yields

ω = ω̂ − ∆b − nr (149)

TR-2005-002 16
which, upon substitution into the above leads to
µ· ¸ · ¸¶ · ¸
˙ 1 ω̂ ω̂ 1 ∆b + nr
δ q̄ = ⊗ δ q̄ − δ q̄ ⊗ − ⊗ δ q̄ (150)
2 0 0 2 0
µ· ¸ · ¸ ¶ · ¸
1 −bω̂ ×c ω̂ +bω̂ ×c ω̂ 1 ∆b + nr
= T · δ q̄ − T · δ q̄ − ⊗ δ q̄ (151)
2 −ω̂ 0 −ω̂ 0 2 0
· ¸ · ¸
1 −2bω̂ ×c 03×1 1 ∆b + nr
= · δ q̄ − ⊗ δ q̄ (152)
2 0T 3×1 0 2 0
· ¸ · ¸ · ¸
1 −2bω̂ ×c 03×1 1 −b(∆b + nr ) ×c (∆b + nr ) δq
= · δ q̄ − · (153)
2 0T 3×1 0 2 −(∆b + nr )T 0 1
· ¸ · ¸
1 −2bω̂ ×c 03×1 1 (∆b + nr )
= · δ q̄ − − O(|∆b||δq|, |nr ||δq|) (154)
2 0T 3×1 0 2 0

Neglecting the second order terms, we can write

· ¸ " ˙ # · ¸
˙
δq 1
δθ −ω̂ × δq − 12 (∆b + nr )
˙
δ q̄ = ˙ = 2 = (155)
δq4 1̇ 0

or finally
˙ = −ω̂ × δθ − ∆b − nr
δθ (156)

The governing equation for the bias error is easily computed from eqs. (133) and (135) as

˙ = ḃ − b̂˙ = nw
∆b (157)

Combining these results, we may write the error state equations as

· ¸ · ¸ · ¸ · ¸ · ¸
˙
δθ −bω̂ ×c −I3×3 δθ −I3×3 03×3 nr
˙ = · + · (158)
∆b 0 3×3 0 3×3 ∆b 03×3 I3×3 nw

or
x̃˙ = Fc · x̃ + Gc · n (159)
where
· ¸ · ¸
−bω̂ ×c −I3×3 −I3×3 03×3
Fc = , Gc = (160)
03×3 03×3 03×3 I3×3

are the system matrix and the noise matrix, and

· ¸ · ¸
δθ nr
x̃ = , n= (161)
∆b nw

denote the error state and the noise vector, respectively.

As discussed in section 2.1, we assume nr and nw to be white and independent, so that the continuous time system
noise covariance matrix is specified by
· ¸ · 2 ¸
£ T
¤ Nr 03×3 σrc · I3×3 03×3
Qc = E n(t + τ )n (t) = = 2 (162)
03×3 Nw 03×3 σw c
· I3×3

2.5 Discrete Time Error State Equations

For implementation of the discrete time Kalman Filter equations, we will need to discretize the above error propagation
model. In particular, we will have to find the state transition matrix Φ and the system noise covariance Qd [4, Chapter
4.9].

TR-2005-002 17
2.5.1 The State Transition Matrix
Since the continuous time system matrix Fc is constant over the integration time step, we may write the state transition
matrix as [4, eq. (2-58a)]
Φ(t + ∆t, t) = exp (Fc ∆t) (163)
1 2 2
= I6×6 + Fc ∆t + F ∆t + . . . (164)
2! c
Straightforward calculation produces the powers of Fc as
· ¸ · ¸
−bω̂ ×c −I3×3 bω̂ ×c2 bω̂ ×c
Fc = , F2c =
03×3 03×3 03×3 03×3
· 3 2
¸ · ¸ (165)
−bω̂ ×c −bω̂ ×c bω̂ ×c4 bω̂ ×c3
F3c = , F4c =
03×3 03×3 03×3 03×3
In keeping with the notation of Lefferts et al. [3], we see that the transition matrix has the following block structure
· ¸
Θ Ψ
Φ(t + ∆t, t) = (166)
03×3 I3×3
The matrix Θ can be written as
1 1
Θ = I3×3 − bω̂ ×c∆t + bω̂ ×c2 ∆t2 − bω̂ ×c3 ∆t3 + . . . (167)
2! 3!
Using the properties of the skew-symmetric matrix bω̂ ×c (cf. section 1.3.1) and reordering the terms yields
µ ¶ µ ¶
1 2 3 1 2 1
Θ = I3×3 + −∆t + |ω̂| ∆t − . . . bω̂ ×c + ∆t − |ω̂| ∆t + . . . bω̂ ×c2
2 4
(168)
3! 2! 4!
µ ¶ µ µ ¶¶
1 1 1 1 1
= I3×3 − |ω̂|∆t − |ω̂|3 ∆t3 + . . . bω̂ ×c + 2
1 − 1 − |ω̂|2 ∆t2 + |ω̂|4 ∆t4 − . . . bω̂ ×c2
|ω̂| 3! |ω̂| 2! 4!
(169)
1 1
= I3×3 − sin (|ω̂|∆t) bω̂ ×c + (1 − cos(|ω̂|∆t)) bω̂ ×c2 (170)
|ω̂| |ω̂|2
Further expansion of bω̂ ×c gives
ω̂ ω̂ ω̂ T
Θ = cos (|ω̂|∆t) · I3×3 − sin (|ω̂|∆t) · b ×c + (1 − cos(|ω̂|∆t)) · (171)
|ω̂| |ω̂| |ω̂|
where comparison with section 1.4 reveals that Θ is in fact a rotational matrix with ω as the axis of rotation and |ω̂|∆t
the corresponding angle.
For small values of |ω̂|, either of the above expressions will lead to numerical instability. By taking the limit and
applying L’Hôpital’s rule, we arrive at
∆t2
lim Θ = I3×3 − ∆tbω̂ ×c + bω̂ ×c2 (172)
|ω|→0 2
Proceeding in similar fashion with the matrix Ψ, we find
1 2 1
Ψ = −I3×3 ∆t + ∆t bω̂ ×c − ∆t3 bω̂ ×c2 + . . . (173)
2!
µ 3! ¶ µ ¶
1 2 1 1 3 1
= −I3×3 ∆t + ∆t − |ω̂| ∆t + . . . bω̂ ×c + − ∆t + |ω̂| ∆t − . . . bω̂ ×c2
2 4 2 5
(174)
2! 4! 3! 5!
µ µ ¶¶
1 1 1
= −I3×3 ∆t + 1 − 1 − |ω̂|2 ∆t2 − |ω̂|4 ∆t4 + . . . bω̂ ×c
|ω̂|2 2! 4!
µ µ ¶¶ (175)
1 1
+ −|ω̂|∆t + |ω̂|∆t − ∆t3 + |ω̂|2 ∆t5 − . . . bω̂ ×c2
3! 5!
1 1
= −I3×3 ∆t + (1 − cos(|ω̂|∆t)) bω̂ ×c − (|ω̂|∆t − sin(|ω̂|∆t)) bω̂ ×c2 (176)
|ω̂|2 |ω̂|3

TR-2005-002 18
For small values of |ω̂|, we can again take the limit and apply L’Hôpital’s rule and obtain

sin (|ω̂|∆t) ∆t (1 − cos (|ω̂|)) ∆t

lim Ψ = −I3×3 ∆t + lim bω̂ ×c − lim bω̂ ×c2 (177)
|ω̂|→0 |ω̂|→0 2|ω̂| |ω̂|→0 3|ω̂|2
cos (|ω̂|∆t) ∆t2 sin (|ω̂|) ∆t2
= −I3×3 ∆t + lim bω̂ ×c − lim bω̂ ×c2 (178)
|ω̂|→0 2 |ω̂|→0 6|ω̂|
∆t2 cos (|ω̂|) ∆t3
= −I3×3 ∆t + bω̂ ×c − lim bω̂ ×c2 (179)
2 |ω̂|→0 6
∆t2 ∆t3
= −I3×3 ∆t + bω̂ ×c − bω̂ ×c2 (180)
2 6

A note regarding the small ω approximation The error committed by using the small angle approximation is
roughly in the order of ∆tn+2 |ω̂|2 , if the non-approximated term is divided by |ω̂|n .
Therefore, at a sampling rate of 100 Hz, the error for Θ above would be in the order of 10−6 · |ω̂ thresh |2 , where
|ω̂ thresh | denotes the threshold for the small ω approximation.

2.5.2 The Noise Covariance Matrix Qd

The covariance of the noise in the discrete time system can be computed according to [4, p. 171]
Z tk+1
Qd = Φ(tk+1 , τ )Gc (τ )Qc GT T
c (τ )Φ (tk+1 , τ ) dτ (181)
tk
Z tk+1 · ¸· ¸ · ¸· ¸
Θ Ψ −I3×3 03×3 −I3×3 03×3 ΘT 03×3
= Qc dτ (182)
tk 03×3 I3×3 03×3 I3×3 03×3 I3×3 ΨT I3×3
Z tk+1 · ¸ · T
¸
−Θ Ψ −Θ 03×3
= Qc dτ (183)
tk 03×3 I3×3 ΨT I3×3
(184)

Assuming the noise to be white and independent (cf. eq. (162)), we can further expand to
Z tk+1 · ¸· 2 ¸· ¸
−Θ Ψ σrc · I3×3 03×3 −ΘT 03×3
Qd = 2 dτ (185)
tk 03×3 I3×3 03×3 σw c
· I3×3 ΨT I3×3
Z tk+1 · 2 2
¸
σr · I3×3 + σw · ΨΨT σw2
·Ψ
= 2 dτ (186)
tk σw · ΨT 2
σw · I3×3

where the last step follows from the fact, that Θ is a rotational matrix, and therefore ΘΘT gives the identity matrix.
The resulting matrix Qd has the following structure
· ¸
Q11 Q12
Qd = (187)
QT12 Q22

and the elements follow after considerable algebra as

à (|ω̂|∆t)3
!
2 2 ∆t3 3 + 2 sin(|ω̂|∆t) − 2|ω̂|∆t 2
Q11 = σr ∆t · I3×3 + σw · I3×3 + · bω̂ ×c (188)
3 |ω̂|5
à (|ω̂|∆t)2
!
2 ∆t2 |ω̂|∆t − sin(|ω̂|∆t) 2 + cos(|ω̂|∆t) − 1 2
Q12 = −σw · I3×3 − · bω̂ ×c + · bω̂ ×c (189)
2 |ω̂|3 |ω̂|4
2
Q22 = σw ∆t · I3×3 (190)

TR-2005-002 19
 Similar to the transition matrix, we can derive the form for small |ω̂| by taking the limit and applying L’Hôpital’s
rule
µ ¶
2 2 ∆t3 2∆t5 2
lim Q11 = σr ∆t · I3×3 + σw I3×3 + · bω̂ ×c (191)
|ω̂|→0 3 5!
µ ¶
2 ∆t2 ∆t3 ∆t4
lim Q12 = −σw · I3×3 − · bω̂ ×c + · bω̂ ×c2 (192)
|ω̂|→0 2 3! 4!

2.6 Propagation Equations

Having defined the propagation of the quaternion (cf. section 1.6), and the discrete time state transition- and noise
covariance matrices (cf. section 2.5), we can now write the Kalman Filter propagation equations.
Assume we receive gyro measurements ω mk and ω mk+1 , and we have an estimate of the quaternion q̄ˆk|k and the
bias b̂k|k at timestep k, together with the corresponding covariance matrix Pk|k . From this, using eq. (136), we also
have an estimate of ω̂ k|k .
We now proceed as follows:
1. We propagate the bias using the discretized form of eq. (135) as

b̂k+1|k = b̂k|k (193)

2. Using the measurement ω mk+1 and b̂k+1|k , we form the estimate of the new turn rate according to eq. (136) as

ω̂ k+1|k = ω mk+1 − b̂k+1|k (194)

3. We propagate the quaternion using a first order integrator (cf. section 1.6.2) with ω̂ k|k and ω̂ k+1|k to obtain
q̄ˆk+1|k .
4. From the formulas in sections 2.5.1 and 2.5.2 we compute the state transition matrix Φ and the discrete time
noise covariance matrix Qd .
5. We compute the state covariance matrix according to the Extended Kalman Filter equation

Pk+1|k = ΦPk|k ΦT + Qd (195)

TR-2005-002 20
3 Update
In the update phase of the Kalman filter, we will use an exteroceptive sensor to measure orientation. Common examples
are star- and sunsensors. We will now consider a simplified model of a sunsensor, a more detailed version of which is
presented in [7].

3.1 Measurement Models

In order to provide exteroceptive information to the attitude filter, we can envisage a multitude of sensors. Each is
characterized by its particular sensor model.
In the following, we will present a selection of sensors models, followed by the general procedure to fuse these
measurements with the filter estimate.

3.1.1 Sun Sensor

In this model, we assume that we are able to measure a projection of the vector from the sensor to the sun with respect
to sensor frame S r¯ , whose representation in the global coordinate frame G r¯ is known.
The relationship between the two is given by the rotation
S
r¯ = SG C G r¯ (196)

The rotational matrix SG C can be decomposed as

S
GC = SL CL
GC (197)

where the transformation SL C between sensor frame {S} and spacecraft frame {L} is known and fixed, and the trans-
formation L
G C is a function of the attitude quaternion.
The actual measurement z will be a projection of the vector r¯ , corrupted by zero-mean, white, Gaussian noise
nm
z = ΠSL CL G
G C r¯ + nm (198)
where Π is the projection matrix.
The the measurement noise is characterized by

E [nm ] = 0 (199)
£ ¤
E nm nm T = R (200)

For the update phase of the Kalman filter, we need to relate the measurement error z̃ to the state vector.
³ ´
z̃ = z − ẑ = ΠSL C L G C(q̄) − L̂
G C( ˆ
q̄ ) · G r¯ + n m (201)

From the error model (cf. section 2.3) and the properties of the rotational matrix (cf. section 1.4) we recall the
following expressions

q̄ = δ q̄ ⊗ q̄ˆ (202)

L L
G C(q̄) = G C(δ q̄ ⊗ q̄ˆ)
L
= L̂
C(δ q̄) · L̂ ˆ
G C(q̄ ) (203)

L
L̂
C(δ q̄) ≈ I − bδθ ×c (204)

TR-2005-002 21
and hence 1 ³ ´
L
G C(q̄) − L̂ ˆ
G C(q̄ ) =
L
L̂
C(δ q̄) − I3×3 · L̂ ˆ L̂ ˆ
G C(q̄ ) = −bδθ ×c · G C(q̄ ) (206)
We can now write
¡L ¢
z̃ = ΠSL C L̂
C(δ q̄) − I L̂ ˆ
G C(q̄ ) · G r¯ + nm (207)
≈ ΠSL C (−bδθ ×c) L̂ ˆ
G C(q̄ ) · G r¯ + nm (208)
= ΠSL C bL̂ ˆG
G C(q̄ ) r¯ ×c · δθ + nm (209)
h i ·δθ ¸
= ΠSL C bL̂
G C( ˆ
q̄ ) G
r¯ ×c 0 · + nm (210)
b̃
so that the measurement matrix H corresponds to
h i
H = ΠSL C bL̂ ˆG
G C(q̄ ) r¯ ×c 0
(211)

3.2 Kalman Filter Update

Given the propagated state estimates q̄ˆk+1|k and b̂k+1|k , as well as their covariance matrix Pk+1|k , the current mea-
surement z(k + 1), and the measurement matrix H, we can update our estimate in the following way:
1. Compute the measurement matrix H(k) according to eq. (211)
2. Compute residual r according to

r = z − ẑ (212)
(213)

3. Compute the covariance of the residual S as

S = HPHT + R (214)

4. Compute the Kalman gain K

K = PHT S−1 (215)

5. Compute the correction ∆x̂(+)

· ¸ · ¸
δ θ̂(+) 2 · δq̂(+)
∆x̂(+) = = = Kr (216)
∆b̂(+) ∆b̂(+)

6. Update the quaternion according to

· ¸
δq̂(+)
δ q̄ = p
ˆ (217)
1 − δq̂(+)T δq̂(+)

or, if δq̂(+)T δq̂(+) > 1, using

·¸
1 δq̂(+)
δ q̄ˆ = p · (218)
1 + δq̂(+)T δq̂(+) 1

q̄ˆk+1|k+1 = δ q̄ˆ ⊗ q̄ˆk+1|k (219)

here the difference to the expression obtained if we were using an additive instead of multiplicative error model, where q̄ = q̄ˆ + ∆q̄ with
1 Note
£ ¤T
∆q̄ = ∆qT ∆q4 :
L ˆ
G C(q̄ + ∆q̄) − L̂ ˆ T
G C(q̄ ) = 4qˆ4 ∆q4 I3×3 − 2qˆ4 b∆q ×c − 2∆q4 bq̂ ×c + 2∆qq̂ + 2q̂∆q
T
(205)

TR-2005-002 22
 7. Update the bias
b̂k+1|k+1 = b̂k+1|k + ∆b̂(+) (220)

8. Update the estimated turn rate using the new estimate for the bias

ω̂ k+1|k+1 = ω mk+1 − b̂k+1|k+1 (221)

9. Compute the new updated Covariance matrix

Pk+1|k+1 = (I6×6 − KH)Pk+1|k (I6×6 − KH)T + KRKT (222)

TR-2005-002 23
References
[1] R. O. Allen and D. H. Chang. Performance Testing of the Systron Donner Quartz Gyro (QRS11-100-420);
Sn’s 3332, 3347 and 3544. Technical Report ENGINEERING MEMORANDUM EM #343-1297, JPL, 1993.
[2] W. G. Breckenridge. Quaternions - Proposed Standard Conventions. Technical Report INTEROFFICE MEMO-
RANDUM IOM 343-79-1199, JPL, 1999.
[3] E. J. Lefferts, F. L. Markley, and M. D. Shuster. Kalman filtering for spacecraft attitude estimation. Journal of
Guidance, Control, and Dynamics, 5(5):417–429, Sept.-Oct. 1982.
[4] P. S. Maybeck. Stochastic Models, Estimation and Control, volume 1. Academic Press, New York, 1979.
[5] M. D. Shuster. A survey of attitude representations. Journal of the Astronautical Sciences, 41(4):439–517,
October–December 1993.
[6] D. Simon. Optimal State Estimation, Kalman, H∞ , and Nonlinear Approaches. John Wiley & Sons, Inc., Hobo-
ken, NJ, 2006.
[7] N. Trawny. Sun sensor model. Technical Report 2005-001, University of Minnesota, Dept. of Comp. Sci. & Eng.,
Jan. 2005.
[8] J. R. Wertz, editor. Spacecraft Attitude Determination and Control. Kluwer Academic, Dordrecht; Boston, 1978.

TR-2005-002 24