JOURNAL OF GUIDANCE, CONTROL, AND DYNAMICS

Vol. 36, No. 6, NovemberDecember 2013

Highly Efficient Attitude-Estimation Algorithm for Star

Trackers Using Optimal Image Matching
Tjorven Delabie, Joris De Schutter, and Bart Vandenbussche
Katholieke Universiteit Leuven, 3001 Heverlee, Belgium

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

DOI: 10.2514/1.61082
This paper presents a novel attitude-estimation algorithm for spacecraft using a star tracker. The algorithm is
based on an efficient approach to match the stars of two images optimally on top of each other, hence the name of the
algorithm: attitude estimation using image matching. In tests, attitude estimation using image matching proved to be
as robust as the most robust existing methods, and faster than the fast iterative methods. On top of this, attitude
estimation using image matching allows, in a lot of cases, the elimination of a computationally intensive coordinate
conversion, which normally precedes the attitude-estimation algorithm. The computational cost of this conversion
step is several times higher than that of the attitude-estimation algorithm itself, and so this elimination yields a huge
increase in efficiency as compared to the existing algorithms. This significant reduction in computational cost allows to
obtain the attitude estimates at a higher rate, implement more accurate centroiding algorithms, or use more stars in
the attitude-estimation algorithm, all of which improve the performance of the attitude estimation. It could also allow
the use of star trackers in the expanding field of small-satellite projects, in which satellite platforms have a limited
computational capability.

Nomenclature
A

bi

b; l

Ecent
F

i; j; k
^
^ j;
^ k

i;
^
^

ir ; jr ; k^r 

=
=
=
=
=

ki , hi

kiin , hiin

NEAcross
npix
ns

=
=
=

Pe

qa

qdat

qdiff

orthogonal attitude matrix, which minimizes the

Wahba cost function
unit vector of star i, observed in the spacecraft
body frame
maximum values for the x and y coordinates of
the focal plane
average centroiding accuracy
focal length of the optical system
unit-vector coordinates of an observed star
unit-vector coordinates of a database star
unit-vector coordinates of a database star, after
rotation by the inverse of qdat
parameters of the polynomials to transform the
measured coordinates to corrected coordinates
for the camera distortions
parameters of the polynomials to transform the
corrected coordinates for camera distortions
to measured coordinates
cross-boresight noise equivalent angle
number of pixels in one row of the image
number of stars used in the attitude-estimation
algorithm
factor with which the variance of the position
error was multiplied for outliers
quaternion representing the true attitude of
the spacecraft
quaternion representing the attitude at which
the database image is selected
quaternion representing the estimated difference
in attitude between qa and qdat

qe

ri
tx

=
=

ty

wi

xi ; yi 

x^ i ; y^ i 

x0 ; y0 
r

=
=
=
=
=

=
=
=
=
=

quaternion representing the estimated attitude

of the satellite
unit vector of star i, observed in a reference frame
distance over which the database stars are
translated in the x direction
distance over which the database stars are
translated in the y direction
nonnegative weight of star i
coordinates of observed star i in the focal plane
coordinates of database star i in the focal plane
intersection of the focal plane and the optical axis
angle in arcseconds over which the coarse estimate,
qdat , is rotated away from the true attitude over
three axes
cost value
field of view of the camera image
pitch angle
roll angle
yaw angle

I.

Introduction

N MANY spacecraft missions, accurate knowledge of the

orientation of the spacecraft in space is crucial. Examples are
space telescopes observing an astronomical target, or communication
satellites that need to accurately point an antenna to a ground station.
Several sensors to determine this orientation, also referred to as the
attitude of the satellite, have been developed. The most accurate of
these sensors is the star tracker. This sensor takes an image of the
surrounding star field and compares it to a database of known star
positions. Typically, this sensor can determine the attitude of the
satellite with an accuracy in the range of a few arcseconds [1] or even
subarcsecond [2].
An autonomous star tracker operates in two different modes [3].
The first mode is the initial attitude acquisition. In this mode, the
star tracker determines the attitude of the satellite without a priori
knowledge [4,5]. Once an initial attitude has been acquired, the star
tracker switches to the tracking mode. In this second mode, in which
the star tracker has a priori knowledge, the database search can be
limited, allowing fast and accurate attitude estimation.
A variety of attitude-estimation algorithms that can be used in
this tracking mode have been proposed. All of these estimate the
spacecraft attitude from vector measurements, and seek the matrix
that minimizes the loss function proposed by Wahba [6]:

Presented as Paper 2012-4478 at the AIAA Guidance, Navigation, and

Control Conference, Minneapolis, Minnesota, 1316 August 2012; received 5
November 2012; revision received 7 January 2013; accepted for publication 7
January 2013; published online 1 July 2013. Copyright 2013 by Tjorven
Delabie. Published by the American Institute of Aeronautics and
Astronautics, Inc., with permission. Copies of this paper may be made for
personal or internal use, on condition that the copier pay the $10.00 per-copy
fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers,
MA 01923; include the code 1533-3884/13 and $10.00 in correspondence
with the CCC.
*Ph.D. Researcher, Department of Mechanical Engineering, Celestijnenlaan 300B. Member AIAA.

Professor, Department of Mechanical Engineering, Celestijnenlaan 300B.

Professor, Institute for Astronomy, Celestijnenlaan 300B.

1672

1673

DELABIE, SCHUTTER, AND VANDENBUSSCHE

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

L
A 

n
1X
w jb Ari j2
2 i1 i i

A. Centroiding

(1)

in which bi is a unit vector observed in the spacecraft body frame, ri is

a corresponding unit vector in a reference frame, and wi is a
nonnegative weight.
The most robust of these attitude-estimation algorithms are
Davenports q method [7] and the singular-value-decomposition
(SVD) method [8]. These methods are slow, but are based on robust
algorithms to calculate the symmetric eigenvalue problem and the
SVD [9,10]. Fast iterative solutions to Wahbas problem have been
developed [11]. These solutions solve the characteristic equation for
the maximum eigenvalue, and use this value to construct the optimal
attitude quaternion. This method was first used in QUEST [12,13],
which is still the most widely used algorithm to solve Wahbas
problem. QUEST was followed by FOAM [14], ESOQ [15], and
ESOQ2 [16]. The main goal of the algorithms following QUEST was
to improve the computational efficiency. However, the improvements
are small, if any, and so these algorithms could be considered to be
equally fast [11].
These attitude-estimation algorithms all share one important
drawback when used with a star tracker. They determine the attitude
of the spacecraft based on observations, which are represented as unit
vectors. In a star tracker, however, the observations are twodimensional (2-D) star centroids on the focal plane. The conversion
of these 2-D coordinates to unit vectors is not straightforward.
Because of optical and electronic distortions, temperature, and
magnetic and star intensity effects, an empirical model based on
laboratory calibrations is often used to convert the coordinates [17].
In this paper, an algorithm is proposed, which is faster than the
algorithms mentioned previously, and more important, in some cases,
eliminates the need for the computationally intensive coordinate
conversion during each attitude-determination step. This way, the
computational cost of the entire attitude-estimation procedure is
highly reduced. This increased efficiency allows us to obtain
the attitude information at a higher rate, or track more stars,
improving the accuracy of the attitude estimation. Furthermore, in the
growing market of small satellites, this method could reduce the
computational expense of the spacecraft bus, allowing to carry a more
demanding payload.
In Sec. II, the algorithm, referred to as attitude estimation using
image matching (AIM), is presented. Section III discusses the
accuracy, speed, and robustness of AIM, compared to the state-ofthe-art algorithms. These comparisons show that the speed of AIM
greatly exceeds that of existing algorithms, while obtaining similar
accuracy and robustness provided by the robust estimators.

The first step in the sequence of algorithms is to determine the

centroids of the observed stars in the camera image. Because the
previous attitude and rate information is known in tracking mode, the
position of stars in the image can be predicted, based on the starimage coordinates of the previous image [3]. This prediction step
eliminates the need to digitize the entire image, and instead allows to
only digitize small windows at the predicted star positions. In those
windows, a centroiding algorithm [18,19] is used to accurately
determine the centroids of the stars in the camera image.
B. Star Identification

The identification of the image stars (i.e., determining to which

database star each observed star corresponds) is referred to as star
identification. Samaan et al. presented two methods for star
identification in [20]. In this paper, an adaptation of the staridentification algorithm of [5] was used. This algorithm calculates the
similarity between the camera image and preprocessed database
images using the shortest distance transform. The database image
with the highest similarity is retained, and the stars can be identified
this way. This algorithm is significantly more robust than other staridentification algorithms. In addition, it is fast and gives a very
reliable performance value, which indicates whether or not the stars
have been correctly identified. In this adaptation, the database images
used in [5] are sorted so that the images lying closest to the previous
attitude are checked first. The algorithm is stopped when one of the
database images yields a performance value above a predetermined
threshold. This is usually already the case in the first image, which
makes this approach extremely fast, on top of being very robust to
distortions in the camera image.
C. Coordinate Conversion

While the previous steps were the same for both AIM and the
existing algorithms, in this step, there is an important difference.
First, the classical approach is described, followed by the approach
used by AIM.
1. Classical Approach

In the current state-of-the-art algorithms, the attitude is determined

using unit vectors
i; j; k. Therefore, the next step is to convert the
focal-plane coordinates
x; y of the observed stars to unit vectors.
This can be done using a simple pinhole model as described in [17]:
8
xx0
F
>
>
xx0 2
yy0 2
> i  p




























1

>
F  
F 
<
yy0
F
j  p




























xx
yy
>
1
F 0 2 
F 0 2
>
>
1
>
: k  p




























xx0 2
yy0 2
1

II.

Algorithm

In this section, the AIM algorithm will be derived. Before

discussing the attitude-estimation algorithm itself, it is important to
examine the preceding algorithms, which calculate the input for the
actual attitude-estimation algorithm. These algorithms require a
number of calculations, which are an order of magnitude higher than
that of the attitude-estimation algorithm itself. Therefore, it is
important to examine the influence of the attitude-estimation
algorithm on these preceding steps, in view of the comparison with
other attitude-estimation algorithms. This section starts with an
overview of the algorithms that lead to the input of the attitudeestimation algorithm (as seen in Fig. 1), before deriving the AIM
algorithm itself.

 

(2)

In these equations,
x; y is the coordinate of the star in the image;

x0 ; y0  is the intersection of the focal plane and the optical axis;

i; j; k is the unit vector of the observed star, described in the sensor
reference frame; and F is the focal length of the optical system.
To account for distortions, however, the image-plane coordinates
are first corrected for distortions using a fifth-order polynomial [17]:
xc  k0  k1 x  k2 y  k3 x
x2  y2   k4 x
x2  y2 2 k5 x2
k6 xy k7 y2

(3)

yc  h0  h1 y  h2 x  h3 y
y2  x2   h4 y
y2  x2 2 h5 y2
h6 yx h7 x2

Fig. 1 Sequence of algorithms to estimate the spacecraft attitude with a

star tracker.

(4)

in which
xc ; yc  are the coordinates after they have been corrected
for camera distortions, and the parameters
ki ; hi  depend on the
camera. The coordinates
xc ; yc  are then converted using Eq. (2).

1674

DELABIE, SCHUTTER, AND VANDENBUSSCHE

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

2. Approach of AIM

The first difference between the state-of-the-art algorithms and

AIM is that AIM determines the attitude using focal-plane
coordinates. While the existing algorithms convert the observed
focal-plane coordinates to unit vectors, AIM will convert the unit
vectors of the database stars, which were selected in the staridentification step, to focal-plane coordinates. The fact that the
conversion is done on the database stars (of which the coordinates do
not change) yields a significant increase in efficiency. This is
explained and quantified in Sec. III.A.
To convert the database unit vectors to coordinates in the focal
plane, the selected database unit vectors first need to be rotated so that
they are centered around the k axis. To perform this rotation, a
quaternion qdat is selected. This quaternion is a coarse estimate of the
current attitude, being the attitude in which the unit vectors are
observed. This leads to the second distinction between AIM and the
state-of-the-art algorithms: AIM requires a coarse estimate of the
current attitude to be able to estimate the current attitude, whereas
the existing algorithms do not. In the tracking mode, the attitude
estimation obtained in the previous time step can be used as qdat
because the attitude changes between subsequent time steps are
generally small. This is further discussed in Sec. III.B.2. In the lostin-space mode, the star-identification algorithm of [5] was used. This
algorithm identifies the stars in the field of view (FOV), and also
presents a coarse estimate of the attitude. This estimate is then used
as qdat .
In the first step of the coordinate conversion for AIM, the unit
vectors are rotated over the inverse of qdat to center them around the k
axis. Because qdat is a unit quaternion, the inverse is equal to its
^ over the
^ j;
^ k
conjugate. To rotate the database unit vector v 
i;
inverse of the quaternion qdat , Eq. (5) is used.
vr  q
dat vqdat

(5)

^ is the unit vector of a star as found in the

^ j;
^ k
In this equation, v 
i;
database, vr 
i^r ; j^r ; k^r  is that unit vector after rotation by the
inverse of qdat , and a bar over a quaternion indicates that the conjugate
of the quaternion is taken.
After this, the conversion from unit-vector coordinates to database
^ y
^ is performed. Using the pinhole model,
focal-plane coordinates
x;
this conversion is performed with the following equations:


x^  x0  Fi^r k^r
y^  y0  Fj^r k^r

(6)

^ y
^ are the focal-plane coordinates an
The resulting coordinates
x;
ideal star tracker would observe if its attitude was qdat . Because a real
star tracker has camera distortions, these coordinates are transformed
to account for these distortions. Using the inverse transformation of
^ y
^ are transformed to the values as
Eqs. (3) and (4), coordinates
x;
they would be measured with the distortions in the camera. The
polynomials to calculate this transformation are given in Eqs. (7)
and (8).
^ x^ 2  y^ 2   k4in x

^ x^ 2  y^ 2 2
x^ d  k0in  k1in x^  k2in y^  k3in x

k5in x^ 2 k6in x^ y^ k7in y^ 2

Fig. 2

(7)

^ y^ 2  x^ 2   h4in y

^ y^2  x^ 2 2
y^ d  h0in  h1in y^  h2in x^  h3in y

h5in y^ 2 h6in y^ x^ h7in x^ 2

(8)

In these equations,
x^ d ; y^ d  are the database-star coordinates after

the camera distortions have been added, and
kiin ; hiin  are the
parameters that are specific for the camera.
D. Attitude-Estimation Algorithm

Once the preceding steps have been finished, the actual attitudeestimation algorithm is executed. At the heart of AIM lies an efficient
optimization to find the transformation that matches the stars of the
camera-star image and of the database-star image optimally on top of
each other. This transformation is then used to determine the
estimated difference qdiff between the attitude at which the camerastar image was taken (being the true attitude qa ) and the attitude at
which the database-star image was selected qdat . Finally, the
estimated attitude of the satellite qe is calculated from this.
1. Construction of the Cost Function

To find the transformation that optimally matches the stars of two

images on top of each other, a cost function that needs to be
minimized is constructed. This cost function is the sum of the
Euclidean distances squared between each pair of image star and their
corresponding transformed database star. These stars were paired
with each other in the star-identification step. Each image star was
identified and matched with a star in the database. If no match is
found, the star is discarded so that the number of image stars and
database stars in this cost function is always the same.
The distance is multiplied by a weight that is specific for each star
pair. This weight could be determined based on, for example, a
confidence measure for the centroid, calculated in the centroiding
algorithm. This way, a star of which the centroid has been determined
with higher confidence in the centroiding algorithm could be given
more value in the tracking step. The cost function is constructed as
follows:

; tx ; ty  

ns
X

wi fxi x^ i cos
  y^ i sin
 tx 2

i1

 yi x^ i sin
 y^i cos
 ty 2 g

(9)

In this function, wi is the weight given to star i;
xi ; yi  are the

coordinates of the observed star i in the focal plane;
x^ i ; y^i  are the
coordinates of the corresponding database star i in the focal plane;
is the angle over which the database stars are rotated with respect to
the origin of the frame in which the database coordinates are
described; tx and ty are the distances over which the database stars are
translated in the x and y directions, respectively; and ns is the number
of stars used in the attitude-estimation algorithm. This is depicted in
Fig. 2. In this figure, the four stars of the camera image
xi ; yi  are
visible in a lighter shade and filled on the bottom-left corner. The
database stars
x^i ; y^ i  are depicted in black. In Fig. 2b, the database
stars are rotated over an angle . In Fig. 2c, they are translated over a
distance tx . Finally, they are translated over a distance ty in Fig. 2d.
The total distance between the camera stars and the transformed
database stars, as seen in Fig. 2d, is the cost function. The database
stars of the previous image are shown in dashed lines.

Transformation to minimize the distances between camera and database stars.

1675

DELABIE, SCHUTTER, AND VANDENBUSSCHE

2. Minimization of the Cost Function

To achieve an optimal matching of the transformed database

stars with the camera-image stars, the total distance between the
corresponding stars needs to be minimized:
min
; tx ; ty 

(10)

The unknown variables , tx , and ty , which minimize this cost

function, can be found by calculating the derivative of this cost
function to each of the unknowns, setting the three obtained
equations equal to zero and solving this system of three equations.

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

8 d
>
< d  0
d
dtx  0
>
: d  0
dty

(11)

in which is the camera FOV. Because most calculations in the

attitude-determination-and-control system are done using quaternions, these Euler angles are then converted to a quaternion, the
difference quaternion qdiff . To convert Euler angles to a quaternion,
Eq. (19) is used:
3
2
cos
h  cos
h  cos
h   sin
h  sin
h  sin
h 
6 sin
h  cos
h  cos
h  cos
h  sin
h  sin
h  7
7
qdiff  6
4 cos
h  sin
h  cos
h   sin
h  cos
h  sin
h  5
cos
h  cos
h  sin
h  sin
h  sin
h  cos
h 
(19)
in which h , h , and h are the half of the angles , , and ,
respectively. Because these angles are small, this conversion is
simplified to Eq. (20):
0

Generally, solving this system of three equations can take a

significant amount of work and lead to very involved expressions, but
this is not the case for the AIM algorithm. By substituting the second
and third equations into the first, the three transformation variables,
which optimally map the database-image stars on top of the cameraimage stars, can be explicitly calculated immediately. This is an
extremely fast procedure. Moreover, because no equation solving or
complex calculations are used, this procedure is very robust. These
three explicit relationships are given in the following equations:

qdiff

1
1  h h h
B h h h C
C
B
@ h  h h A
h h h

(20)

This quaternion represents the estimated rotation needed to rotate the

database quaternion qdat to the true-attitude quaternion qa . Finally, by
multiplying qdiff with qdat , the estimated quaternion qe is obtained:
qe  qdiff qdat

(21)

^
 arctan 2
sx:sy^ sy:sx^ sxy^  syx;
^
sx:sx^ sy:sy^  sxx^  syy

(12)




b
b
b
tx  sx : cos
  sy : sin
  sx^
2
2
2

(13)





l
l
l
ty  sx : sin
  sy : cos
  sy^
2
2
2

(14)

Here, b and l are the maximum values the pixel coordinates can have
in the focal plane;
b2; l2 is therefore the coordinate of the center
of the focal plane. The other variables in these three equations are
calculated as follows:
sx 

ns
X

wi :xi

sxx^ 

n1

sxy^ 

ns
X

ns
X
n1

wi :xi :x^ i

sy 

n1

wi :xi :y^ i

sx^ 

n1

syx^ 

ns
X

ns
X

ns
X

wi :yi

n1

wi :x^i

(15)

n1

wi :yi :x^ i

sy^ 

ns
X

wi :y^i

syy^ 

n1

ns
X

wi :yi :y^i

4. Small-Angle Approximation

The three Euler angles calculated in Eqs. (1618) represent the

difference between the estimated satellite attitude and the attitude of
the database image. This database image is taken at a previously
estimated satellite attitude qdat . Because the updating frequency of a
star tracker is generally between 1 and 10 Hz, and because the
rotational velocity of a satellite is generally under 1 deg s, the
difference between the estimated satellite attitude qe and the attitude
of the database image qdat is generally smaller than 0.1 deg around
each axis. Because the Euler angles are close to zero, the
trigonometric functions in Eqs. (1214), (17), and (18) were
approximated by their Taylor series up to the first order:
sin


(22)

cos
 1

(23)

arctan


(24)

Up to an angle of 0.25 deg, the error remains under one-thousandth

of a percent, which is negligible for this application. Because
trigonometric functions are computationally expensive, this
approximation was implemented in the AIM algorithm.

n1

III.
3. Calculation of the Attitude Quaternion

The three transformation values, which are readily calculated from

Eqs. (1214), are then converted to three Euler angles:


(16)



ty
 arctan :
l

(17)



t
 arctan : x
b

(18)

Testing

The performance of AIM was determined during a series of tests,

which validated the speed, accuracy, and robustness of AIM. These
results are compared to the results of QUEST and the q-Davenport
method, respectively, one of the fast iterative solutions and one of the
robust solutions. These three algorithms were implemented in
MATLAB and C++. The implementation of QUEST and the qDavenport method is based on the formulas given in [6]. For the
QUEST algorithm, the adapted version of [21] was used because this
solved some problems with robustness. Stars up to a magnitude of 6
are used in the calculations. The tests were performed on a Dell
Latitude laptop with a 2.60 GHz Intel Core i7 processor and 8 GB
RAM. The C++ code was executed in Microsoft Visual C++ 6, with
the /O2 flag to maximize speed and the /ML library compiler option
for static single-threaded libraries.

1676

DELABIE, SCHUTTER, AND VANDENBUSSCHE

Table 1

Floating-point operations for the different

attitude-estimation algorithms
AIM

ns
4
9
25

125
205
461

Zeroa
188
293
629

QUEST
Onea
225
330
666

Twoa
236
341
677

Number of iterations.

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

Table 2 C++ execution times for the different

attitude-estimation algorithms
ns

AIM, s

4
9
25

0.13
0.19
0.36

QUEST, s
Zeroa Onea Twoa
0.20 0.23 0.24
0.29 0.31 0.33
0.57 0.58 0.60

q-Davenport, s
5.38
5.58
5.67

Number of iterations.

A. Speed

A faster attitude-estimation algorithm presents the advantage of

allowing the control system to achieve the attitude information at a
higher rate [22]. For this reason, the algorithms following QUEST
were developed to increase the speed. A good overview of the speed
of the most common estimation algorithms can be found in [11].
1. Flop Count

The amount of floating-point operations (flops) for each of the

existing algorithms was determined in [11]. In Table 1, the amount of
flops for AIM is added to those results. The amount of flops is a
function of the number of stars ns used in the tracking algorithm. For
QUEST, it is also a function of the number of NewtonRaphson
iterations. In the q-Davenport method, the symmetric eigenvalue
problem is solved. This can be done with different algorithms, which
are generally iterative in nature. Because of this, a straightforward
result cannot be given for the q-Davenport, and it is therefore omitted
from Table 1.
From Table 1, we see that, thanks to the very efficient procedure of
AIM, the number of flops needed is lower than the flops needed by

QUEST. When 25 stars are used, AIM reduces the number of flops
with 25% compared to QUEST.
2. C++ Execution Times

The execution time of AIM, QUEST, and q-Davenport was

measured using the C++ timer QueryPerformanceCounter during
simulations in C++. Simulations of 100 million estimations were run
with a different number of stars in the image (ns ). For QUEST,
simulations were run with a different number of NewtonRaphson
iterations to calculate the largest eigenvalue. The calculation times
per execution are given in microseconds in Table 2.
AIM is significantly faster than both QUEST and q-Davenport.
The computational time is around one-third lower than that of the fast
method, QUEST, and an order of magnitude lower than that of the
robust method, q-Davenport. The calculation time also rises slower
when the number of stars increases. This is because AIM estimates
the attitude of the satellite using 2-D coordinates instead of threedimensional (3-D) coordinates. Another important reason for this
higher speed is that AIM does not require computationally intensive
operations (such as an eigenvalue calculation for q-Davenport).
Because the coordinate-conversion step is different for AIM and
the state-of-the-art algorithms, it is also important to assess the speed
of the algorithms with the coordinate-conversion step included. This
is quantified in Table 3, which shows the speed results of AIM,
QUEST, and q-Davenport in microseconds, with the coordinateconversion step included. To convert the coordinates, a pinhole
model was used. For AIM, this is Eq. (6); for QUEST and qDavenport, this is Eq. (2). Camera distortions were also taken into
account in this step using the polynomials of Eqs. (7) and (8) for AIM,
and Eqs. (3) and (4) for QUEST and q-Davenport. The parameters of
these polynomials were chosen to be the same as the parameters used
in the Herschel star tracker, to get a realistic result [17]. These
parameters are given in Table 4.
The coordinate-conversion step precedes the attitude-estimation
algorithm and is different for AIM and the state-of-the-art methods.
When the calculation time of this coordinate conversion is taken into
account, it becomes clear from Table 3 that the procedure of AIM is
faster than that of QUEST and q-Davenport. The calculation time of
the AIM procedure is around one-third lower than that of QUEST. It
can be concluded that AIM yields a large speed increase.
In some cases, AIM allows to eliminate the coordinate-conversion
step, this way speeding up the procedure greatly. This is discussed in
the next section.
3. Elimination of a Preceding Algorithm

Table 3 C++ execution times for the different

attitude-estimation algorithms with the coordinateconversion step included
ns

AIM, s

4
9
25

0.36
0.62
1.48

QUEST, s
Zeroa Onea Twoa
0.52 0.54 0.56
0.92 0.95 0.97
2.25 2.27 2.28

q-Davenport, s
5.74
6.26
7.42

Number of iterations.
Table 4
i
0
1
2
3
4
5
6
7

When the tracked stars remain the same for some period, the same
transformed database stars can be used for several exposures. This
means the unit vectors of the database stars can be converted to image
coordinates one time, and be used for a number of following
exposures. This in contrast to the current algorithms, in which the
camera-star coordinates need to be converted to unit vectors every
time. When the same database stars can be used for several exposures,
the algorithm sequence of Fig. 1 changes for the AIM algorithm to the
shorter algorithm sequence of Fig. 3. The database image then only
needs to be updated when stars enter or leave the focal plane, or when
the observed stars and database stars have strayed too much from
each other. In the next section, it is validated when a database image
should be updated in this case.

Parameters of the polynomials used in the camera-distortion model

ki
1.27866006e 5
1
8.57471666e 5
2.19016482e 4
5.20164707e 7
6.53773139e 5
5.29575523e 5
2.24629447e 5

hi
2.0675277e 5
1
1.60933572e 5
2.14748994e 4
2.26158663e 7
4.54670784e 5
4.89014428e 5
2.27073830e 5

kiinv
1.3025586e 4
0.99999929
8.5120363e 5
2.1871748e 4
6.3269083e 7
6.4147788e 5
5.2131004e 5
2.2034513e 5

hiinv
1.875591e 5
0.99999960
1.6173263e 5
2.1458371e 4
3.4805732e 7
4.4630672e 5
4.8010159e 5
2.2337249e 5

1677

DELABIE, SCHUTTER, AND VANDENBUSSCHE

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

Fig. 3 Shortened sequence of algorithms during database reuse.

The increased efficiency is especially important when the star

tracker is in fine-pointing mode. In this mode, the payload of the
satellite needs to stay fixed on a certain point for a long time, and the
attitude of the satellite changes slowly. It is in this mode, which is very
common for satellites, that AIM can very efficiently use the same
database image for a very long period of time, this way totally
eliminating a very computationally intensive calculation.
In Table 5, the calculation times of AIM are compared to those of
the state-of-the-art algorithms, for the case in which AIM can reuse
its database image. Because of the relatively large computational
expense of converting the coordinates in the state-of-the-art approaches, the speed increase yielded by AIM is very high in this case.
When 25 stars are tracked, the time needed to estimate the spacecraft
attitude after the star identification is more than six times lower for
AIM than for the state-of-the-art fast method, QUEST.
B. Accuracy

The more accurate a star tracker is, the more precisely it allows a
spacecraft to point the payloads to specific targets. This allows for
better observations and more valuable data. A star tracker can
typically determine the attitude of the spacecraft with an error of a few
arcseconds or even subarcsecond cross boresight, and an error that is
typically 616 times larger around the roll axis [3].
1. Test Setup

For each simulation of 10,000 star-tracker exposures distributed

randomly over the sky, values were chosen for the FOV , the
accuracy of the centroiding Ecent , the number of pixels in one row of
the camera npix , and the number of stars used in the estimation
algorithm ns . These values will always be mentioned when the results
are presented. To test the accuracy, noise was added to the camera
image. This was done by adding Gaussian white noise with a variance
equal to Ecent to the known correct pixel coordinates. To simulate
camera distortions, the correct pixel coordinates were transformed
using the polynomials of Eqs. (7) and (8) with the parameters of
Table 4.
2. Projection Distortions

The AIM algorithm optimally matches 2-D star images on top of

each other. A problem that arises here is that some distortion is always
induced when a 3-D image is projected onto a plane [23]. This means
that the position of stars relative to each other is slightly different
when the star image is taken with a different camera center. These
distortions are larger if the difference between the centers is larger.
The projection distortions decrease the accuracy when the difference
between the attitude at which the camera image was taken qa and the
attitude at which the database stars were taken qdat increases. In other
words, the accuracy decreases when the coarse estimate of the current
attitude, qdat , is further off from the true attitude qa . It will be shown,
however, that this decrease in accuracy is insignificant in a realistic
environment.

To quantify this error, the increase in rms errors of AIM is shown

with respect to the optimal q-Davenport solution as a function of the
difference in attitude of database and image center in Table 6.
The attitude-difference value r represents the difference in
arcseconds between the Euler angles , , and of the attitude at
which the image was taken qa and the attitude at which the database
stars were selected qdat . The last three columns show the percentage
with which the rms attitude-angle errors of AIM are higher than the qDavenport error for the three attitude angles. (The first two are cross
boresight and the last column is roll.) Up to a difference of
100 arcseconds, the rms errors for both methods are almost identical.
This means that AIM and q-Davenport output an equally accurate
estimate of the attitude, when the difference in Euler angles of the
actual attitude and the attitude at which the database was selected is
lower than 100 arcseconds. With a difference in Euler attitude angles
of 500 arcseconds, AIM yields an error in cross boresight, which is
around 1% higher than that of q-Davenport. The error around the roll
axis only slightly increases.
To validate the practical relevance of this effect, we note that during
normal operation of the star tracker, the database stars are selected
around the previous attitude. The difference between the attitude at
which the star image was observed (the current spacecraft attitude)
and the attitude at which the database was taken (the previous
spacecraft attitude) depends on the rotational speed of the spacecraft
and the frequency at which images are taken.
As a real-life example, from the technical specifications of the CT602 star tracker of Ball Aerospace [24], it can be seen that the star
tracker offers full performance up to a tracking rate of 0.3 deg s, and
reduced performance up to 1.5 deg s while having an update rate of
10 Hz. Therefore, the difference in attitude between two consecutive
exposures can, for this star tracker, maximally be 0.3 deg s
3600 arcseconds deg 10
1s  108 arcseconds with full
performance, and 0.5 deg s 3600 arcseconds deg 10
1s 
540 arcseconds when reduced performance is accepted.
With these differences, it can be concluded from Table 6 that the
errors induced by the projection distortions are insignificant, even for
worst-case scenarios.
The results of this table can be used to assess which amount of
attitude difference is allowed before a new database image is selected.
To increase the speed, the same database image should be used
as long as possible, to eliminate the computationally intensive
coordinate conversion. A threshold value could be determined so that
a new database image is created when the attitude difference exceeds,
for example, 500 arcseconds, to keep the estimation error within 1%
of the optimal estimation error. As can be seen from the real-life
example given previously, this amount of allowed attitude difference
should allow to use the database image for several exposures, even
when a fast maneuver is executed. In the case of the fine-pointing
mode, the attitude difference remains very small (in the order of a few
arcseconds), so that the same database image can be used for a very
long time and induced errors are definitely insignificant.

Table 6 Percentage of the rms attitude-angle-error increase of AIM

compared to the q-Davenport error as a function of the difference in
attitude at which the image was observed and the attitude at which the
database stars were takena
Attitude difference r , arcsecond

Table 5 C++ execution times for the different

attitude-estimation algorithms when AIM can reuse its
database image
ns

AIM, s

4
9
25

0.13
0.19
0.36
a

QUEST, s
Zeroa Onea Twoa
0.52 0.54 0.56
0.92 0.95 0.97
2.25 2.27 2.28

Number of iterations.

0
10
100
200
500
700

q-Davenport, s
5.74
6.26
7.42

0
0
0
0.19
1.33
4.64

Error increase, %

0
0
0
0.11
0.56
2.97

0
0
0.01
0.06
0.09
0.18

 8 deg, npix  1024, Ecent  0.5, leading to an rms error of

around 4.9 arcseconds in cross boresight, and around
90 arcseconds around the roll axis.

1678

DELABIE, SCHUTTER, AND VANDENBUSSCHE

Table 7
Scenario

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

 8, npix  1024,
Ecent  0.5, ns  9
 8, npix  1024,
Ecent  0.1, ns  9
 8, npix  2048,
Ecent  0.5, ns  9
 8, npix  4096,
Ecent  0.5, ns  9
 8, npix  1024,
Ecent  0.5, ns  15
 8, npix  1024,
Ecent  0.1, ns  15

RMS attitude-angle errors of AIM, QUEST, and q-Davenport in various scenarios

RMS error
, arcsecond
4.94

AIM
RMS error
, arcsecond
4.95

QUEST
RMS error
, arcsecond
4.95

RMS error
, arcsecond
88.03

RMS error
, arcsecond
4.94

q-Davenport
RMS error
, arcsecond
4.95

RMS error
, arcsecond
87.98

1.00

0.95

18.86

0.99

0.97

19.11

0.99

0.97

18.87

2.57

2.48

44.96

2.57

2.49

44.95

2.57

2.48

44.94

1.20

1.24

23.67

1.20

1.26

23.62

1.20

1.26

23.62

3.69

3.74

66.66

3.68

3.76

67.34

3.68

3.75

66.67

0.77

0.74

13.44

0.77

0.77

13.42

0.77

0.77

13.42

RMS error
RMS error
, arcsecond , arcsecond
87.99
4.94

3. Accuracy Results

of a star can be determined, typically ranging from 0.05 to 0.5), and

npix is the number of pixels in the image.
Equation (25) presents us with a convenient calculation to estimate
the accuracy of the attitude estimation, given the number of pixels and
FOV of the camera, the number of stars used, and the centroiding
accuracy.
From Eq. (25), it can be deduced that improved centroiding is
linearly proportional to the accuracy. The accuracy also scales
inversely proportional with the root of the number of stars used in the
tracking algorithm. Using Eq. (25), it is estimated that using an extra
star yields a decrease in error, which can be calculated as

The accuracy of AIM, QUEST, and q-Davenport was tested during

simulations of 10,000 star images distributed randomly over the sky.
The values that are given for each algorithm and each simulation
scenario are the rms errors around the three axes. For QUEST, no
NewtonRaphson iterations were used because these are not
necessary to improve the accuracy according to Shuster [25]. For
AIM, the difference between the attitude at which the image was
taken and the attitude at which the database stars were selected was
chosen to be 100 arcseconds around each axis. In other words, the
coarse estimate of the attitude qdat was rotated 100 arcseconds around
each axis from the known true attitude. In Sec. III.B.2, it was
calculated that this is the largest difference (worst-case scenario)
at which the example star tracker is still required to offer full
performance. In realistic scenarios, this difference will be a lot
smaller.
The experiments of which the results are depicted in Table 7 show
that the accuracy of AIM is almost exactly the same as the accuracy of
QUEST and q-Davenport. The deviation is in the orders of onehundredths of an arcsecond. While QUEST is expected to perform
equally good as q-Davenport, there is a slight reduction in accuracy.
This is because the systematic error induced by the camera distortions
lowers the accuracy of QUEST, which is less robust to such a
systematic error. The robustness of QUEST will be discussed more
in-depth in Sec. III.C. If one NewtonRaphson iteration is performed,
the accuracy of QUEST is equal to that of q-Davenport, at the cost
of an increase in computational time. From these results, we can
conclude that AIM is as accurate as the state-of-the-art algorithms.

This decrease in error is a function of the number of stars used in

the tracking algorithm. This accuracy improvement was also
implemented in MATLAB and verified in simulations. During these
simulations, AIM determined the attitude using one or two extra stars.
The results are given in Table 8. Simulation results to validate the
accuracy improvement caused by better centroiding can be found in
Table 7, in which the results of simulations performed with different
values for Ecent are shown.
The obtained error reduction corresponds well with the predicted
error reduction of Eq. (26). The decrease in error of AIM compared to
QUEST or q-Davenport is in the order of a few arcseconds now
(around the roll axis), which is a significant improvement.

4. Accuracy Improvements

C. Robustness

One way to benefit from the large increase in the computational

efficiency of AIM is to improve the accuracy of the attitude
estimation. This can be done, on the one hand, by using the increased
efficiency to increase the speed of the attitude estimation and by
getting the attitude estimation at a higher frequency. This improves
the performance of the entire attitude-determination-and-control
system, but this effect is hard to quantify in tests. On the other hand,
because the computational load is decreased significantly, there is
room to implement changes in the algorithms, such as an improved
centroiding algorithm or using more stars in the tracking algorithm.
An error representation, which is interesting to give an order-ofmagnitude idea of the effect of these accuracy improvements, is the
noise equivalent angle (NEA). This is the star trackers ability to
reproduce the same attitude provided the same optical stimulation.
This error is independent of software, algorithmic errors, and
calibration [3]. The NEA is approximately

A distortion in the optics of the star camera, dead pixels, or errors in

the centroiding algorithm could lead to a large error in the calculation
of the centroid of one or more stars. This is especially so for smaller
satellites, in which there is less budget to buy expensive optics, and
there is less redundancy and room for error checks to limit the number
of calculations. It is therefore important that the attitude-estimation
algorithm is robust to such errors. This is validated in the first section.
As opposed to the existing attitude-estimation algorithms, AIM
requires a coarse estimate of the attitude qdat to be able to estimate the
attitude of the spacecraft. The estimated attitude of the previous time
step can be used as a coarse estimate. During large slew maneuvers,
this coarse estimate might be far off because the spacecraft is rotating
rapidly. In the second section, the robustness of AIM is validated for
increasingly inaccurate coarse estimates.

NEAcross 

:Ecent
p





npix : ns

(25)

in which NEAcross is the cross-boresight NEA, Ecent is the average

centroiding accuracy (i.e., the fraction of a pixel to which the centroid

1

r














p








ns 1
ns
ed  q




ns  1
1

(26)

ns

1. Robustness to Outliers

The robustness of AIM is compared to that of QUEST (with zero

iterations) and q-Davenport in MATLAB simulations using 10,000
star-tracker exposures distributed randomly over the sky. In the
various scenarios, one of the image stars, the so-called outlier, was
given a position error with a variance, which was a factor Pe higher
than the position errors of the other stars. The distance between this

1679

DELABIE, SCHUTTER, AND VANDENBUSSCHE

Table 8

RMS attitude-angle errors of AIM, QUEST, and q-Davenport in various scenarios in which AIM uses more starsa

Scenario

 8, npix  1024, Ecent  0.5, ns  9
10

 8, npix  1024, Ecent  0.5, ns  9
11
 8, npix  1024, Ecent  0.5, ns  15
16
 8, npix  1024, Ecent  0.5, ns  15
17

RMS error ,
arcsecond
4.68
4.39
3.66
3.65

AIM
RMS error ,
arcsecond
4.59
4.37
3.61
3.52

RMS error ,
arcsecond
85.82
79.77
65.15
65.03

RMS error ,
arcsecond
4.94
4.94
3.68
3.68

QUESTb/q-Davenport
RMS error ,
RMS error ,
arcsecond
arcsecond
4.95
87.98
4.95
87.98
3.75
66.67
3.75
66.67

a
The
b

number of stars used by AIM is given in brackets.

One iteration.

Table 9

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

Scenario

 8, npix  1024,
Ecent  0.5, ns  9,
Pe  0
 8, npix  1024,
Ecent  0.5, ns  9,
Pe  25
 8, npix  1024,
Ecent  0.5, ns  9,
Pe  50
 8, npix  1024,
Ecent  0.5, ns  9,
Pe  100

AIM
QUEST
q-Davenport
RMS error , RMS error , RMS error , RMS error , RMS error , RMS error , RMS error , RMS error , RMS error ,
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
4.94
4.95
87.99
4.94
4.95
88.03
4.94
4.95
87.98
41.3

39.5

84

81

1542

137

87

165

159

3198

208

226

Table 10
Scenario

 8, npix  1024,
Ecent  0.5, ns  9,
r  0
 8, npix  1024,
Ecent  0.5, ns  9,
r  500
 8, npix  1024,
Ecent  0.5, ns  9,
r  1000
 8, npix  1024,
Ecent  0.5, ns  9,
r  3600
 8, npix  1024,
Ecent  0.5, ns  9,
r  10; 800

RMS attitude-angle errors of AIM, QUEST, and q-Davenport with outliers in the image

736

41.3

39.7

811

41.4

39.5

736

13,097

84

81

1541

19,382

165

159

3197

RMS attitude-angle errors of AIM, QUEST, and q-Davenport when qdat is inaccurate

AIM
QUEST
q-Davenport
RMS error , RMS error , RMS error , RMS error , RMS error , RMS error , RMS error , RMS error , RMS error ,
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
arcsecond
4.94
4.95
87.99
4.94
4.95
88.03
4.94
4.95
87.98
4.97

5.04

93.2

4.88

5.01

93.0

4.88

5.01

93.0

5.12

5.24

95.6

4.92

5.08

95.3

4.92

5.08

95.3

7.37

6.93

110.43

4.99

5.05

103.9

4.99

5.05

103.9

316.1

6.33

6.14

147.8

6.33

6.14

147.8

35.1

21.6

image star and its corresponding database star will therefore be

significantly higher. The results of these simulations are presented in
Table 9.
The results in Table 9 show that AIM is as robust to distortions as is
q-Davenport, which is one of the robust methods. QUEST performs
worse than AIM or q-Davenport in the presence of outliers. For the
roll error, this difference in error can amount to more than a degree.
This problem QUEST has with robustness is solved by performing two NewtonRaphson iterations. In simulations, QUEST then
performed similar to q-Davenport and AIM, at the cost of an increase
in the number of calculations.
2. Robustness to an Inaccurate Database Attitude qdat

In the tracking mode, the attitude estimation of the previous time

step can be used as a coarse estimation for the attitude in the current
attitude estimation. During large slew maneuvers, the spacecraft may
have significantly rotated away from the attitude of the previous time
step. Because of this, the coarse estimate of the attitude may be far off.

The effect of this was verified by validating the performance of AIM

when the coarse estimate was increasingly inaccurate, or in other
words, when the spacecraft had increasingly rotated away from
the previously estimated attitude. The results are given in Table 10,
in which r is the angle in arcseconds over which the coarse estimate,
qdat , is rotated away from the true attitude over each of the
three axes.
From Table 10, it is clear that for increasingly inaccurate coarse
estimates, qdat , the accuracy of AIM decreases. However, even for an
r of 10,800 arcseconds, AIM still converges to a reasonable
solution. Considering the real-life example of a star tracker with an
update frequency of 10 Hz, an r of 10,800 arcseconds would result
in the spacecraft rotating 30 deg s around each of its three axes. At
this point, it is likely that the star-identification algorithm would fail
because these algorithms generally also use a coarse estimate of the
attitude to efficiently identify the stars [20]. In this case, the attitudeestimation procedure would fail regardless of the used estimation
algorithm.

1680

DELABIE, SCHUTTER, AND VANDENBUSSCHE

It can be concluded that, in normal operating conditions, using the

previous attitude as qdat will not pose robustness issues, but might
reduce the accuracy when large maneuvers are performed.

Downloaded by RESEARCH CENTER IMARAT ( RCI) on March 27, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.61082

IV.

Conclusions

In this paper, a novel attitude-estimation algorithm, referred to as

attitude estimation using image matching (AIM), was proposed. At
the base of AIM lies an algorithm that optimally maps the stars of two
images on top of each other. Because this optimization problem can
be reduced to explicit equations from which the unknown variables
can be calculated, AIM is both extremely fast and robust. When AIM
is compared to the fast iterative and the robust attitude-estimation
algorithms that exist now, it is clear that AIM is faster than the fastest
algorithms, as robust as the most robust methods, and has a similar
accuracy as the existing algorithms. Furthermore, AIM allows in a lot
of cases to eliminate a very computationally expensive coordinate
conversion in the algorithms preceding the attitude-estimation
algorithm. This way, the computational cost of the attitude estimation
is reduced very significantly.
The proposed algorithm can allow to improve the performance of
the attitude estimation by using the reduction in computational cost to
acquire the attitude estimates at a higher rate, use more stars in the
attitude-estimation algorithm, or improve the centroiding algorithm.
It is also a valuable contribution to the expanding field of smallsatellite projects, in which platforms have a limited computational
capability, and the reduced computational complexity of AIM could
allow the implementation of star trackers on these platforms.

Acknowledgment
Tjorven Delabie is funded by the Agency for Innovation by
Science and Technology in Flanders, Belgium (Agentschap voor
Innovatie door Wetenschap en Technologie).

References
[1] Sidi, M. J., Spacecraft Dynamics and Control: A Practical Engineering
Approach, Cambridge Univ. Press, Cambridge, England, U.K., 1997,
pp. 353374.
[2] Percival, J. W., Nordsieck, K. H., and Jaehnig, K. P., The ST5000: A
High-Precision Star Tracker and Attitude Determination System,
Proceedings of the SPIE, Space Telescopes and Instrumentation 2008:
Optical, Infrared and Millimeter, Vol. 7010, Society of Photo-Optical
Instrumentation Engineers, Bellingham, WA, 2008, pp. 70104H
70104H-6.
[3] Liebe, C. C., Accuracy Performance of Star TrackersA Tutorial,
IEEE Transactions on Aerospace and Electronic Systems, Vol. 38,
No. 2, 2002, pp. 587599.
doi:10.1109/TAES.2002.1008988
[4] Spratling, B. B. IV., and Mortari, D., A Survey on Star Identification
Algorithms, Algorithms, Vol. 2, No. 1, 2009, pp. 93107.
doi:10.3390/a2010093
[5] Delabie, T., Durt, T., and Vandersteen, J., A Highly Robust Lost in
Space Algorithm Based on the Shortest Distance Transform,
Proceedings of the AIAA Guidance, Navigation, and Control
Conference, AIAA Paper 2011-6435, Portland, Oregon, Aug. 2011.
[6] Markley, F. L., and Mortari, D., How to Estimate Attitude from Vector
Observations, Proceedings of the AAS/AIAA Astrodynamics Specialist
Conference, AAS Paper 1999-427, Girdwood, Alaska, Aug. 1999.

[7] Lerner, G. M., Three-Axis Attitude Determination, Spacecraft

Attitude Determination and Control, edited by Wertz, J. R., Springer
Verlag, Berlin/New York, NY, 1978, pp. 420428.
[8] Landis, M. F., Attitude Determination Using Vector Observations and
the Singular Value Decomposition, Journal of the Astronautical
Sciences, Vol. 36, No. 3, JulySept. 1988, pp. 245258.
[9] Horn, R. A., and Johnson, C. R., Topics in Matrix Analysis, Cambridge
Univ. Press, Cambridge, England, U.K., 1991, p. 134.
[10] Golub, G. H., and Loan, C. F. V., Matrix Computations, 3rd ed., Johns
Hopkins Univ. Press, Baltimore, MD, 1983, pp. 7274.
[11] Cheng, Y., and Shuster, M. D., The Speed of Attitude Estimation,
Proceedings: Advances in the Astronautical Sciences AAS/AIAA,
Vol. 127, American Astronautical Society, Springfield, VA, 2007,
pp. 101116.
[12] Shuster, M. D., Approximate Algorithms for Fast Optimal Attitude
Computation, AIAA Guidance and Control Conference, AIAA, New
York, NY, 1978, pp. 8895.
[13] Shuster, M. D., and Oh, S. D., Three-Axis Attitude Determination from
Vector Observations, Journal of Guidance and Control, Vol. 4, No. 1,
1981, pp. 7077.
doi:10.2514/3.19717
[14] Markley, F. L., Attitude Determination Using Vector Observations: A
Fast Optimal Matrix Algorithm, Journal of the Astronautical Sciences,
Vol. 36, No. 3, AprilJune 1993, pp. 261280.
[15] Mortari, D., ESOQ: A Closed-Form Solution to the Wahba Problem,
Journal of the Astronautical Sciences, Vol. 45, No. 2, AprilJune 1997,
pp. 195204.
[16] Mortari, D., ESOQ2: Single-Point Algorithm for Fast Optimal Attitude
Determination, Advances in the Astronautical Sciences, Vol. 97, No. 2,
1997, pp. 803816.
[17] Feuchtgruber, H., Herschel STR-A CCD Sub-Pixel Structure, Max
Planck Inst. for Extraterrestrial Physics, PICC-ME-TN-041, Garching,
Germany, March 2012.
[18] Quine, B. M., Tarasyuk, V., Mebrahtu, H., and Hornsey, R.,
Determining Star-Image Location: A New Sub-Pixel Interpolation
Technique to Process Image Centroids, Computer Physics
Communications, Vol. 177, No. 9, 2007, pp. 700706.
doi:10.1016/j.cpc.2007.06.007
[19] Rufino, G., and Accardo, D., Enhancement of the Centroiding
Algorithm for Star Tracker Measure Refinement, Acta Astronautica,
Vol. 53, No. 2, 2003, pp. 135147.
doi:10.1016/S0094-5765(02)00199-6
[20] Samaan, M. A., Mortari, D., and Junkins, J. L., Recursive Mode Star
Identification Algorithms, IEEE Transactions on Aerospace and
Electronic Systems, Vol. 41, No. 4, 2005, pp. 12461254.
doi:10.1109/TAES.2005.1561885
[21] Cheng, Y., and Shuster, M. D., Robustness and Accuracy of the
QUEST Algorithm, AAS/AIAA 17th Space Flight Mechanics
Meeting, American Astronautical Society, Springfield, VA, 2007,
pp. 4661.
[22] Mortari, D., Second Estimator of the Optimal Quaternion, Journal of
Guidance, Control, and Dynamics, Vol. 23, No. 5, 2000, pp. 885887.
doi:10.2514/2.4618
[23] Feeman, T. G., Portraits of the Earth: A Mathematician Looks at Maps,
American Mathematical Society, Providence, RI, 2002, p. 13.
[24] Ball Aerospace & Technologies Corp., CT-602 Star Tracker, http://
www.ballaerospace.com/file/media/D0540_CT-602.pdf
[retrieved
6 Nov. 2012].
[25] Shuster, M. D., The Quest for Better Attitudes, Journal of
Astronautical Sciences, Vol. 54, Nos. 34, JulyDec. 2006,
pp. 657683.
doi:10.1007/BF03256511