Illumination effects on satellite articulation
characterization from a trajectory matrix using
optimization
David Curtis
Air Force Institute of Technology
2950 Hobson Way
Wright-Patterson AFB, OH 45433
937-255-6565 x6117
david.curtis@afit.edu
Abstract—Autonomous on-orbit satellite servicing and inspec-
tion benefits from an inspector satellite that can autonomously
gain as much information as possible about the primary satellite.
This includes performance of articulated objects such as solar
arrays, antennas, and sensors. This paper presents a method
of characterizing the articulation of a satellite using resolved
monocular imagery. A simulated point cloud representing a
nominal satellite with articulating solar panels and a complex
articulating appendage is developed and projected to the image
coordinates that would be seen from an inspector following
a given inspection route. A method is developed to analyze
the resulting trajectory matrix. The developed method takes
advantage of the fact that the route of the inspector satellite is
known to enable the reprojection error to be used as an objective
function for optimization of a model that best describes the
feature points seen. Once the model is calculated, it can be
compared to the known truth. Simulating the input data allows
different inspection routes to easily be evaluated. Particularly,
the effects of fly-by and natural motion circumnavigation inspec-
tion routes with sub-optimal illumination are investigated using
a set of newly proposed evaluation metrics.
TABLE OF C ONTENTS
1. I NTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2. M ETHODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3. R ESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4. C ONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
R EFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
B IOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1. I NTRODUCTION
Due to the nature of the space environment, direct human
interaction with on-orbit satellites is incredibly dangerous
and costly. Therefore, with the exception of a few manned
missions, interaction with satellites is traditionally limited
to windows of radio communications and observations from
telescopes. These methods are inadequate for inspection
and monitoring of a satellite to determine performance. To
perform these monitoring tasks, an inspector satellite in close
proximity to the primary satellite could use sensors, such
as a camera, to characterize the primary satellite and verify
the performance of articulated objects such as sensors, solar
arrays, robotic arms, and communications antennas. Imagery
contains a lot of information, however it is also difficult to
transmit to the ground due to the large size of video files. An
U.S. Government work not protected by U.S. copyright
1
Richard Cobb
Air Force Institute of Technology
2950 Hobson Way
Wright-Patterson AFB, OH 45433
937-255-6565 x4559
richard.cobb@afit.edu
articulated model could provide the necessary information to
users on the ground with a fraction of the data transmission
requirement of the raw imagery. This motivates the creation
of algorithms that can characterize the performance of the
target satellite autonomously using computer vision. Devel-
oping autonomous methods is also important for autonomous
repair missions [1] or deep space missions that have little or
no ability to communicate with human operators on Earth.
Research in the area of computer vision for proximity op-
erations is very diverse, ranging from stereo vision used
to estimate target satellite moment of inertia (MOI)[2] to
demonstration of fully autonomous rendezvous and capture
[1]. Some of these algorithms rely upon markers on the
primary satellite that assist the computer vision algorithm in
determining pose [3],[4]. Others rely on prior knowledge of
the primary satellite’s configuration [5],[6]. Some methods
rely on stereo vision systems,[2],[7] some on monocular
vision systems,[8],[9],[6] and some use laser illumination
of reflective markers [1]. Regardless of the method, they
all use computer vision to identify features in images and
then matches those features from frame to frame (or in
corresponding frames in the case of stereo systems). The
relative position of the features is then estimated using some
type of estimation filter such as an extended Kalman filter
(EKF) or a particle filter. The attitude of the primary satellite
is then estimated by using feature points to define a pri-
mary reference frame,[8],[10] or by using a known model
of the primary satellite [5]. Alternatively, if the orientation
of feature points is known, the attitude can be calculated
directly and used as the measurement in the EKF [4],[8],[10].
Feature points can also be used to create a 3D model of the
satellite [11]. Apart from their own previous work[12],[13],
the authors have not discovered previous research attempting
to use computer vision to determine if the primary satellite
is in-fact a rigid body, or if it is undergoing some type of
articulation.
While articulation has not been investigated for space ap-
plications specifically, there has been extensive work on
identifying and characterizing articulation in the computer
vision field. The seminal works of Yan and Pollefeys [14]
and Trasadern and Reid [15] extend the concepts of structure
from motion through factorization of a trajectory matrix
[16] from rigid bodies only to multiple linked rigid bodies.
Numerous researchers have further extended these concepts
to characterize articulated motion.
Yucer et al.[17] developed a method of reconstructing an
object with multiple articulated motions using optimization
techniques. First, a ‘ray-space optimization’ method is used
to convert each 2D image trajectory into a 3D trajectory. This
method requires knowledge of the camera motion a priori,
however this requirement is not necessarily restraining in
an application in which the relative trajectory between the
inspector and the primary satellite is known. The 3D trajecto-
ries of each point are then used to segment the motion into N
rigid bodies. Once segmented, an optimization routine is used
to find a rotation matrix, shape, and translation for each rigid
body that minimizes the error in reprojecting the points to
image coordinates. Next, the kinematic chain is estimated and
another optimization is done with the articulation constraints
enforced.
Paladini et al.[18] present a method in which the motion
matrix (identifying the motion of the camera) and shape ma-
trix (identifying the shape of the object) are solved for in an
alternating fashion with least squares. First, an initial solution
to the motion matrix is found using the method outlined by
Tresadern and Reid[15]. This is then refined by optimizing
a cost function that includes the metric constraints and the
articulation constraints. They use the method in Marques
and Costeira[19] to enable their method to handle missing
data. Results suggest the algorithm is capable of handling
trajectory matrices in which over 60% of the columns contain
missing data without detriment to accuracy of the recovered
shape.
Zhang and Hung[20] approach the problem of recovering
articulated structure from motion as an ellipsoid fitting prob-
lem. They begin by noting that each subset of the trajectory
matrix representing the points on a rigid body can be repre-
sented as a 3D ellipsoid. Since points corresponding to the
same motion will lie on the same ellipsoid, this can be used
to segment the points. They use the error in fitting a point
to an ellipsoid as a metric with which to segment the feature
points into separate motions.
Much of the literature on recovery of articulated structure
from motion is focused on capturing the motion of humans.
Some of this research can be applied to articulated motion
of any kind. For instance, Fayad et al.[21] present a method
of automatically recovering the 3D shape and structure of an
articulated body. Their method uses an optimization routine
to assign feature points to the motion that minimizes the
total re-projection error of the solution. The method allows
points to belong to more than one motion. This overlap of
points on multiple motions defines the joint between the two
objects and allows the kinematic chain to be built within
the optimization framework. Russell et al.[22] expand on
the method of Fayad et al. to include the capability to
segment motion into independent objects and further segment
independent objects into multiple parts based on dependent
motions such as articulation.
Understanding articulated motion is also important to enable
robots to learn about how objects move in order to allow
the robots to manipulate the objects in the future. With this
motivation, Pillai et al.[23] developed a training process by
which the parameters of everyday articulate objects, such as
doors and drawers are determined from video of a person
using those objects.
The method in this work builds on existing work in the
computer vision field and applies it to the inspection of
a satellite in space. An algorithm is developed to build
an articulated model from simulated imagey taken from an
inspector satellite on a particular inspection route. Evaluation
metrics to quantify the quality of the model are developed and
2
the effect of inspection routes with sub-optimal illumination
on these metrics is investigated.
2. M ETHODS
Method overview
The method developed herein extend [12] and [13] in attempt-
ing to detect and characterize the articulation and shape of
a satellite from a simulated trajectory matrix that would be
seen from a monocular camera on an inspector satellite. The
method begins with an inspection route and the time history
of a point cloud representing a nominal satellite, shown in
Figure 1a, with articulated motion. Given the inspection
route and the point cloud, a trajectory matrix is created
using a pinhole camera model, taking into account occlusion
and shadowing. The trajectory matrix is a 2F × P matrix
where F is the number of image frames and P is the number
of feature points. The trajectory matrix and the inspector
satellite trajectory are then used to characterize the satellite
and its articulated motion. Figure 1a provides an overview of
the current method. This method is a combination of multiple
static optimization routines similar to the process outlined by
Yucer et al.[17].
Simulation set-up
For this work a nominal satellite was developed for testing.
This satellite consists of a main body, two articulating panels,
and a 6-component articulating arm. Each of the joints is a
revolute joint meaning that it articulates about a constant axis.
Figure 1b shows the model satellite which is approximately
2 units deep (x-direction), 10 units wide (y-direction), and 10
units tall (z-direction) with 8 articulation joints with labels
and articulation axes shown. Points were randomly spread
over each face to build the point cloud shown in Figure 1c.
In this work the two panels and only 3 of the joints in the
articulating arm were activated. The effect of different num-
bers of active appendage joints is assessed in previous work
[13] where all 8 joints are active with 6 in a single kinematic
chain. Adding links to the chain increases the complexity,
however with 6 joints in the appendage this method was
demonstrated to be capable of building an accurate model of
the satellite [13]. For this work, the effects of sub-optimal
lighting conditions are investigated. To minimize variation,
the motion of the satellite is the same in all cases. Over the
length of the simulation, joints 1 and 2 articulate linearly from
-1.26 radians to 1.26 radians while joints 3, 5, and 6 articulate
linearly from -0.94 radians to 0.94 radians.
Two types of inspection routes are investigated in this work: a
2 × 1 elliptical natural motion circumnavigation (NMC),[24]
and a fly-by in which the inspector route consists of a straight
line path in the relative frame. In each case, the route can
be phased for the ‘best’ illumination conditions or it can
be offset by some illumination offset angle (θ). The ‘best’
illumination conditions are assumed to be those in which
the primary satellite is most illuminated as viewed by the
inspector satellite, however, due to the nature of the space
environment, even these illumination conditions are sub-
optimal as compared to a terrestrial application in which an
object can be lit from all sides. Figure 2 shows the meaning
of θ for fly-by routes and NMCs. Fly-by routes consist of
the inspector satellite moving past the primary satellite on a
linear path in the relative frame that is parallel with the in-
track direction. The route is simulated to occur over 1 hour
(assuming the primary satellite is in a 24 hour circular orbit).
In this case, θ is related to the location in the orbit where
 Figure 1. a) Outline of the proposed method. b) Diagram showing component and articulation axis naming. c)
Example point cloud representation of the satellite.

Figure 2. Inspection route diagrams illustrating the illumination offset parameter (θ). a) Fly-by inspection route. b)
NMC inspection route.

3
the fly-by occurs. For an NMC the zero illumination offset
condition (θ = 0) represents a route in which the phasing
of the NMC is such that the inspector is always between the
primary satellite and the Sun. In this case, θ represents sub-
optimal phasing of the NMC with the primary satellite orbit.
Given the inspector route, the trajectory matrix is created by
translating visible points from the point cloud to the camera
image plane at each position in the inspection route where an
image is taken (100 images, an arbitrarily selected number,
are taken for each route simulated in this work). Points are
visible if they are illuminated by the Sun vector, they are on
a face that is pointed toward the camera, and they are not
occluded by another part of the satellite. An error function is
used to gradually fade points in and out of view as the angle
between the Sun or camera and the normal vector translates
through 90◦ . Visible points are translated to the camera
image plane using a pinhole camera model. For simulation
purposes, it is assumed that throughout the trajectory the cam-
era is at a distance from the primary satellite that produces
resolved imagery of appropriate resolution for feature point
extraction and tracking. In all cases, the primary satellite is
simulated to be in geostationary orbit at equinox. The effect
of Earth shadowing is not simulated. Note that while the
nominal satellite contains panels that resemble solar arrays,
no effort was taken to ensure their motion is consistent with
a solar array. Both sides of the array may be illuminated at
different portions of the route.
Ray space optimization
The ray space optimization technique outlined in Yucer et
al.[17] provides an excellent method of estimating 3D shape
from a 2D image coordinates when the camera motion is
known. The method parametrizes the 3D location of each
point (p) in a given frame (i) using the camera center (Ci ),
the direction vector from the camera center to the point (Dip )
and the distance along the ray from the camera center to the
point (µpi ) using the following equation:
Sip = Ci + µpi Dip
where Sip is the 3D location of point p in frame i. There are
many valid 3D paths that can result in the same 2D image
coordinates, therefore, an added assumption is made that the
points move smoothly from frame to frame. This leads to a
cost function that is only a function of the depth µpi at each
frame.
F −1
X as LSA [26] use a metric based on the angles between the
µp ) =
Ers (~ ωip k(Ci + µpi Dip ) − (Ci+1 + µpi+1 Di+1
p
)k2
i=1
The term ωip is a weighting term that is based on the 2D
distance of the point in frame i + 1 from the epipolar line
corresponding to the point location in frame i. This weighting
method effectively keeps the 3D location of the point when
it is moving near its location when it is static [17]. The
distance from the epipolar line can also be used to determine
if the point is stationary. With the assumption that the
camera motion is known with respect to the main body of
the satellite, points that are stationary are points on the main
body (or components that are rigidly attached to the main
body throughout the inspection). Points that have an average
distance from the epipolar line below a threshold (γs ) are
considered stationary and are segmented to the main body
[13].
Knowing that a point is stationary also allows the calculation
of its position to be simplified to a triangulation problem that
can be solved using linear least squares. For a stationary
point S1p = S2p = ... = SFp , therefore equation (1) becomes
S p = Ci + µpi Dip . When written for each of the F frames
there are 3F equations. Since S is no longer different for each
frame there are only F+3 unknowns. This set of equations
(see equations (3)) can be written in the form of Ax = b
(equation (4)) and solved with linear least squares. Note that
in this framework, points are independent, and therefore the
optimization is performed on each point separately using only
the frames in which the point is visible.
S p − µp1 D1p = C1
S p − µp2 D2p = C2
.. .
. = ..
S − µF DFp
p p
= CF (3)
 Sp
 
I −D1p 0 ··· 0
  
p C1
 µ1 
I 0 −D2p ··· 0   µp   C2 
  .  =  ...  (4)
. .. .. .. ..   2  
.
. . . . .  .. 
I 0 0 ··· −DFp CF
µpF
Note the appropriate selection of γs will depend on the noise
in the inspection trajectory and feature point locations. In
the presence of noise, a simple threshold may not be capable
of segmenting points that are stationary, in which case this
method should not be used to segment points on the main
body.
Segment points
The nature of a circumnavigation inspection route yields a
trajectory matrix with a significant amount of missing data.
(1) This makes motion segmentation challenging. Figure 3
shows an example trajectory matrix where empty elements
are shown in black and elements containing data are shown in
white. Spectral clustering is a popular method of segmenting
data [25]. While there are multiple methods of spectral clus-
tering, they all operate on some type of similarity matrix. For
P points, a similarity matrix is P × P in dimension with each
element representing how similar the point corresponding to
its column is to the point corresponding to its row. The
type of similarity metric used varies widely. Some, such
(2) subspaces of nearest neighbors, while others use a metric
based on the range and velocity between points [20]. Once
the similarity metric is chosen, there are numerous ways of
segmenting the data. For this work we investigated both
spectral clustering using k-means [27] as well as recursive
2-way spectral clustering [28],[17]. Both methods involve
solving for the eigenvectors of the Laplacian matrix. The
Laplacian matrix is the similarity matrix minus a diagonal
matrix of the row sums of the similarity matrix. The reader
is directed to [29] for additional information on spectral
clustering.
Particularly in the case of an NMC, there are many points
on the same rigid component that do not share any common
frames. Therefore it is difficult to segment them onto the
same rigid component. When added to the fact that the
number of rigid components is not assumed to be known a
priori, over segmentation is necessary to ensure each segment
4

Figure 3. Example trajectory matrix mask for 1
complete NMC.
consists of points primarily from the same component. The k-
means clustering technique requires the number of segments
(k) to be given while the 2-way spectral clustering technique
used by Yucer et al. [17] only requires an estimate for the
number of segments. In testing for this application both
methods produced similar results with approximately 90% of
points segmented correctly. For this work, a similarity matrix
based on the range and 2D velocity [21] between image points
was used with spectral clustering using k-means [27],[30].
The value of k was judiciously set to 10 for fly-by routes and
22 for NMC routes.
Rigid body optimization
Once the points are segmented, the next step is to find how
all the rigid bodies are moving. Similar to Yucer et al.[17] we
sought the translation, rotation, and shape that minimized the
reprojection error and a smoothness constraint.
F from
X
min kWin − Pi (Rin Ωn + T̃n 2 motion matrix containing the first two rows of the rotation
i )k +
R,T,Ω
i=1
| {z } time step, Ω is the 3 × P shape matrix in the body frame, and
Reprojection Error
XF
λrb

acos(.5(trace(Rin (Ri−1 n
)T ) − 1)) + kTin − Ti−1
n
k
 scale factor (α) and the rotation matrix between the camera
i=2
| {z } route, the scaled orthographic projection equation can be used
Smoothness Constraint
(5) translation as follows.
This optimization is performed for each segment. W n are the
columns in the trajectory matrix for the n-th segment. P is
the camera matrix that projects points in the world frame to
the camera image plane. Ωn is the 3D location of the points
in the body frame. Rn is the rotation matrix that rotates body
frame for segment n to the world frame. T̃ n is the translation
from the origin of the world frame to the center of the points
in Ωn while T n is the translation from the origin of the world
frame to the origin of the body frame. They are related by:
T̃in = Tin − Rin Ω̄n (6)
Using T̃ n in the reprojection error portion of the cost function
helps to decouple the rotation from translation [13].
There are multiple rigid body shape and motion combinations
that can minimize the reprojection error. Since the rigid
bodies are most likely to follow smooth paths, the smooth-
ness constraint is added to encourage solutions that follow
a smooth rotational and translational paths. λrb is used to
weight the smoothness constraint. The appropriate value of
λrb will be dependent on the scaling of the trajectory matrix.
Setting it too low will not enforce smoothness and may lead
to a jittery path while setting it too high will minimize the
rigid body motion at the expense of reprojection error. To
allow the method to be robust to scaling, the value of λrb
is choosen adaptively. Before optimization begins, and after
every 100 iterations during optimization, the value of λrb is
adjusted so that the smoothness cost is no more than 25% and
no less than 5% of the reprojection error.
The rotation matrix (Rn ) is parameterized using an Euler axis
(~a) and an Euler angle (φ) and represents the rotation matrix
from the segment’s body frame (b) and the world frame (w).
The reference frames used in this work are shown in Figure
4. The Euler axis is not constrained to be unit length, instead
it is normalized before being used to calculate Rn .
To begin optimization, an initial guess must be supplied for
all the optimization variables. In many cases, the ray space
optimization results provide a good method of initializing
the shape (Ωn ) and the translation (T n ). However, when
the camera rotation is less than the articulated motion some
issues were discovered [13]. Since the ray space optimization
routine attempts to find the shortest path that models the
observed motion, when the camera motion is less than the
articulated motion the solution tends toward a path in which
the point follows a trajectory similar to the camera motion
rather than the articulated motion. If this solution is used
to initialize the rigid body optimization, the error propagates
through the process resulting in an inaccurate model.
To resolve this issue, another method was developed to
initialize the shape and rotation parameters using the scaled
orthographic projection model [31]. Using a scaled ortho-
graphic camera model, the trajectory matrix can be written as
W = αRΩ + T where α is a scaling based on the distance
the camera to satellite (α = |rWf C | ), R is the 2F × 3
matrix between the camera frame and the body frame at each
T2D is the translation from the world frame origin to the body
frame origin projected into the 2D image plane. Since the
frame and the world frame are known from the inspection
to solve for the resulting shape from any body rotation and
W̃ = RΩ = 1
αW − T~2D (7)
Ω=R W †
(8)
Wcalc = α(RS + T~2D ) (9)
E = kW − Wcalc k (10)
This set of equations is used to initialize the rotation parame-
~ The motion matrix R is determined for 5,000
ters (~a and φ).
seeds consisting of a randomly selected constant Euler axis ~a
and a linear set of Euler angles φ ~ of a random slope. The
translation (T2D ) is taken from the ray space optimization
results. For each R the error (E) between Wcalc and W
is calculated. The seed with the lowest error is used to
~ the resulting Ω is used to initialize shape.
initialize ~a and φ;
The translation is initialized to be the center of the available
3D estimates at each frame (Tin = S̄in ) of the ray space
optimization results. This is a highly non-linear optimization
problem with numerous local minima. While this method of
initialization worked well for the cases run, alternatives, such
as using a heuristic optimization solver for initialization or a
global optimization algorithm may produce superior results.
Outlier rejection is conducted every 100 iterations during
rigid body optimization. Points that have a reprojection error
higher than some multiplier (γf ) times the average per point
reprojection error and points that have an average range to all
other points higher than γr standard deviations from the mean
range between points are rejected as outliers.
5
 Figure 4. Reference frames used in this work. The ‘world frame’ (w) corresponds to the CW frame and moves with
the primary satellite in its orbit. Each component has a ‘body frame’ (b) attached to it which is arbitrarily assigned.
The ‘camera frame’ (cam) is attached to the inspector camera with the positive z axis along the imaging axis and the
positive x axis aligned with the camera’s direction of travel in the world frame.

Combine segments
Since the data has been over-segmented, by choosing a larger
than expected number of segments, the next step is to com-
bine any segments (merge) that are part of the same rigid
body. To do this, four adjacency matrices that compare the
location and rotation of each segment to each other segment
are created. If two segments are in fact the same rigid
body, the rotations should be the same, therefore the rotation
b ,b
between their body frames (Ri n1 n2 ) should be constant for
all common image frames.
b ,bn2
Ri n1 = (Ri1 )T Ri2 (11)
b ,b
The average variance of the elements of Ri n1 n2 over all
common frames (CF) is used to define the (n1 , n2 ) position
of an N × N adjacency matrix where N is the number of
segments.
For each segment, the closest two segments from each of
the four adjacency matrices are selected for testing to deter-
mine if they should be merged. Rigid body optimization is
performed on the combined segments as outlined in section
2. If the resulting function value is below 200% (γc ) of
the average function values of the segments independently,
the two segments are merged and the adjacency matrices are
recalculated. This is continued until all segments have been
checked without triggering a merge [13].
Build kinematic chain
The next step is to determine the kinematic chain that best
describes which components are linked to each other. This
is done by evaluating the average range adjacency matrix
mentioned in the previous section. The kinematic chain is the
minimum spanning tree of a graph (G) with each component
as a node and edges defined by the average distance between
components in common frames (CF ).

It is assumed that segments that are part of the same rigid

body are close to one another, therefore an adjacency matrix CF
1 X j
is created that compares the range between translations in G(j, k) = kT − Tik k (12)
common frames. Segments that are the same rigid body CF i=1 i
will also remain at the same distance from one another
over common frames. A third adjacency matrix is created
comparing the distance between the segments’ average loca- To ensure the root of the kinematic chain corresponds to the
tion, allowing comparison of segments that do not appear in main body, the segments are first ordered by their average
common frames. Finally, a fourth adjacency matrix is created translation.
to compare the variance of the range between the translations.
These four adjacency matrices are used to determine which The kinematic chain can be expressed as a table of dimension
pairs of segments to test for merging. M × 2 where M is the number of joints (M = N − 1)
6
containing parent-child relationships. J p and J c are the location of the joint in the parent’s body
frame and the location of the joint in the child’s body frame
respectively. Parameterizing using the joint locations forces
 
parent1 child1
 parent2 child2  linkage between the components and reduces the number of
H= .. ..  (13) optimization variables [13]. Notice the rotation and transla-
 . .  tion of the main body is set to identity and zero respectively.
parentM childM This method assumes the main body is stationary in the rela-
tive frame. This assumption is in place for convenience and
This table is re-ordered as required to ensure that each may not be valid in many realistic situations. To remove this
segment appears as a child before it appears as a parent. assumption, additional optimization variables could be added
Components that do not share common frames with their to encompass the rotation and translation of the main body
parent component are eliminated. Note that this method is with respect to the world frame. This adjustment would also
not robust to cycles in the kinematic chain [13]. all allow the method to accommodate errors in the inspection
route which would show up as part of the calculated main
Articulation parameter optimization body motion in the world frame.
Once the kinematic chain is known, the articulation param- x ,x
eters can be calculated. For this application the joints are The rotation between linked frames Ri k−1 k is parameter-
assumed to be revolute. Yucer et al. [17] use a similar method ized by two Euler axes, one for the parent (xk−1 ) and one
to the one presented here, however they assume universal for the child (xk ), and an Euler angle φi . Since body frames
joints and do not include the joint locations as optimization are chosen arbitrarily during rigid body optimization of each
variables. The articulation parameters are the articulation component, there is no reason the articulation axis would be
the same between the body frame of the parent and the body
axes between linked parts, the angle of articulation, and the frame of the child. This means the frames must be aligned
joint locations. For N components, there are N-1 joints. For before the articulation angle is applied. An intermediate
each joint, p identifies the parent component and c identifies frame (aa) is used to align the frames and then a rotation
the child component. The w superscript represents the world matrix is formed using the articulation axis of the parent (âxp k )
frame defined by the center of the main body of the primary and the articulation angle (φi ).
satellite. The first term of equation (14) is the same as the
reprojection error in equation (5), however the rotation and x ,xk x ,aa
Ri k−1 = Ri k−1 Raa,xk
translations are now dependent on the higher components
in the kinematic chain. The last term is also similar to the Raa,xk = cos(ψa )I + (1 − cos(ψa ))âa (âa )T − sin(ψa )â×
a
smoothness constraint in equation (5), however due to the ~aa = âxk−1 × âxk
articulation constraints the articulation angle is the only tem-
porally dependent variable. As with rigid body optimization, ~aa
âa =
the weights λjp and λsc are chosen adaptively to keep the k~aa k
joint penalty below the reprojection error and the smoothness ψa = acos(âxk−1 · âxk ) (17)
penalty below between 25% and 5% of the reprojection error.
x ,aa
N
X −1  X
CF Ri k−1 = cos(φi )I + (1 − cos(φi ))âxk−1 (âxk−1 )T
kWi − Pi (Riw,c Ωc + T w,c )k2
 
min
R,J,Ω −sin(φi )â×
xk−1 (18)
j=1 i=1
| {z }
Reprojection Error
12
The joint penalty in equation (14) encourages the joints to
X
+ λjp ln(1 + edk η )
remain close to the point cloud. For each component, a
k=1
| {z } bounding box (BB) containing the points in Ωn is created
Joint Penalty and scaled by a factor of 1.5 (γBB ). A soft plus function uses
CF  the difference between joint location and the bounding box
+ λsc
X
(1 − cos(φi − φi−1 ))2
(14) (d) and a multiplier (η) to penalize the cost function if the
joints (Jp and Jc ) are outside of the bounding box [13].
i=2
| {z }
Smoothness Constraint To initialize the optimization for each joint, the information
from the rigid body optimization is used. Since each joint
is revolute it consists of rotation about a constant axis. A
The rotation matrices and translations are calculated based rotation matrix relating the body frame of each component to
on the hierarchy of the kinematic chain and the articulation the world frame is available from rigid body optimization.
constraints. Define X = [x1 , x2 , · · · , xl ] as a single chain To find the articulation parameters (âp , âc , φi ) that most
starting at the root and ending with a component that has no closely match the calculated rotation between components
children where each element of X identifies a component and (Rip,c ) over common frames the following cost function is
l is the length of the chain. The rotations and translations can minimized:
be written as follows for the joint between xk and xk−1 .
 CF
I , k=1
X
w,x
Ri k = w,x x ,x (15) min [acos(0.5trace(Rip,c (Mip,c )T ) − 1)
Ri k−1 Ri k−1 k
, otherwise M,φ
 i=1
w,x 0 , k=1 CF −1
Ti k = w,x w,x w,x (16)
Ri k−1 J p + Ti k−1 − Ri k J c , o.w.
X
+ λa (φi+1 − 2φi + φi+1 )2 ] (19)
i=2

7
The first term enforces the single axis of rotation constraint matched to the most appropriate truth joint the articulation
while the second term encourages smooth motion by only parameters can be compared. However, the articulation
penalizing articulation that is not at a constant rate. Rip,c is parameters of the calculated model are in a body frame that
calculated from Riw,p and Riw,c . Mip,c is calculated from the is unrelated to the body frame used to create the truth model.
To alleviate this, the articulation axis and joints are compared
optimization variables (âp , âc , φi ) by first aligning the axes in the world frame (w) as follows (equations (22)), where N
and then rotating about it as outlined in equation set (17) [13]. is the number of joints, and F is the number of frames with
data. Since the joints are revolute, any location on the axis
As expressed in Yucer et al.[17], the joint location translated is acceptable, therefore for the joints the error is measured as
into the world frame will coincide for two linked parts. the distance from the joint to the line represented by the true
joint location and articulation axis in the world frame.
Rip J p + Tip = Ric J c + Tic , ∀ i ∈ CF
" #
N Fj
For components linked by a universal joint there is a single Eâ = 1
P 1
P
acos(|(âjtrue,w )T âjcalc,w |)
Jp and Jc that best satisfy this constraint. For a revolute N
j=1
Fj
i=1
joint, any point on the axis will satisfy this constraint. To " #
N Fj
mathematically favor the joint locations that are closest to 1 1
kâjtrue,w j j
P P
the component center, the norm of the joint locations is also EJ = N Fj
× (Jcalc,w − Jtrue,w )k (22)
j=1 i=1
minimized giving the following cost function [13].
CF The articulation angle is evaluated using the error in the range
X of angles covered, Eφrange , and the rate of change of the
min [k(Rip J p + Tip ) − (Ric J c + Tic )k2 angle Eφ̇ .
Jp ,Jc
i=1
+ λb (kJp k + kJc k)] (20) N
1
|∆jtrue − ∆jcalc |
P
Eφrange = N
(23)
j=1
As the kinematic chain grows, the complexity of the problem ~ j ) − min(φ ~j )
increases, as does the number of variables. The number of ∆j = max(φ (24)
" #
optimization variables in equation (14) is (4 × 3 × (N − 1)) + 1
PN
1
F
Pj i i
(CF ×(N −1))+(3×P ). Each articulation parameter (âp , âc , Eφ̇ = N Fj
|φ̇true − φ̇calc | (25)
j=1 i=2
φ, Jp , Jc ) for each joint has an influence on the components
and joints below it in the kinematic chain. This means the φ̇i = φi−1 − φi (26)
optimization variables are highly interdependent resulting in
a complex non-linear problem. However, the articulation
parameters do not have an influence on the components above
the joint in the kinematic chain. Using this relationship, a In addition to evaluating the results directly, it may be inter-
method termed Incremental Joint Addition (IJA) was devel- esting to understand how the satellite is capable of moving,
oped in [13]. IJA determines the articulation parameters in or more generally, what volume around the satellite are the
order, according to the kinematic chain by optimizing at every appendages capable of reaching. This can be termed the
step. Equation (14) is optimized N-1 times, each time another satellite’s workspace. To calculate the satellite’s workspace,
joint is added to the chain. This decreases the complexity the space around the satellite is discretized into a grid of
since each time a set of parameters is introduced into the m × m × m = M points. Each joint is moved through
optimization they only influence the reprojection error for one its range of articulation angles at Y increments. A convex
component. It also provides a method of further combining hull is created encompassing each shape in the world frame
components that are likely to be the same rigid body. If the in the current increment and the previous increment. The
change in Euler angle for a joint is low, the two components grid points in M that fall within the convex hull [32] are
likely can be combined into a single component. Results annotated as covered. The collection of points in M that
using the IJA method as well as optimizing over all joint are covered represents the workspace of the satellite. Both
parameters together (AJP) are presented in previous work the calculated workspace and the true workspace are found
[13]. For the work in this paper the IJA method is used in this manner. They are compared using three percentages.
exclusively. Pwc is the percentage of the truth workspace covered by the
calculated model, Pwoc is the percentage of additional space
Evaluate results covered by the calculated model, and Pgc is the percentage
Multiple metrics are used to evaluate the results. Because the of the grid points where the true workspace matches the
trajectory matrix is simulated from a known point cloud, the calculated workspace. These values are calculated as follows
position of each point at each time step is known. This allows where B is an M × 1 vector with Bi = 1 when grid point i
calculation of the normalized reconstruction error [21] (E3D ) is covered and Bi = 0 when grid point i is not covered. 
which is direct comparison of the calculated point cloud to denotes element-wise multiplication.
the known point cloud in the world frame using equation calc
B truth
P
BP
(21) where Struth
i are the true world frame point locations Pwc = B truth
(27)
of all points calculated in image frame i and Scalci are the P calc
B P |B truth
−1|
calculated world frame point locations in image frame i. Pwoc = B truth
(28)
P calc truth
F 1−|B −B |
P
kStruth
i −Scalc
i k Pgc = M (29)
i=1
E3D = F
(21)
kStruth Note that the workspace calculation does not consider phys-
P
i k
i=1
ical limits to motion imposed by other components. In other
The calculated kinematic chain may contain more or fewer words, two components may occupy the same physical space
joints than the truth model. Once the calculated joints are during some combination of motions [13].
8
 Figure 5. Example results. a) Truth (blue ·) and calculated (red ∗) 3D position of points for an example frame. b)
Calculated points, kinematic chain, joint locations, and axes for an example frame. Points in each component are a
different color. c) Grid positions in both true and calculated workspace (blue ·), Grid positions in only true workspace
(red ·), Grid positions in only calculated workspace (green ·).
3. R ESULTS
Figure 5 shows an example of the results available from
each trial. There are numerous aspects of this method that
are non-deterministic such as the placement of points in
the truth model, the segmentation algorithm, and the rigid
body optimization initialization method. Therefore, numer-
ous simulations were conducted to evaluate the effect of the
illumination offset angle on the evaluation metrics. Results
for the fly-by inspection route are shown in Figure 6. Each
plot shows a fourth-order polynomial fit along with bounds
containing at least 50% of predicted values calculated using
Matlab’s polyfit command. While some metrics (Eφrange ,
Eφ̇ , and Pwoc ) do not seem to be effected by the illumination
offset angle, most metrics begin to degrade around θ = ±60◦ .
This is consistent with when the percentage of visible points
begins to decrease as shown in Figure 7.
Trials were also conducted for full NMCs with various illu-
mination offsets. The results are shown in Figure 8. As with
the fly-by, the results begin to diminish as the illumination
offset approaches θ = ±60◦ . This is again consistent with
the decrease in the percentage of visible points as shown in
Figure 7. This suggests that for these two types of inspection
routes the percentage of visible points does correspond to the
quality of the results. However, the overall results for the fly-
by are better than the overall results of the NMC for most
metrics even though the NMC views a much higher percent-
age of points. This suggests that generally there is a point at
which a higher percentage of visible points does not produce
better results. This is likely because more points being visible
results in a trajectory matrix with more missing data which
adds complexity to the ‘Combine Segments’ portion of the
method (section 2).
To further investigate this, additional simulations were run
in which the percentage of the NMC completed was varied
along with the illumination offset angle. As the percentage
of the NMC increases, more points will be visible, but the
amount of missing data in the trajectory matrix will also
increase. More visible points means more information is
available about the overall shape of the primary satellite,
however, more missing data complicates the segmentation
process and the combination of segments. This suggests there
9
is limit at which more data is not beneficial to the success of
the algorithm. To investigate this limit, an experiment was
developed to test the affect of the illumination offset and the
percentage of the NMC completed.
Results are shown in Figure 9. For ease of evaluation
and presentation, a single metric (Qall ) is used to quantify
the results from each simulation. Qall is the summation
of the metrics described in section 2 normalized by their
maximum. The percentage metrics (Pwc , Pwoc , Pgc ) are
normalized by their maximum while the error metrics (E3D ,
Eâ , EJ , Eφrange , Eφ̇ ) are normalized by their maximum and
subtracted from 1 so a better model is represented by a larger
number. The contour represents a surface created by a fourth-
order polynomial in each of the test variables fit to the Qall
metric. The points represent each of the test points with a
color corresponding to the value of Qall for that simulation.
After 50% of the NMC is completed, the results suggest
that additional data causes an overall decreasing trend in
the quality of the resulting model. This suggests that if the
inspection route covers more than 50% of the NMC, only the
data from the first half of the NMC should be used to create
the model. Any additional data could be processed using
some type of model refinement method in which new points
and frames are appropriately added to the existing model. The
results also suggest that an inspection route limited to 50% of
an NMC may be as capable of producing an adequate model
as a route of longer duration.
Limitations and future work
This method is capable of characterizing satellite articulation,
however limitations exist.
• There are a number of parameters that require adjustment
based on the specifics of the problem. These parameters are
shown in Table 1 along with the method of determining an
appropriate setting. Noise analysis (NA) means that the best
parameter setting is based on the amount of noise in the tra-
jectory matrix and the inspector satellite route. Prior problem
knowledge (PPK) means that some information regarding the
particular problem, such as an estimate on the number of
components, should be used to select an appropriate value.
 Figure 6. Fly-by trial results for each evaluation metric (outlined in section 2). Blue dots represent the value for each
trial, black lines are a 4th order polynomial fit to the data, and dashed lines contain 50% of future predictions
Figure 7. Percentage of truth model points that are
visible during the inspection route.
Cost function ratios (CFR) means that the ratio of the dif-
ferent terms in the cost function should be used to scale the
value appropriately. Point density (PD) means the number of
points found/tracked should dictate the criteria for removing
points. Risk assessment (RA) means the risk in having more
or less than the correct number of components should be
assessed. Investigating methods of automatically adjusting
these parameters is an area for future work.
10
• The current implementation only calculates parameters
for a component when the component is visible and when
all components above it in the kinematic chain are visible.
This could be highly limiting for certain cases in which a
component is shadowed or occluded for a significant number
of frames. Methods of reducing this limitation is an area for
future work.
• As currently implemented, this method requires the main
body of the satellite to be motionless in the world frame.
While this is certainly a limitation, this limitation could be
removed within the same framework. This will be investi-
gated in future work.
• It is assumed that the motion of the inspector satellite is
known. While this is likely to be the case generally, the exact
relative position may not be known. The robustness of this
method to noise in the inspector satellite route has not been
investigated.
• Real imagery and the challenges of feature point recogni-
tion and tracking in space illumination conditions have not
been investigated in this work.
• The current method performance decreases as more points
are visible. Future work will incorporate a method of using
additional information sequentially to improve the model and
continue to track articulated motion.
4. C ONCLUSIONS
The method presented successfully builds an articulated
model from a trajectory matrix simulated from a known
satellite inspection route. The method assumes revolute joints
and cannot accommodate cycles in the kinematic chain. The
 Figure 8. Full NMC trial results for each evaluation metric (outlined in section 2). Blue dots represent the value for
each trial, black lines are a 4th order polynomial fit to the data, and dashed lines contain 50% of future predictions
Figure 9. Results from Illumination offset angle and
Percentage of completed NMC testing. Contour plot
represents a 4th order polynomial fit of both
experimentation variables to the equally weighted
combination of all evaluation metrics. Dots represent
each trial colored by the value of the combined metric at
that point.
11
results in this work are for a model with 5 articulated joints,
3 of which are in a single chain, however previous work
demonstrated success of this method for a 6 joint chain
[13]. The method assumes the inspection route produces
resolved imagery of appropriate resolution for feature point
tracking and does not investigate the difficulties of the space
illumination environment on feature point recognition and
tracking. Eight new performance metrics are proposed and
used to evaluate inspection routes and the resulting models.
In this work, the effect of fly-by routes and NMCs with sub-
optimal illumination phasing were investigated. As expected,
for both fly-by and the NMC, the results degrade as the illu-
mination offset angle increases, with significant degradation
occuring beyond approximately ±75◦ . Results show that the
additional point visibility gained by routes with over 50%
NMC completion does not contribute to improved quality of
the model. This suggests that NMC inspection routes could
be cut-off after 50% completion, or if the route is continued,
data from the first 50% of the NMC should be used to create
the model and any remaining data should be added with
some type of model refinement method. Further work will
investigate model refinement methods.
ACKNOWLEDGMENTS
The authors would like to acknowledge Dr Frank Chavez and
others at the Air Force Research Laboratory Space Vehicles
directorate for their guidance in this research.
Table 1. Parameters. Method of determination: NA = R. T. Howard and R. D. Richards, Eds., vol. 6220, 2006,
Noise analysis; PPK=Prior problem knowledge; pp. 1–12.
CFR=Cost function ratios; PD=Point density; RA=Risk
assessment [7] G. Fasano, M. Grassi, and D. Accardo, “A stereo-
vision Based System for Autonomous Navigation of
Parameter Name/description Value Method an In-Orbit Servicing Platform,” in AIAA Infotech@
of Aerospace, Seattle, Washington, 2009, pp. 1–10.
determi-
nation [8] N. K. Philip and M. R. Ananthasayanam, “Relative
γs Stationary threshold: determines 1−6 NA position and attitude estimation and control schemes for
when a point is considered sta- the final phase of an autonomous docking mission of
tionary spacecraft,” Acta Astronautica, vol. 52, pp. 511–522,
k Number of segments: determines 10 PPK 2003.
the number of segments for spec- or
tral clustering. (sec. 2) 22 [9] S. J. Kelly, “A Monocular SLAM method to Estimate
γf Reprojection error fair share mul- 8 PD Relative Pose During Satellite Proximity Operations,”
tiplier for outlier rejection. (sec. Master’s Thesis, Air Force Institute of Technology,
2) 2015.
γr Range rejection multiplier for out- 4 PD,
lier rejection. (sec. 2) PPK [10] F. Yu, Z. He, B. Qiao, and X. Yu, “Stereo-vision-based
γc Segment merge criteria: allow- 2 RA relative pose estimation for the rendezvous and docking
able increase in function value for of noncooperative satellites,” Mathematical Problems in
acceptable merge. (sec. 2) Engineering, vol. 2014, p. 12, 2014.
γBB Joint bounding box expansion: 0.5 PPK [11] V. Ghadiok, J. Goldin, and D. Geller, “Gyro-Aided
determines the size of the bound- Vision-Based Relative Pose Estimation for Autonomous
ing box around the component
shape. (sec. 2) Rendezvous and Docking,” Advances in the Astronauti-
η Coefficient in joint penalty. (eqn. 100 CFR cal Sciences, vol. 149, pp. 713–728, 2013.
(14)) [12] D. H. Curtis and R. G. Cobb, “Satellite Articulation
λa Rotation parameter initialization 5 CFR Sensing using Computer Vision,” in Proceedings of
smoothness weight. (eqn. (19)) AIAA SciTech 2017, 2017.
λb Joint location initialization close- 5 CFR
ness weight. (eqn. (20)) [13] ——, “Satellite articulation characterization from an
γa Minimum articulation angle mean 0.007 RA image trajectory matrix using optimization,” in Ad-
rate. (IJA algorithm [13]) rad. vanced Maui Optical Space Surveillance Technologies
per Conference, 2017.
frame
γm Segment merge criteria: allow- 5 RA [14] J. Yan and M. Pollefeys, “A factorization-based ap-
able increase in function value for proach to articulated motion recovery,” in Proceedings
acceptable merge. (IJA algorithm of the IEEE Computer Society Conference on Computer
[13]) Vision and Pattern Recognition, vol. 2, 2005, pp. 815–
821.
[15] P. Tresadern and I. Reid, “Articulated structure from
R EFERENCES motion by factorization,” in Proceedings of the IEEE
[1] R. T. Howard, A. F. Heaton, R. M. Pinson, and C. K. Computer Society Conference on Computer Vision and
Carrington, “Orbital Express Advanced Video Guid- Pattern Recognition, vol. 2, 2005, pp. 1110–1115.
ance Sensor,” in IEEE Aerospace Conference Proceed- [16] C. Tomasi and T. Kanade, “Shape and motion from
ings, 2008, pp. 1–10. image streams under orthography: a factorization
[2] B. E. Tweddle and D. W. Miller, “Computer Vision- method.” International Journal of Computer Vision,
Based Localization and Mapping of an Unkown, Unco- vol. 9, no. 2, pp. 137–154, 1992.
operative and Spinning Target for Spacecraft Proximity [17] K. Yucer, O. Wang, A. Sorkine-Hornung, and
Operations,” Dissertation, Massachusetts Institute of O. Sorkine-Hornung, “Reconstruction of Articulated
Technology, 2013. Objects from a Moving Camera,” in IEEE International
[3] C.-C. J. Ho and N. H. McClamroch, “Automatic space- Conference on Computer Vision Workshops, 2015, pp.
craft docking using computer vision-based guidance 28–36.
and control techniques,” Journal of Guidance, Control, [18] M. Paladini, A. Del Bue, J. Xavier, L. Agapito,
and Dynamics, vol. 16, no. 2, pp. 281–288, 1993. M. Stošić, and M. Dodig, “Optimal metric projections
[4] B. E. Tweddle and A. Saenz-Otero, “Relative Computer for deformable and articulated structure-from-motion,”
Vision-Based Navigation for Small Inspection Space- International Journal of Computer Vision, vol. 96, no. 2,
craft,” Journal of Guidance, Control, and Dynamics, pp. 252–276, 2012.
vol. 38, no. 5, pp. 969–978, 2015. [19] M. Marques and J. Costeira, “Optimal shape from mo-
[5] B. J. Naasz, J. Van Eepoel, S. Z. Queen, C. M. South- tion estimation with missing and degenerate data,” in
ward, and J. Hannah, “Flight results from the HST SM4 2008 IEEE Workshop on Motion and Video Computing,
Relative Navigation Sensor system,” in 33rd Annual WMVC, no. 1, 2008, pp. 2–7.
AAS Guidance and Control Conference, Breckenridge, [20] P. B. Zhang and Y. S. Hung, “Articulated Structure
Colorado, 2010. from Motion through Ellipsoid Fitting,” in International
[6] S. J. Hannah, “A relative navigation application of UL- Conference of Image Processing, Computer Vision and
TOR technology for automated rendezvous and dock- Pattern Recognition, 2015, pp. 179–186.
ing,” in Proceedings of SPIE, Spaceborne Sensors III, [21] J. Fayad, C. Russell, and L. Agapito, “Automated articu-
12
 lated structure and 3D shape recovery from point corre- B IOGRAPHY [
spondences,” in Proceedings of the IEEE International
Conference on Computer Vision, 2011, pp. 431–438.
David Curtis received his B.S. degree in
[22] C. Russell, R. Yu, and L. Agapito, “Video Pop-up: Mechanical Engineering from Clarkson
Monocular 3D Reconstruction of Dynamic Scenes,” in University in 2005 and his M.S. degree
European Conference on Computer Vision. Springer in Aeronautical engineering from the Air
International Publishing, 2014, pp. 583–598. Force Institute of Technology in 2009.
[23] S. Pillai, M. R. Walter, and S. Teller, “Learning He is currently a Ph.D. candidate at
Articulated Motions From Visual Demonstration,” in the Air Force Institute of Technology.
Robotics: Science and Systems, 2014. His research interest are in the fields
of computer vision and optimal control,
[24] C. Sabol, R. Burns, and C. A. McLaughlin, “Satel- particularly for space applications.
lite formation flying design and evolution,” Journal of
Spacecraft and Rockets, vol. 38, no. 2, pp. 270–278,
2001. Richard Cobb received his B.S. from the
[25] R. Vidal, “Subspace Clustering,” IEEE Signal Process- Pennsylvania State University in 1988,
ing Magazine, no. March, pp. 52–68, 2011. his M.S. from the Air Force Institute of
Technology in 1992, and a Ph.D. from
[26] J. Yan and M. Pollefeys, “A general framework for the Air Force Institute of Technology in
motion segmentation: Independent, articulated, rigid, 1996. He is currently a professor of
non-rigid, degenerate and non-degenerate,” in European Aerospace Engineering at the Air Force
Conference on Computer Vision. Springer Berlin Institute of Technology. His research fo-
Heidelberg, 2006, pp. 94–106. cuses on dynamics and control of space
[27] A. Y. Ng, M. I. Jordan, and Y. Weiss, “On Spectral structures for space-based remote sens-
Clustering: Analysis and an Algorithm,” Advances in ing, and optimization and control for aerospace applica-
Neural Information Processing Systems, pp. 849–856, tions.
2002.
[28] J. Shi and J. Malik, “Normalized cuts and image seg-
mentation,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 22, no. 8, pp. 888–905, 2000.
[29] U. V. Luxburg, “A Tutorial on Spectral Clustering A Tu-
torial on Spectral Clustering,” Statistics and Computing,
vol. 17, no. March, pp. 395–416, 2006.
[30] I. Buerk, “SpectralClustering,”
p. 2, 2012. [Online]. Available:
https://www.mathworks.com/matlabcentral/fileexchange/34412-
fast-and-efficient-spectral-clustering
[31] J. L. Mundy and A. Zisserman, Geometric Invariance
in Computer Vision, J. Mundy and A. Zisserman, Eds.
The MIT Press, 1992.
[32] J. D’Errico, “Inhull,” 2006. [Online]. Available:
https://www.mathworks.com/matlabcentral/fileexchange/10226-
inhull

13