Conference PaperPDF Available

On-Ground Validation of a CNN-Based Monocular Pose Estimation System for Uncooperative Spacecraft

Abstract and Figures

The estimation of the relative pose of an inactive spacecraft by an active servicer spacecraft is a critical task for close-proximity operations, such as In-Orbit Servicing and Active Debris Removal. Among all the challenges, the lack of available space images of the inactive satellite makes the on-ground validation of current monocu-lar camera-based navigation systems a challenging task, mostly due to the fact that standard Image Processing (IP) algorithms, which are usually tested on synthetic images, tend to fail when implemented in orbit. This paper reports on the testing of a novel Convolutional Neural Network (CNN)-based pose estimation pipeline with realistic lab-generated 2D monocular images of the European Space Agency's Envisat spacecraft. Following the current need to bridge the reality gap between synthetic and images acquired in space, the main contribution of this work is to test the performance of CNNs trained on synthetic datasets with more realistic images of the target spacecraft. The validation of the proposed pose estimation system is assured by the introduction of a calibration framework, which ensures an accurate reference relative pose between the target spacecraft and the camera for each lab-generated image, allowing a comparative assessment at both keypoints detection and pose estimation level. By creating a laboratory database of the Envisat spacecraft under space-like conditions, this work further aims at facilitating the establishment of a standardized on-ground validation procedure that can be used in different lab setups and with different target satellites. The lab-representative images of the Envisat are generated at the Orbital Robotics and GNC lab of ESA's European Space Research and Technology Centre (ESTEC). The VICON Tracker System is used together with a KUKA robotic arm to respectively track and control the trajectory of the monocular camera around a scaled 1:25 mockup of the Envisat spacecraft.
Content may be subject to copyright.
L. Pasqualetto Cassinis(1), A. Menicucci(1) , E. Gill(1), I. Ahrns(2) , and J. Gil-Fern´
(1)Delft University of Technology, Kluyverweg 1 2629 HS, Delft, The Netherlands, Email: {L.PasqualettoCassinis,
A.Menicucci, E.K.A.Gill}
(2)Airbus DS GmbH, Airbusallee 1, 28199, Bremen, Germany, Email:
(3)ESTEC, Keplerlaan 1, 2201 AZ, Noordwijk, The Netherlands, Email:
The estimation of the relative pose of an inactive space-
craft by an active servicer spacecraft is a critical task for
close-proximity operations, such as In-Orbit Servicing
and Active Debris Removal. Among all the challenges,
the lack of available space images of the inactive satel-
lite makes the on-ground validation of current monocu-
lar camera-based navigation systems a challenging task,
mostly due to the fact that standard Image Processing (IP)
algorithms, which are usually tested on synthetic images,
tend to fail when implemented in orbit. This paper re-
ports on the testing of a novel Convolutional Neural Net-
work (CNN)-based pose estimation pipeline with realis-
tic lab-generated 2D monocular images of the European
Space Agency’s Envisat spacecraft. Following the cur-
rent need to bridge the reality gap between synthetic and
images acquired in space, the main contribution of this
work is to test the performance of CNNs trained on syn-
thetic datasets with more realistic images of the target
spacecraft. The validation of the proposed pose estima-
tion system is assured by the introduction of a calibra-
tion framework, which ensures an accurate reference rel-
ative pose between the target spacecraft and the camera
for each lab-generated image, allowing a comparative as-
sessment at both keypoints detection and pose estimation
level. By creating a laboratory database of the Envisat
spacecraft under space-like conditions, this work further
aims at facilitating the establishment of a standardized
on-ground validation procedure that can be used in differ-
ent lab setups and with different target satellites. The lab-
representative images of the Envisat are generated at the
Orbital Robotics and GNC lab of ESA’s European Space
Research and Technology Centre (ESTEC). The VICON
Tracker System is used together with a KUKA robotic
arm to respectively track and control the trajectory of the
monocular camera around a scaled 1:25 mockup of the
Envisat spacecraft.
Keywords: Convolutional Neural Networks; On-ground
Validation; Monocular Pose Estimation; Calibration Pro-
Nowadays, the safety and operations of satellites in or-
bit has become paramount for key Earth-based applica-
tions, such as remote sensing, navigation, and telecom-
munication. In this context, advancements in the field of
Guidance, Navigation, and Control (GNC) were made in
the past years to cope with the challenges involved in In-
Orbit Servicing (IOS) and Active Debris Removal (ADR)
missions [27, 29]. For such scenarios, the estimation of
the relative pose (position and attitude) of an uncoopera-
tive spacecraft by an active servicer spacecraft represents
a critical task. Compared to cooperative close-proximity
missions, the pose estimation problem is indeed compli-
cated by the fact that the target satellite is not functional
and/or not able to aid the relative navigation. Hence, opti-
cal sensors shall be preferred over Radio Frequency (RF)
sensors to cope with a lack of navigation devices such as
Global Positioning System (GPS) sensors and/or anten-
nas onboard the target.
In this framework, pose estimation systems based solely
on a monocular camera are recently becoming an attrac-
tive alternative to systems based on active sensors or
stereo cameras, due to their reduced mass, power con-
sumption and system complexity [23, 14]. However, a
significant effort is still required to comply with most
of the demanding requirements for a robust and accu-
rate monocular-based navigation system. Notably, the
aforementioned navigation system cannot rely on known
visual markers, as they are typically not installed on an
uncooperative target. Since the extraction of visual fea-
tures is an essential step in the pose estimation process,
advanced Image Processing (IP) techniques are required
to extract keypoints (or interest points), corners, and/or
edges on the target body. In model-based methods, the
detected features are then matched with pre-defined fea-
tures on an offline wireframe 3D model of the target to
solve for the relative pose. This is usually achieved by
solving the Perspective-n-Points (PnP) problem [21]. In
other words, a reliable detection of key features under
adverse orbital conditions is highly desirable to guaran-
tee safe operations around an uncooperative spacecraft.
Unfortunately, standard IP algorithms usually lack of fea-
ture detection robustness when applied to space images
[3], undermining the overall navigation system and, in
turn, the whole close-proximity operations around the un-
cooperative target. From a pose initialization standpoint,
the extraction of target features can in fact be jeopardized
by external factors, such as adverse illumination condi-
tions, low Signal-to-Noise ratio (SNR) and Earth in the
background, as well as by target-specific factors, such as
the presence of complex textures and features on the tar-
get body. Moreover, most of the IP methods are based
on the image gradient, detecting textured-rich features
or highly visible parts of the target silhouette. As such,
the detected features are image-specific and can vary in
number and typology depending on the image histogram.
This means that most of these techniques cannot accom-
modate an offline feature selection step, which translates
into a computationally expensive image-to-model corre-
spondence process to ensure that each detected 2D fea-
ture is matched with its 3D counterpart on the available
wireframe model of the target object.
In recent years, Convolutional Neural Networks (CNNs)
are emerging as a valid and robust alternative to stan-
dard monocular-based pose estimation systems, with two
main CNN-based architectures currently being investi-
gated. Initially, end-to-end architectures in which a single
CNN replaced the entire pose estimation pipeline were
more exploited [20, 22, 24, 25]. However, since the pose
accuracies of these systems proved to be lower than the
accuracies returned by standard PnP solvers, especially
in the estimation of the relative attitude [20], keypoints-
based architectures stood out as the preferred option.
Specifically, average orientation errors of 1.31o±2.24o
were achieved by keypoints-based methods as opposed to
the average orientation errors of 9.76o±18.51oachieved
by end-to-end methods. These averages were computed
across test images of the TANGO spacecraft as part of
the Spacecraft Pose Estimation Dataset challenge [8]. In
keypoints-based CNN systems, a CNN is used only at a
feature detection level to replace standard IP algorithms,
and the output features are fed to a PnP solver together
with their body coordinates, which are made available
through the wireframe 3D model of the target body. Due
to the fact that the trainable features can be selected of-
fline prior to the training, the matching of the extracted
feature points with the features of the wireframe model
can be performed without the need of a large search
space for the image-model correspondences, which usu-
ally characterizes most of the edges/corners-based meth-
ods [3]. However, due to a lack of availability of repre-
sentative space images, these CNN systems often need to
be trained with synthetic renderings of the available tar-
get model. As a result, their feature detection robustness
on more realistic images is usually unknown and difficult
to predict.
In this context, the on-ground validation of the CNNs’
performance shall be sought by testing their robustness
against representative images of the target spacecraft,
generated in a laboratory environment which recreates
space-like illumination conditions. In this way, the per-
formance of synthetically-trained CNNs on lab-generated
images can be tested. Moreover, a calibration framework
shall be established which returns an accurate reference
for the relative pose between the monocular camera and
the target mockup for each generated image, in order to
be able to quantify the CNN performance at both key-
points detection and pose estimation levels.
Several laboratory setup exist to recreate rendezvous ap-
proaches around a mockup of a target spacecraft with a
monocular camera [30], e.g. the Space Rendezvous Lab-
oratory (SLAB) at Stanford University [8], the Orbital
Robotics & GNC laboratory (ORGL) at the European
Space Research and Technology Centre (ESTEC) [31],
and the Testbed for Robotic Optical Navigation (TRON)
at the German Aerospace Agency (DLR) [9]. However,
only a few detailed calibration procedures were recently
described which allow the accurate estimation of the ref-
erence relative pose between camera and target [28]. Be-
sides, the calibration of the target spacecraft highly de-
pends on the presence (cooperative target) or not (unco-
operative target) of visual markers, as well as on the ren-
dezvous trajectory that shall be recreated (static or rota-
tional target). As a result, the existing calibration proce-
dures usually require adaptations to the specific setup.
In relation to the on-ground validation of CNN-based
pose estimation systems, an additional challenge also
arises from bridging the gap between the synthetic ren-
derings and the lab-representative images. If the syn-
thetic dataset used to train the CNN fails in representing
the textures of the target mockup as well as the specific
illumination in the laboratory setup, the performance on
lab-generated images will in fact result in inaccurate de-
tections and lead to low pose estimation accuracies. To
overcome this, recent works addressed the impact of aug-
mented synthetic datasets on the CNN performance in ei-
ther lab-generated or space-based imagery [13, 1]. These
augmented datasets are built on a backbone of purely syn-
thetic images of the target by adding noise, randomized
and real Earth background, and randomized textures of
the target model. However, the synthetic and laboratory
environment are usually tuned to return a high representa-
tiveness of the synthetic images. Furthermore, the same
3D model is usually used in bot the synthetic rendering
and in the laboratory setup. As a result, the CNN detec-
tion robustness against variations in the target model has
not been fully addressed yet.
In this framework, the main objectives of this paper are:
To propose a calibration procedure capable of
estimating accurate reference poses between the
monocular camera and the target spacecraft
To investigate the impact of datasets augmentation
and randomization on the CNN training, validation
and testing
To improve the performance of synthetically-trained
CNNs on lab-generated images.
Specifically, the main novelty of this work is to inves-
tigate the performance of the proposed pose estimation
system when the mockup of the target spacecraft differs
from the rendering model used to synthetically-train the
The paper is organized as follows. Section 2 introduces
the proposed pose estimation framework. The laboratory
setup and the calibration procedure are described in Sec-
tions 3-4. In Section 5, the CNN training, validation and
testing phases are detailed. Special focus is given to the
augmentation and randomization pipeline. Section 6 il-
lustrates the adopted pose estimation methods, whereas
the results are presented in Section 7. Finally, Section 8
provides the main conclusions and recommendations.
From a high-level perspective, a model-based monocular
pose estimation system receives as input a 2D image and
matches it with an existing wireframe 3D model of the
target spacecraft to estimate the pose of such target with
respect to the servicer camera. Referring to Figure 11,
the pose estimation problem consists in determining the
position of the target’s centre of mass tCand its orienta-
tion with respect to the camera frame C, represented by
the rotation matrix RC
B. The Perspective-n-Points (PnP)
p= (ui, vi) = xC
relate the unknown pose with a feature point pin the im-
age plane via the relative position rCof the feature with
respect to the camera frame. Here, rBis the point loca-
tion in the 3D model, expressed in the body-frame co-
ordinate system B, whereas fxand fydenote the focal
lengths of the camera and (Cx, Cy) is the principal point
of the image.
From these equations, it can be seen that an important
aspect of estimating the pose resides in the capability of
the IP system to extract features pfrom a 2D image of
the target spacecraft, which in turn need to be matched
with pre-selected features rBin the wireframe 3D model.
Notably, such wireframe model of the target needs to be
made available prior to the estimation.
The on-ground validation pipeline of the proposed pose
estimation system is shown in Figure 2 and consists of
the following main stages:
1. Calibration procedure and Image Acquisition:
laboratory images of a scaled 1:25 mockup model
of the Envisat spacecraft are generated by mounting
the camera on a robotic arm which performs a tra-
jectory around the mockup. Besides, the camera is
Figure 1. Schematic of the PnP problem using a monoc-
ular image (Figure adapted from [23]).
intrinsically and extrinsically calibrated with respect
to the Envisat mockup in order to associate refer-
ence labels of the relative pose between the adopted
monocular camera and the mockup for each gener-
ated image
2. Dataset Generation and CNN Training: a
keypoints-based CNN is trained and validated on
augmented datasets. The augmentation is performed
by introducing image noise, artificial lights, random
background and random textures into synthetically-
generated images of a rendering model of the En-
visat spacecraft
3. Online Inference: the keypoints-based CNN is
tested on both synthetic and lab-generated images.
The relative pose is estimated by feeding a PnP
solver with the detected keypoints as well as with
the intrinsic camera parameters and 3D model of En-
4. Validation of Pose Estimation Results: the CNN-
based pose estimation results on the lab-generated
images are validated against the reference pose la-
bels, derived from the calibrated objects.
The adopted laboratory setup is illustrated in Figure 3
and makes use of the GNC Rendezvous, Approach and
Landing Simulator (GRALS) testbed of the ORGL fa-
cility at ESTEC. The setup is constituted of the follow-
ing elements: (a) a 1:25 scaled mockup of the Envisat
spacecraft mounted on a black-painted, static tripod; (b) a
Prosilica GT4096 monocular camera mounted on a fixed
aluminum plate; (c) a ceiling KUKA robotic arm, used
to move the camera around the mockup; (d) the VICON
Figure 2. Illustration of the proposed on-ground validation of the CNN-based pose estimation system.
Figure 3. GRALS facility with the scaled 1:25 Envisat mockup, the VST and the monocular camera mounted on the KUKA
robotic arm. One of the VTS cameras and the markers object used in the extrinsic calibration of the camera are also
Tracker System (VTS), used to track objects with retro-
reflective markers and to provide estimates of their pose
with respect to a user-defined reference frame; (e) an ex-
ternal computer providing the software interface between
the monocular camera, the VTS and the KUKA robotic
3.1. VICON Tracking System
The VTS is a highly accurate motion capture system ca-
pable of tracking dynamic objects with millimeter accu-
racy [11]. The system includes a set of calibrated IR cam-
eras, some retro-reflecting spherical markers which can
be detected and tracked by the cameras, and a software
interface to stream telemetry to the external computer. In
the current setup, a subset of 10 cameras is selected such
that the total field of view covers the operating volume in
which the image acquisition is carried out.
3.2. KUKA Software and Hardware Elements
The KUKA robotic arm is controlled from the external
computer via a Robot Software Interface (RSI) connec-
tion. The arm can translate along a ceiling rail and ro-
tate around its six joints, thus guaranteeing the execution
of an elliptical trajectory around the Envisat mockup at
around 1-2 m distance.
The calibration setup consists of the elements described
in Section 3 and is inspired by the calibration procedure
reported in [28]. The objective is to estimate the rela-
tive pose between the monocular camera and the Envisat
mockup for each generated image.
4.1. Reference Frames Definition
Referring to Figure 4, the following reference frames are
VRT Reference Frame O: this is the reference frame
in which all the objects tracked by VTS are ex-
Camera Frame C: this frame is defined such that the
third axis is perpendicular to the image plane and
is aligned with the optical axis of the camera, with
the other two axes planar to the focal plane of the
Plate Reference Frame I: this reference frame is
built from retro-reflective VTS markers and is
rigidly attached to the camera mounting plate
Figure 4. Illustration of the reference frames adopted
during the calibration procedure.
Envisat Body Frame B: this is a rigid frame oriented
with its axis parallel to the principal axes of inertia
of the Envisat mockup and centered on the service
Markers Object Frame M: this frame is built from
retro-reflective VTS markers attached to a planar
The transformation between each of these frames can be
expressed by a roto-translation matrix T, which incorpo-
rates the relative rotation matrix Rand the relative posi-
tion vector t,
T=R t
4.2. Camera Intrinsic Calibration
The first step of the calibration procedure consists of the
estimation of the camera intrinsic parameters, such as the
focal length, the principal point and the tangential and ra-
dial distortion coefficients. This is accomplished by tak-
ing images of a chessboard with different camera views
and using the estimateCameraParameters Matlab built-in
function. The function estimates for the intrinsic param-
eters and the distortion coefficients of a single camera,
whilst also returning the images used to estimate the cam-
era parameters and the standard estimation errors for the
single camera calibration.
4.3. Extrinsic Calibration
Once the camera intrinsic parameters are estimated, the
relative roto-translation matrix TB
Cbetween the camera
frame C and the Envisat body frame B shall be estimated.
The procedure consists of the following steps:
Figure 5. Illustration of the location of reference frame I
with respect to the camera frame C. Note that the exact
location of the C frame is unknown prior to calibration.
Estimation of the roto-translation matrix TI
C- Cam-
era Extrinsic Calibration
Estimation of the roto-translation matrix TB
Mockup-to-VTS Calibration
Estimation of the roto-translation matrix TC
Mockup-to-Camera Calibration
4.3.1. Estimation of the roto-translation matrix TI
The first task is to recreate the objects M and I in Figure 4
by placing some retro-reflective markers onto the camera
mounting plate and a planar surface, respectively. Based
on similar setups [28], 15 markers were used to recreate
the object M, whereas a total of 9 markers were chosen
for the camera mounting plate in order to guarantee a re-
liable tracking by the VTS throughout the whole image
acquisition trajectory. Figure 5 illustrates the location of
the I frame with respect to the camera frame C. Only four
out of the nine markers are shown for clarity.
Next, the planar object M is moved in order to gener-
ate pictures of the retro-reflective markers under differ-
ent camera views. The pixel location of each marker
is then extracted by using the Matlab built-in Circular
Hough Transform (CHT) algorithm. This is shown on
the left-hand side of Figure 6. A manual 2D-3D point
correspondence is performed in order to associate each
detected marker with its three-dimensional location in
the M frame. At this stage, the Efficient Perspective-n-
Points (EPnP) algorithm is used to solve the PnP problem
and obtain an estimate of the roto-translation between the
camera frame C and the object frame M. The estimation
result is shown in the right-hand side of Figure 6. At this
stage, an initial estimate of the constant roto-translation
matrix TI
Ccan be obtained from the roto-translations TC
Oand TI
O, the latter two being returned by the VTS.
Subsequently, several pictures of the object M are taken
with different camera views, and the CHT is applied to
each of them to extract the pixel location of the retro-
reflective markers. For each frame, the 2D-3D point cor-
respondence can be made by using the initial estimate of
(a) (b)
Figure 6. Estimation of the roto-translation between the
camera frame C and the markers frame M. The markers
detection by the CHT algorithm (a) is shown beside the
estimated roto-translation of the camera with respect to
the the M object (b).
C. The PnP problem can then be solved by means of
a non-linear least squares solver, by minimizing the fol-
lowing sum of squares [28]:
pf,i (k)πMO
f,i (k),TI
f,i (k),TI
O= xC
f,i =xC
f,i yC
f,i zC
f,i T=RC
f,i +RC
where Nmis the number of fiducial markers, Nais the
number of frames and MO
f,i represents the location of the
ithmarker in the marker frame M. The output of the min-
imization is a refined estimate of TI
4.4. Estimation of the roto-translation matrix TB
Mockup Calibration
The adopted procedure to estimate the roto-translation
matrix TB
Odoes not require the placement of retro-
reflective markers on the Envisat mockup, taking advan-
tage of the fact that the mockup is kept fixed throughout
the image acquisition. The first step consists in acquir-
ing a few images of the Envisat from different camera
views (Figure 7). For each frame, the pixel location of
pre-selected natural features of Envisat is hand-picked
and a 2D-3D point correspondence is created with the
three-dimensional points of an available 3D model. In
this work, the corners of the Envisat body and of the SAR
antenna were considered. Next, the EPnP is used to es-
timate the camera-to-Envisat roto-translation matrix TC
for each frame.
Figure 7. Example of a camera view of the Envisat
mockup used for the estimation of the roto-translation
matrix TB
O. The visible hand-picked corners are marked
with red circles.
Figure 8. Reprojection of the hand-picked corners of En-
visat with the refined estimate of TB
Ofor a representative
By knowing the roto-translation of the I frame with
respect to the VICON origin O as well as the roto-
translation matrix TI
C, it is then possible to obtain a raw
estimate of the roto-translation matrix TB
Ofor each frame.
Due to the inaccuracies involved in the manual selection
of the Envisat features as well as of the EPnP estimates
of TC
B, the estimates of the constant matrix TB
Owill be
different from each other and will require an additional
refinement. This is accomplished once again by mini-
mizing the total reprojection error of the selected features
in a fashion similar to the one adopted in Section 4.3.1.
Figure 8 shows the hand-picked features correctly repro-
jected with the refined estimate of TB
Ofor a representa-
tive frame. Notably, and possibly due to inaccuracies still
present in the estimates of TI
Cand TB
O, a larger reprojec-
tion error was observed for a few frames. As such, future
adaptations of the calibration procedure, such as a more
accurate calibration of both the mockup and the camera,
should be considered to increase the validity of the refer-
ence pose labels for each generated image.
4.5. Estimation of the roto-translation matrix TC
Once the constant roto-translation matrices TC
Iand TB
are estimated, they can be used together with the VTS es-
timates of TI
Oto return the desired roto-translation TC
between the Envisat frame B and the camera frame C
throughout the entire image acquisition trajectory.
As already mentioned in Section 1, CNNs are currently
emerging as a promising features extraction method. This
is mostly due to the capability of their convolutional lay-
ers to extract high-level features of objects with improved
robustness against image noise and illumination condi-
tions [15]. Referring to Figure 9, the first essential step of
keypoints-based CNN systems is represented by an Ob-
ject Detection Network (e.g. Faster R-CNN [18], R-FCN
[17] or MobileNet [6]) placed before the main CNN. The
ODN regresses the coordinates of a bounding box around
the target object, in order to crop a Region Of Interest
(ROI) and to allow robustness to scale, variation, and
background textures. The cropped ROI is then fed into
a Keypoint Detection Network, which convolves with the
input image and outputs a set of feature maps. These so-
called heatmaps are detected around pre-selected features
on the target object, such as corners or interest points.
The 2D pixel coordinates of the heatmap’s peak intensity
characterize the predicted feature location, with the inten-
sity and the shape indicating the confidence of locating
the corresponding keypoint at this position [16]. Notably,
the selection of the CNN will drive the achievable key-
points detection accuracy and robustness. Some archi-
tectures, such as the stacked Hourglass [12] and the U-
Net [19], perform a downsampling of the input followed
in series by an upsampling, in order to detect features
at different scales. However, recent advancements in the
field [2] demonstrated that by using parallel sub-networks
across multiple resolutions, rather than multi-resolution
serial stages, the CNN can manage to maintain a richer
feature representation, facilitating more accurate and pre-
cise heatmaps. For this reason, the HRNet [26] architec-
ture currently represents the state-of-the-art in keypoint
detection, and it is chosen in the proposed pose estima-
tion system.
5.1. Augmentation and Randomization Pipeline
Referring to Figure 10, the first step of the proposed
pipeline for the datasets augmentation and randomiza-
tion consists in generating ideal synthetic images of the
Envisat 3D model. A highly-textured, realistic Envisat
model is rendered in the Cinema 4D©software by keep-
ing the virtual camera (Table 1) fixed and by randomly
varying the pose of the rendering model with respect
to the camera. Besides, the Azimuth and Elevation of
the Sun are randomly varied by ±40 deg around the
Figure 9. Proposed CNN architecture and interface with the PnP solver.
Table 1. Parameters of the camera used to generate the
synthetic images in Cinema 4D©.
Parameter Value Unit
Image resolution 256×256 pixels
Focal length 3.9·103m
Pixel size 1.1·105m
ideal camera-Sun relative position, in order to recreate
favourable as well as more adverse illumination condi-
tions. Next, a randomization pipeline is introduced which
adds the following effects to the rendering:
Texture randomization. This is performed in order
to increase the CNN robustness against texture vari-
ations between the synthetic and lab models of En-
visat. The randomization is achieved in two differ-
ent ways, by either adding a shader to each material
in order to noise the textures, or by directly shuffling
the textures of the materials.
Light randomization. Four additional lights are in-
troduced in random locations, aside from the main
Sun illumination, in order to increase the CNN ro-
bustness against the illumination conditions recre-
ated in the laboratory setup.
• Background randomization. Random scenes are
used as image background in order to increase the
CNN robustness against the laboratory environment.
Specifically, external disturbance sources in the lab
are likely to return non-zero pixel values in the im-
age background, leading to inaccurate CNN detec-
tions in case the training dataset would consist of
only black backgrounds.
Following the Cinema4D©rendering, an additional
pipeline is used to further augment the generated images.
This is performed by introducing the Earth in the back-
ground in some of the images and by corrupting the im-
ages with the following noise models:
Gaussian, shot, impulse and speckle noise
Gaussian, defocus, motion and zoom blurs
Spatter, color jitter and random erase
Table 2 lists all the augmentation techniques together
with the number of generated images. A total of 24,400
images were rendered and further split into training
(70%), validation (15%) and test (15%) datasets.
5.2. Training, Validation and Test
During training, the validation dataset is used beside the
training one to compute the validation losses and avoid
overfitting. The Adam optimizer [7] is used with a co-
sine decaying learning rate with initial value of 103and
decaying factor of 0.1. The network is trained for a to-
tal of 210 epochs. Finally, the network performance after
training is assessed with the synthetic test dataset.
Figure 10. Dataset Augmentation Pipeline.
Table 2. Augmentation Breakdown.
Description number of Images
No augmentations 1000
Random lights 550
Random lights & textures 2000
Random lights & background 350
Randomization & Noise & Earth 20,500
Total 24,400
(a) Shader effect (b) Randomized textures
(c) Random background (d) Motion blur
Figure 11. Output examples of the randomization
The performance is assessed in terms of Root Mean
Squared Error (RMSE) between the ground truth (GT)
and the x,ycoordinates of the extracted features, which
is computed as
[(xGT,i xi)2+ (yGT ,i yi)2]
The CNN performance on the test dataset shows a mean
detection accuracy of 0.97, with a RMSE mean µ= 2.78
pxl and a Mean Absolute Deviation (MAD) of 2.87 pxl.
Overall, this proves that the network is capable of accu-
rately detecting the pre-trained keypoint features in most
of the test images. Figure 12 shows a mosaic of key-
point detection results on a subset of the test dataset. No-
tably, wrong detections occur when the solar panel com-
pletely hides the main Envisat body. However, the CNN
returns good detection accuracies when only parts of En-
visat are occluded, demonstrating the capability of learn-
ing the relative position between features during partial
observability (Figure 13).
Following the promising results presented by the authors
in [15], the CEPPnP method [4] is selected to estimate the
relative pose from the detected features. In this method,
the CNN heatmaps around the detected features are ex-
ploited to derive feature covariance matrices and capture
the statistical distribution of the detected features.
The first step of the CEPPnP algorithm is to rewrite the
PnP problem in Eqs.1-2 as a function of a 12-dimensional
vector ycontaining the control point coordinates in the
camera reference system,
M y =0,(8)
Figure 12. Mosaic of keypoints detection results on a subset of the test dataset.
(a) acc = 0.95, RMSE = 0.47 pxl (b) RMSE =93 pxl
Figure 13. Example of high (a) and low (b) detection
accuracies during poor visibility or occlusion.
where Mis a 2n×12 known matrix. This is the funda-
mental equation in the EPnP problem [10]. The likeli-
hood of each observed feature location uiis then repre-
sented as
P(ui) = k·e1
where uiis a small, independent and unbiased
noise with expectation E[∆ui] = 0and covariance
i] = σ2Cuiand kis a normalization con-
stant. Here, σ2represents the global uncertainty in the
image, whereas Cuiis the 2x2 unnormalized covariance
matrix representing the Gaussian distribution of each de-
tected feature, computed from the CNN heatmaps. Af-
ter some calculations [4], the EPnP formulation can be
rewritten as
This is an eigenvalue problem in which both Nand L
matrices are a function of yand Cui. The problem is
solved iteratively by means of the closed-loop EPPnP
solution for the four control points, assuming no feature
uncertainty. Once yis estimated, the relative pose
is computed by solving the generalized Orthogonal
Procrustes problem used in the EPPnP [5].
To derive Cuifor each feature, each heatmap distribution
is used to compute a weighted covariance between x, y,
Cui=cov(x, x)cov(x, y)
cov(y, x)cov(y, y),(11)
cov(x, y) =
and nis the number of pixels in each feature’s heatmap.
In order to represent a distribution around the peak of the
detected feature, rather than around the heatmap’s mean,
the mean is replaced by the peak location p= (px, py).
This is particularly relevant when the heatmaps are asym-
metric and their mean does not coincide with their peak.
In this section, the pose estimation results are presented
for both the synthetic test dataset and the mockup images
generated at the ORGL facility of ESTEC. Two separate
error metrics are adopted in the evaluation, in accordance
with Kisantal et al. [8]. Firstly, the translational error be-
tween the estimated relative position ˆ
tCand the ground
truth tis computed as
This metric is also applied for the translational and rota-
tional velocities estimated in the navigation filter. Sec-
ondly, the attitude accuracy is measured in terms of the
Euler axis-angle error between the estimated quaternion
qand the ground truth q,
β= (βsβv) = qˆ
ER= 2 arccos (|βs|).(15)
7.1. Synthetic Test Dataset
The CNN-detected keypoints and the heatmaps-derived
covariances are fed into the CEPPnP solver together
with the intrinsic camera parameters and the Envisat 3D
model to solve for the relative pose. Besides, the perfor-
mance of the CEPPnP algorithm is evaluated against the
covariance-free EPnP solver [10], to assess the impact of
the feature covariance on the pose estimation accuracy.
Table 3 shows the results across the test dataset in terms
of mean µand MAD.
Table 3. Pose Estimation performance results for the syn-
thetic test, expressed in terms of µ±MAD
ET[m] 2.25 ±3.3 5.83 ±10.2
ER[deg] 2.7 ±2.8 2.3 ±2.6
Additionally, the estimated relative position and attitude
are expressed as a function of the relative range in order
to capture the trend of the pose estimation accuracy for
increasing relative distances. This is represented in Fig-
ure 14. As can be seen, including the heatmaps-derived
covariances results in a more robust and accurate pose
estimation, specifically due to an improved estimate of
the relative position (Figure 14a). In other words, the
CEPPnP position estimate is characterized by a more ac-
curate mean µand a smaller MAD, indicating that the es-
timation performance is improved in those scenarios for
which the CNN detections are less accurate. These con-
siderations confirm the results reported by the authors in
earlier works [15].
7.2. ORGL Dataset
The CNN performance on the ORGL dataset is evalu-
ated at both keypoints detection and pose estimation lev-
els. Firstly, the keypoints detection of the proposed CNN,
30 40 50 60 70 80 90 100
Relative range [m]
Mean Position Error [m]
30 40 50 60 70 80 90 100
Relative range [m]
Mean Attitude Error [deg]
Figure 14. Pose Estimation Results - The standard devia-
tion of the position (a) and attitude (b) errors is depicted
as the length of each error bar above and below the mean
errors ET,ER.
trained with the randomized training dataset, is compared
with the keypoints detection of the same CNN trained on
a subset of the augmented dataset, characterized only by
Earth in the background and noise. This is shown in Fig-
ure 15 for a sample image. Due to a lack of background,
light and texture randomization in the training dataset,
the CNN trained only on the partially-augmented dataset
is overfitted on the textures learned on the synthetic En-
visat model. As a result, the network cannot associate
the correct texture to each feature, and the detected key-
points are randomly scattered around the image (Figure
15a). Conversely, the CNN trained on the randomized
dataset proves to be more robust against variations in tex-
ture and light between the synthetic and the lab images,
inferring the correct shape of the mockup and detecting
most of the keypoints in the correct location (Figure 15b).
This improved robustness is mostly linked to the capabil-
ity of the CNN to learn shapes rather than textures, which
can be traced back to the textures randomization step in-
cluded in the augmentation and randomization pipeline.
Remarkably, the features are detected even without a high
synthetic-lab representativeness, showing the CNN ca-
pability to transfer to images which considerably differ
from the training ones.
Next, the pose estimates are compared with the reference
pose labels computed from the calibration procedure de-
scribed in Section 4. Table 4 lists the pose estimation
error across 100 ORGL images. As can be seen, the pose
estimates result in a large mean attitude error, despite an
(a) No randomization (b) Light/textures/background randomization
Figure 15. Impact of light, textures and background randomization on the CNN detection performance for a sample ORGL
image. Notably, the randomization of the training dataset improves the CNN robustness against different light, texture
and background conditions.
Table 4. Pose Estimation performance results for the
ORGL images, expressed in terms of µ±MAD
ET[m] 0.77 ±0.68 0.78 ±0.99
ER[deg] 30.5 ±22.7 20.8 ±13.7
accurate estimation of the relative position. Moreover,
the inclusion of feature covariances in the PnP solver does
not seem to improve the estimation accuracy. There are
at least two potential causes of this behaviour. On the
one hand, the relative distance between the monocular
camera and the Envisat mockup is of approximately 1 m,
meaning that relatively small pixel errors can lead to large
attitude errors. On the other hand, slight differences be-
tween the shape of the rendering model and the mockup
could affect the attitude estimate more than the position.
Besides, the reference attitude labels could be inaccurate
for some of the generated images, mostly due to some
inaccurate VTS telemetry or as a result of inaccurate es-
timates of the roto-translation matrices in Section 4.3.1.
Nevertheless, pose estimation results for a subset of the
ORGL dataset (including Figure 15b) are characterized
by relative attitude errors of <4 deg and relative position
errors of 5 cm, even in presence of adverse illumina-
tion conditions. This shows not only that accurate pose
estimates can be returned by the proposed CNN-based
pose estimation system, but also that most of the refer-
ence pose labels can be used to validate the performance
of the system on lab-generated images.
This paper introduces a framework for the on-ground
validation of a CNN-based monocular pose estimation
system for uncooperative spacecraft. A calibration
procedure is proposed to support the generation of
realistic laboratory images of a 1:25 scaled mockup of
the Envisat spacecraft. These images are used to test
the capability of the proposed CNN to bridge the gap
between synthetic training and laboratory testing whilst
returning accurate pose estimates.
The adopted CNN is validated at different levels of the
proposed pose estimation system, by assessing its perfor-
mance both in terms of keypoints detection and pose esti-
mate. At a keypoint detection level, the system proves to
benefit from the augmentation and randomization of the
dataset used during the CNN training. The results show
that the robustness of the CNN against lab-generated im-
ages can be increased by randomizing lights, material
textures and image background. This also allows to in-
crease the robustness against differences between the ren-
dering model and the mockup. At a pose estimation
level, the results on the synthetic test dataset indicate that
the covariant-based CEPPnP solver returns more accurate
pose estimates than the standard EPnP solver, thanks to
the capability of the heatmaps-based covariances to cap-
ture the statistical information of the detected features.
Besides, the accurate pose estimates reported for a sub-
set of the ORGL dataset indicate that the system built on
a synthetic training can transfer to more realistic images
of the target. Furthermore, it demonstrates that the cal-
ibration procedure returns accurate pose labels for most
of the generated images.
However, further work is still required. First of all, the
calibration procedure shall be revisited and improved in
order to guarantee an accurate and reliable pose label for
each generated image. Secondly, different augmentation
and randomization pipelines shall be investigated in order
to further improve the CNN performance. Finally, a more
comprehensive dataset of lab-generated images shall be
created in order to address the impact of including realis-
tic images on the training on the CNN performance.
This study is funded and supported by the European
Space Agency and Airbus Defence and Space under
Network/Partnering Initiative (NPI) program with grant
number NPI 577 - 2017. The first author would like
to thank Martin Schwendener and Irene Huertas for the
help during the image acquisition campaign at ORGL,
and Kuldeep Barad for the adaptation of the HRNet.
1. K. Black, S. Shankar, D. Fonseka, J. Deutsch,
A. Dhir, and M. Akella. Real-time, flight-ready, non-
cooperative spacecraft pose estimation using monoc-
ular imagery. In 31st AAS/AIAA Space Flight Me-
chanics Meeting, 2021.
2. B. Chen, J. Cao, A. Parra, and T. Chin. Satellite
Pose Estimation with Deep Landmark Regression
and Nonlinear Pose Refinement. In International
Conference on Computer Vision, Seoul, South Ko-
rea, 2019.
3. S. D’Amico, M. Benn, and J. Jorgensen. Pose Esti-
mation of an Uncooperative Spacecraft from Actual
Space Imagery. International Journal of Space Sci-
ence and Engineering, 2(2):171–189, 2014.
4. L. Ferraz, X. Binefa, and F. Moreno-Noguer. Lever-
aging Feature Uncertainty in the PnP Problem. In
Proceedings of the British Machine Vision Confer-
ence, Nottingham, UK, 2014.
5. L. Ferraz, X. Binefa, and F. Moreno-Noguer. Very
Fast Solution to the PnP Problem with Algebraic
Outlier Rejection. In IEEE Conference on Com-
puter Vision and Pattern Recognition, Columbus,
OH, USA, 2014.
6. A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko,
W. Wang, T. Weyand, M. Andreetto, and H. Adam.
MobileNets: Efficient Convolutional Neural Net-
works for Mobile Vision Applications. In ArXiv
Preprint, 2017.
7. D.P. Kingma and J. Ba. Adam: A method for
stochastic optimization. In 3rd International Confer-
ence for Learning Representations, San Diego, CA,
USA, 2015.
8. M. Kisantal, S. Sharma, T.H. Park, D. Izzo,
M. Martens, and S. D’Amico. Satellite Pose Esti-
mation Challenge: Dataset, Competition Design and
Results. IEEE Transactions on Aerospace and Elec-
tronic Systems, 2020.
9. H. Kr´
uger and S. Theil. TRON - hardware-in-the-
loop test facility for lunar descent and landing optical
navigation. In IFAC-ACA 2010 Automatic Control in
Aerospace, 2010.
10. Lepetit, F. Moreno-Noguer, and P. Fua. EPnP: an
accurate O(n) solution to the PnP problem. Inter-
national Journal of Computer Vision, 81:155–166,
11. P. Merriaux, Y. Dupuis, R. Boutteau, P. Vasseur, and
X. Savatier. A study of vicon system positioning per-
formance. Sensors, 17(7):1591, 2017.
12. A. Newell, K. Yang, and J. Deng. Stacked hourglass
networks for human pose estimation. In B. Leibe,
J. Matas, N. Sebe, and M. Welling, editors, Com-
puter Vision - ECCV 2016, volume 9912, pages 483–
499. Springer, Cham, 2016.
13. T.H. Park, S. Sharma, and S. D’Amico. Towards Ro-
bust Learning-Based Pose Estimation of Noncooper-
ative Spacecraft. In AAS/AIAA Astrodynamics Spe-
cialist Conference, Portland, ME, USA, 2019.
14. L. Pasqualetto Cassinis, R. Fonod, and E. Gill. Re-
view of the Robustness and Applicability of Monoc-
ular Pose Estimation Systems for Relative Naviga-
tion with an Uncooperative Spacecraft. Progress in
Aerospace Sciences, 110, 2019.
15. L. Pasqualetto Cassinis, R. Fonod, and E. Gill. Eval-
uation of tightly- and loosely-coupled approaches in
CNN-based pose estimation systems for uncooper-
ative spacecraft. Acta Astronautica, 182:189–202,
16. G. Pavlakos, X. Zhou, A. Chan, K.G. Derpanis, and
K. Daniilidis. 6-DoF Object Pose from Semantic
Keypoints. In IEEE International Conference on
Robotics and Automation, 2017.
17. S. Ren, K. He, R. Girshick, and J. Sun. Object detec-
tion via region-based fully convolutional networks.
Advances in neural information processing systems,
pages 379–387, 2016.
18. S. Ren, K. He, R. Girshick, and J. Sun. Faster R-
CNN: Towards Real-Time Object Detection with Re-
gion Proposal Networks. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 39(6):1137
– 1149, 2017.
19. O. Ronneberger, P. Fischer, and T. Brox. U-
net: Convolutional networks for biomedical im-
age segmentation. In Medical Image Computing
and Computer-Assisted Intervention, pages 234–
241. Springer, 2015.
20. S. Sharma, C. Beierle, and S. D’Amico. Pose Esti-
mation for Non-Cooperative Spacecraft Rendezvous
using Convolutional Neural Networks. In IEEE
Aerospace Conference, Big Sky, MT, USA, 2018.
21. S. Sharma and S. D’Amico. Comparative Assess-
ment of Techniques for Initial Pose Estimation Using
Monocular Vision. Acta Astronautica, 123:435–445,
22. S. Sharma and S. D’Amico. Pose Estimation for
Non-Cooperative Spacecraft Rendezvous using Neu-
ral Networks. In 29th AAS/AIAA Space Flight Me-
chanics Meeting, Ka’anapali, HI, USA, 2019.
23. S. Sharma, J. Ventura, and S. D’Amico. Ro-
bust Model-Based Monocular Pose Initialization for
Noncooperative Spacecraft Rendezvous. Journal of
Spacecraft and Rockets, 55(6):1–16, 2018.
24. J.F. Shi, S. Ulrich, and S. Ruel. CubeSat Simu-
lation and Detection using Monocular Camera Im-
ages and Convolutional Neural Networks. In 2018
AIAA Guidance, Navigation, and Control Confer-
ence, Kissimmee, FL, USA, 2018.
25. S. Sonawani, R. Alimo, R. Detry, D. Jeong, A. Hess,
and H. Ben Amor. Assistive relative pose estimation
for on-orbit assembly using convolutional neural net-
works. In AIAA Scitech 2020 Forum, Orlando, FL,
USA, 2020.
26. K. Sun, B. Xiao, D. Liu, and J. Wang. Deep high-
resolution representation learning for human pose es-
timation. In 2019 IEEE Conference on Computer
Vision and Pattern Recognition, Long Beach, CA,
USA, 2019.
27. A. Tatsch, N. Fitz-Coy, and S. Gladun. On-orbit Ser-
vicing: A brief survey. In Proceedings of the 2006
Performance Metrics for Intelligent Systems Work-
shop, pages 21–23, 2006.
28. A. Valmorbida, M. Mazzucato, and M. Pertile. Cal-
ibration procedures of a vision-based system for rel-
ative motion estimation between satellites flying in
proximity. Measurement, 151, 2020.
29. M. Wieser, H. Richard, G. Hausmann, J-C. Meyer,
S. Jaekel, M. Lavagna, and R. Biesbroek. e.deorbit
mission: OHB debris removal concepts. In ASTRA
2015-13th Symposium on Advanced Space Technolo-
gies in Robotics and Automation, Noordwijk, The
Netherlands, 2015.
30. M. Wilde, C. Clark, and M. Romano. Historical sur-
vey of kinematic and dynamic spacecraft simulators
for laboratory experimentation of on-orbit proxim-
ity maneuvers. Progress in Aerospace Sciences, 110,
31. M. Zwick, I. Huertas, L. Gerdes, and G. Ortega.
ORGL - ESA’s test facility for approach and contact
operations in orbital and planetary environments. In
International Symposium on Artificial Intelligence,
Robotics and Automation in Space, Madrid, Spain,
Conference Paper
Full-text available
Autonomous Optical Navigation is essential for the proximity operations of space missions to asteroids that usually have irregular gravity fields. One core component of this navigation strategy is the Image Processing (IP) algorithm that extracts optical observables from images captured by the spacecraft's on-board camera. Among these observables, the centroid of the asteroid is important to determine the position between the spacecraft and the body, which is the focus of this research. However, the performance of standard IP algorithms is affected and constrained by the features of the images, such as the shape of the asteroid, the illumination conditions and the presence of additional bodies, therefore, the quality of the extracted optical observables is influenced. To address the latter two challenges, this paper develops a Convolutional Neural Networks (CNN)-based IP algorithm and applies it to the Early Characterization Phase (ECP) of the European Space Agency's HERA mission with the target body of binary asteroid Didymos. This algorithm is capable of estimating the centroid of the primary body successfully with high accuracy and without being affected by the presence of the secondary body or the illumination in the input images. In addition, it can also estimate the centroid of the secondary body when the two bodies are in the same image, which increases the robustness of the overall navigation strategy.
Full-text available
The relative pose estimation of an inactive spacecraft by an active servicer spacecraft is a critical task in the design of current and planned space missions, due to its relevance for close-proximity operations, such as In-Orbit Servicing and Active Debris Removal. This paper introduces a novel framework to enable robust monocular pose estimation for close-proximity operations around an uncooperative spacecraft, which combines a Convolutional Neural Network (CNN) for feature detection with a Covariant Efficient Procrustes Perspective-n-Points (CEPPnP) solver and a Multiplicative Extended Kalman Filter (MEKF). The performance of the proposed method is evaluated at different levels of the pose estimation system. A Single-stack Hourglass CNN is proposed for the feature detection step in order to decrease the computational load of the Image Processing (IP), and its accuracy is compared to the standard, more complex High-Resolution Net (HRNet). Subsequently, heatmaps-derived covariance matrices are included in the CEPPnP solver to assess the pose estimation accuracy prior to the navigation filter. This is done in order to support the performance evaluation of the proposed tightly-coupled approach against a loosely-coupled approach, in which the detected features are converted into pseudomeasurements of the relative pose prior to the filter. The performance results of the proposed system indicate that a tightly-coupled approach can guarantee an advantageous coupling between the rotational and translational states within the filter, whilst reflecting a representative measurements covariance. This suggest a promising scheme to cope with the challenging demand for robust navigation in close-proximity scenarios. Synthetic 2D images of the European Space Agency’s Envisat spacecraft are used to generate datasets for training, validation and testing of the CNN. Likewise, the images are used to recreate a representative close-proximity scenario for the validation of the proposed filter.
Full-text available
Reliable pose estimation of uncooperative satellites is a key technology for enabling future on-orbit servicing and debris removal missions. The Kelvins Satellite Pose Estimation Challenge aims at evaluating and comparing monocular visionbased approaches and pushing the state-of-the-art on this problem. This work is based on the Satellite Pose Estimation Dataset, the first publicly available machine learning set of synthetic and real spacecraft imageries. The choice of dataset reflects one of the unique challenges associated with spaceborne computer vision tasks, namely the lack of spaceborne images to train and validate the developed algorithms. This work briefly reviews the basic properties and the collection process of the dataset which was made publicly available. The competition design, including the definition of performance metrics and the adopted testbed, is also discussed. The main contribution of this paper is the analysis of the submissions of the 48 competitors, which compares the performance of different approaches and uncovers what factors make the satellite pose estimation problem especially challenging.
Full-text available
The relative pose estimation of an inactive target by an active servicer spacecraft is a critical task in the design of current and planned space missions, due to its relevance for close-proximity operations, i.e. the rendezvous with a space debris and/or in-orbit servicing. Pose estimation systems based solely on a monocular camera are recently becoming an attractive alternative to systems based on active sensors or stereo cameras, due to their reduced mass, power consumption and system complexity. In this framework, a review of the robustness and applicability of monocular systems for the pose estimation of an uncooperative spacecraft is provided. Special focus is put on the advantages of multispectral monocular systems as well as on the improved robustness of novel image processing schemes and pose estimation solvers. The limitations and drawbacks of the validation of current pose estimation schemes with synthetic images are further discussed, together with the critical trade-offs for the selection of visual-based navigation filters. The state-of-the-art techniques are analyzed in order to provide an insight into the limitations involved under adverse illumination and orbit scenarios, high image contrast, background noise, and low signal-to-noise ratio, which characterize actual space imagery, and which could jeopardize the image processing algorithms and affect the pose estimation accuracy as well as the navigation filter's robustness. Specifically, a comparative assessment of current solutions is given at different levels of the pose estimation process, in order to bring a novel and broad perspective as compared to previous works.
Relative Measurements systems represent a key technology for next generation space missions that require proximity operations between satellites. Before on-orbit validation, the sensors and algorithms need to be validated in laboratory employing a good fiducial reference of the relative motion and specific calibration procedures concerning the estimation of roto-translation matrices between different reference frames. This paper presents a set of calibration procedures that allow to assess the accuracy in estimating the relative pose between a Target spacecraft, equipped with a set of square markers, and an Inspector satellite moving in proximity and hosting a monocular camera. An external Motion Capture system is used to track the motion of a set of spherical markers attached to both the Target and the Inspector, providing a reliable fiducial reference for the relative pose between the two spacecraft. The proposed calibration procedures were tested using the SPARTANS hardware facility of the University of Padova.
Hardware-in-the-loop simulation and test has been an essential part in the development of spacecraft formation flight, rendezvous, capture/docking, and spacecraft robotics systems since the Gemini project. The need to recreate the kinematics and/or dynamics of spacecraft motion on the ground has led to numerous simulators and testbeds at academic institutions, government facilities, and industry laboratories. The simulation facilities range from small air-bearing tables at universities to building-sized simulators for full-sized systems tests at NASA centers. For the first time to the best knowledge of the authors, the paper presents a systematic classification of spacecraft maneuver simulators into kinematic, dynamic, hybrid, and kino-dynamic systems and discusses the design alternatives, trade-offs and limitations to be considered for each type of simulator. The paper also lists current and past systems reported in literature, along with their primary characteristics. It is thus complementary to existing literature focused on air-bearing spacecraft attitude simulators and air-bearing maneuver simulators. The goal of the paper is to inform designers of new facilities of the current state of the art and the existing experience in the field, and to inform spacecraft developers of existing testbeds in order for them to be able to plan test and experiment campaigns.
This work addresses the design and validation of a robust monocular vision-based pose initialization architecture for close-range onorbit-servicing and formation-flying applications. The aim is to rapidly determine the pose of a passive space resident object using its known three-dimensional wireframe model and a single low-resolution two-dimensional image collected on board the servicer spacecraft. In contrast to previous works, the proposed architecture is onboard executable and capable of estimating the pose of the client without the use of fiducial markers and without any a priori range measurements or state information. A novel feature detection method based on the filtering of the weak image gradients is used to identify the true edges of the client in the image, even in presence of the Earth in background. The detected features are synthesized using simple geometric constraints to dramatically reduce the search space of the feature correspondence problem, which is solved using the EPnP method. This approach is proven to be an order of magnitude faster than the state-of-the-art random sample consensus methods. A fast Newton–Raphson method that minimizes the fit error between corresponding image and model features is employed to refine the pose estimate and to resolve pose ambiguity. The proposed methodology is tested using actual space imagery collected during the PRISMA mission at about a 700 km altitude and a 10 m interspacecraft separation.