Autonomous On-board Near Earth Object Detection
P. Rajan P. Burlina
Dept. of Computer Science Applied Physics Laboratory
Johns Hopkins University Johns Hopkins University
purnima@cs.jhu.edu
B. Jedynak N. Mehta
Department of Mathematics Applied Physics Laboratory
and Johns Hopkins University
Statistics
Portland State University
Abstract—Most large asteroid population discovery has been
accomplished to date by Earth-based telescopes. It is speculated
that most of the smaller Near Earth Objects (NEOs) that are less
than 100 meters in diameter, whose impact can create substantial
city-size damage, have not yet been discovered. Many asteroids
cannot be detected with an Earth-based telescope given their size
and/or their location with respect to the Sun. We are investigating
the feasibility of deploying asteroid detection algorithms on-
board a spacecraft, thereby minimizing the expense and need
to downlink large collection of images. Having autonomous on-
board image analysis algorithms enables the deployment of a
spacecraft at approximately 0.7 AU heliocentric or Earth-Sun
L1/L2 halo orbits, removing some of the challenges associated
with detecting asteroids with Earth-based telescopes. We describe
an image analysis algorithmic pipeline developed and targeted
for on-board asteroid detection and show that its performance
is consistent with deployment on flight-qualified hardware.
I. I NTRODUCTION
NASA has a congressional mandate to discover all Near-
Earth Objects (NEOs) at least 1 kilometer in diameter. For-
tunately, 95% of the NEOs larger than 1 km that have been
discovered are likely not to impact Earth. Near-Earth Object
search programs [1] are currently almost exclusively accom-
plished by Earth-based telescopes such as MIT’s LINEAR [2]
project, the NEAT [3] program, the Catalina Sky Survey [4],
or Pan-STARRS [5]. An exception is JPL’s spacecraft-based
WISE telescope brought out of hibernation to characterize
NEOs in the 3.4 and 4.6 micron infrared bands [6] (called
NEOWISE).
It is notable, however, that the NEO that impacted
Chelyabinsk, Russia on 15 February 2013 was only about 17
meters in diameter. It is estimated that only a relatively small
fraction of those so called “city-killing” asteroids, particularly
objects less than 100 meters in diameter, have been discovered
to date [7]. Because of their size, atmospheric effects and their
location with respect to the Sun, some of these NEOs cannot
easily be detected with an Earth-based telescope.
Our focus here is therefore on developing an algorithm
that can be hosted on-board a spacecraft. The two nominal
978-1-4673-9558-8/15/$31.00 ©2015 IEEE
M. Chen D. Edell
Applied Physics Laboratory Applied Physics Laboratory
Johns Hopkins University Johns Hopkins University
A. Sinha G. Hager
Dept. of Electrical Dept. of Computer Science
and Johns Hopkins University
Computer Engineering
Johns Hopkins University
approaches that we are investigating include deploying the
algorithm on a BAE RAD750 with 8 MB of EEPROM and
32 MB RAM in addition to Virtex 5 FPGAs. Understanding
that there are processing and memory constraints on-board,
our algorithm design needs to not only meet the performance
objectives of detecting and identifying asteroids, but its imple-
mentation needs to be optimized for the respective platforms.
This paper describes an agile algorithm candidate that is being
investigated as well as our use of representative real and
simulation imagery for testing it.
Notable prior work includes [5], describing a reference
processing system for detection and identification of asteroids
named the Pan-STARRS Moving Object Processing System
(MOPS). This pipeline identifies moving objects in our solar
system and links those detections within and between night ob-
servations. It attributes detections to known objects, calculates
initial and differentially corrected orbits for linked detections,
recovers detections when they exist, and performs orbit iden-
tification. Most proposed pipelines for Earth-based detection
include a step to combine images into a high signal to noise
ratio (SNR) static-sky image that is subtracted from the current
master image to obtain a difference image containing only
transient sources. Examples include [8], where a shift-and-add
technique is used to improve the SNR and then synthetically
creating long exposure images to facilitate the detection of
trajectories. A related shift-and-add method using a median
image rather than an average image is reported in [9]. A match
filter is used for asteroid detection and matching in [10]. In
[11], [12], [13], tree based searches (including KD-trees) are
used for efficient linking of successive asteroid detection and
finding sets of observation points that can be fitted with an
inherent motion model, through an exhaustive search for all
possible linkages that satisfy the model constraints.
After careful evaluation, it was found that none of these
pipelines were suitable for an on-board system . We also
evaluated alternative approaches for efficient line detection.
The Hough transform, see [14], allows to search for 3-
dimensional lines in parametric space. However, our exper-
iments, in accordance with [12], showed that the algorithm
is sensitive to the discretization of the parameters unless a
very fine grid is used, but in this case the memory and
computational cost requirements are prohibitive. RANSAC
(Random Sample Consensus), see [15], is a generic algorithm
for identifying parametric shapes. In our case, however, the
randomization is not appropriate since the image is typically
dominated by noisy structures. In [16], the authors describe a
framework for multiscale analysis, named beamlet analysis, in
which line segments play a role analogue to the role played
by points in wavelets analysis. However, the resulting pipeline
was not efficient enough and fairly complex. In [17], the
authors use a generic discrete optimization algorithm: the
branch and bound algorithm for optimizing among sub-images
and perform efficient object localization. Since line segments
and sub-images are both identified by two pixels, we have
explored modifications of the algorithm in [17] adapted to
line segments. However, if the average computation time and
memory requirements are competitive, the worst case is not in
the range of performance compatible with the requirements of
an on-board system. As a consequence we have not pursued
this approach.
When compared to prior studies, novel and salient con-
tributions of our work include: (1) a simple and efficient
algorithmic pipeline that can be deployed on flight-qualified
hardware; (2) a simulation engine relying on principled optics
models allowing to generate realistic space-based imagery.
II. A PPROACH
The main components of our image processing pipeline
addresses detection and linking of asteroids and takes as
input a sequence of 3 or 4 time-lapse images acquired from
space-based platforms, as described below (also detailed in
Algorithm 1). The pipeline is a bottom-up brute force search.
The first 3 stages: Image Pre-Processing, Image Registra-
tion and Image Logical Differencing allow for identifying
movers, effectively removing the background noise as well
as the astronomical objects that are not moving. The final
stage: Trajectory Linking allows to explore systematically
all the rectilinear trajectories connecting the movers along the
sequence of images.
Image Pre-Processing Median filtering is applied to reduce
impulsive noise and artifacts that depend on the acquisition.
Image Registration Image registration is used to bring the
images into alignment with a common image of reference so
that the stars in the background line up in all images. In
the case of a triplet or a quadruplet image set, the second
image is used as reference. We use a similarity transformation
(translation, rotation, scaling) and/or skew (full affine trans-
formation), depending on the image acquisition protocol, to
align the images in the sequence. Our registration uses mutual
information [18] to estimate the transformation parameters.
Image Logical Differencing A global thresholding is
applied to the registered image for detecting asteroids and
suppressing background noise. As asteroids are typically very
faint compared to the surrounding stars, the selection of the
detection threshold impacts false alarm rate. Thresholding
yields binary detection images. The set of all binary images
is then used to generate an intersection image that contains
above threshold pixels that occur in at least two images
in the sequence. This is followed by logical differencing
whereby we produce a set of difference images by intersecting
the corresponding binary image with the logical complement
of the common intersection image. This operation provides
a list of candidate detections (movers) for each image in
the sequence. While the logical differencing results in good
detections, additional artifacts such as crater-like formations
(see Fig. 9) are seen as a result of some celestial bodies
being over-exposed. To mitigate this artifact, we find the
connected components (using 8-neighborhood connectivity) in
each difference image and filter out hollow objects based on
connected component size and extent, and store the centroids
of the selected components for the next stage of the pipeline.
Trajectory Linking The list of centroids of moving objects
obtained from image differencing defines a set of candidate
rectilinear trajectories. The goal is to find a subset of centroids
that fit a linear model. The set of filtered centroids potentially
has a high number of noisy points (falsely detected movers),
and the cardinality of the set of all candidate trajectories
increases exponentially with the number of detections, thus
requiring subsequent pruning. Given a sequence of images,
we find all the possible trajectories connecting the detections
between all pairs of images. There are k2 pairs of end images,


where k is the number of images in the sequence, k = 3 or
k = 4 in the experiments, and there are O(k 2 n2 ) possible
trajectories, where n is the average number of detections per
image. We arrange the images along a cuboid, where the
z axis is the time of image acquisition. We then find the
points of intersection of each of these trajectories with all the
remaining images. Here we assume that the velocity of the
asteroid is constant within the sequence. If there is at least one
detection in a 5x5 window around the point of intersection,
we increment the vote for that candidate trajectory. If the vote
exceeds a line confidence threshold, the trajectory is marked
as valid. This arrangement allows for missing detections in
some of the images in the sequence. This is pertinent as
the asteroids spin on their axis and their brightness can vary
drastically from one image to the next. This is especially true
for small asteroids that this mission is intended to discover.
Unlike MOPS [5] and CSS[4], we do not set an upper limit
on the velocity of the asteroid, and hence do not risk missing
fast movers. Finally, we reduce false detections using Support
Vector Machine (SVM) based classification.
III. E XPERIMENTS
We detail the experiments performed using a range of
simulated space-based imagery generated using the JHU/APL
developed Renderer and Camera Emulator (RCE) as well as
real imagery from the NEAT and Catalina Sky Survey dataset.
Fig. 1. (Left) An image simulated by RCE. (Right) 31 simulated MWIR images super-imposed in order to visualize the trajectory of the asteroid in a single
image. The true trajectory can be seen as a faint line towards the top center of the image.

Fig. 2. The trajectories found by the pipeline are shown in green. The true location of the asteroid is marked in red. (Left) Trajectory Detection on a simulated
triplet. (Right) Trajectory Detection on a quadruplet. Adding one more image to the triplet eliminates the false positives.
Algorithm 1 Detection and Linking Algorithm for a sequence
of images
1: Apply a median filter to each image in the sequence.
2: Align all the images to a reference image (e.g. the second image
for a triplet) using Mutual Information based registration.
3: Perform thresholding on the registered images using a pre-defined
threshold to generate binary detection images.
4: Compute the logical (soft) intersection image from the binary
detection images. A pixel has value one if and only if at least 2
detection images have value 1 at the same location.
5: Compute the logical differencing between each binary detection
image and the intersection image.
6: Find the connected components in the difference images. Replace
each connected component by its center of gravity.
For each of the k2 pairs of images in the sequence, compute


7:
the set of all trajectories connecting the candidate moving object
detections from each pair of end images.
8: For each trajectory do
• Find the points of intersection of each trajectory with the
rest of the images.
• Check for detections within a 5x5 neighborhood of the point
of intersection in each image. Increment the line vote by 1
if there are detections.
• Mark the trajectory as a valid line if the vote is higher than
a preset line threshold.
A. Simulated Space-Based Imagery
Using RCE, asteroids are modeled as spherical blackbody-
like emitters, with a cross-sectional area that approximates the
sizes of actual asteroids and surface temperatures typical of
Sun-illuminated asteroids in an Earth-like orbit. Similar to the
way stars are modeled, the radiation emitted is modeled using
Planck’s equation.
2hc2 1
Bλ (T ) = ǫ 5 hc (1)
λ e λkB T − 1
where Bλ (T ) is the spectral radiance at a given wavelength λ
and temperature T The value ǫ is the emissivity of the asteroid,
which essentially converts the blackbody spectral radiance into
spectral irradiance. The constant, h is the Planck’s constant,
c is the speed of light, λ is wavelength, kB is the Boltzmann
constant, and T is the temperature.
The asteroids are assumed to have a nominal temperature
of 200 K due to solar heating and emissivities in the range
from 0.9 to 0.98. The RCE uses stellar data available as part
of the Two Micron All Sky Survey (2MASS) [19], a stellar
survey that scanned the entire sky in three IR bands (centered
at 1.25 µm, 1.65 µm, and 2.17 µm, respectively). The 2MASS
catalog also incorporates data in two visible bands from
other surveys. An example of the simulated Mid-wavelength
Infrared (MWIR) image and the ground truth trajectory derived
from one set of simulated imagery is shown in Fig. 1. Fig. 2
shows the final trajectories detected by our algorithm for one
triplet and one quadruplet of the simulated MWIR dataset.
As is shown in Fig. 2, using trajectory verification on a
greater number of images in the sequence allows us to quickly
disambiguate and reject false trajectories. In this case this trend
is readily apparent when going from a triplet to a quadruplet
of images.
When testing on a set of 2100 triplets or quadruplets or
quintuplets taken from simulated space-based imagery, we
characterize the algorithm with regard to detections at each
of the following three successive stages: the initial detection
of objects, the detection of moving objects and the detection of
trajectories. In Fig. 3, we plot the number of detected objects at
each step of the pipeline for each selected threshold value. The
right plot in Fig. 3 shows that the use of quadruplets allows for
a single trajectory to be found irrespective of the number of
detections found at prior stages, echoing the results displayed
in Fig. 2. The false positive rate was completely abated in this
case.
Fig. 4 shows the True positive vs False positive curve
for asteroid detections at the final stage of the pipeline for
simulated image sequences generated with integration time of
90 seconds and aperture size of 0.5 m. The various points
along the curves were obtained by varying the threshold in
the step 3: of the Algorithm 1. The size of the asteroids in the
simulated dataset ranges from 30 meters to 1000 meters. Fig. 5
shows the same information stratified by asteroid radius and
SNR(signal to noise ratio). The larger asteroids are always
detected. Among the smaller ones (<100 meters), some are
undetected. SNR for an image sequence is computed as the
ratio of the median of the maximum pixel value on the
asteroid in each image in the sequence to the median of the
trimmed(upper 10%) mean noise level in each image. Note that
the SNR is a good indicator of the overall performance as all
asteroids with sequence SNR larger than 1.3 were detected.
Fig. 7 depicts the detection percentage as a function of SNR.
The plot on the right of Fig. 7 shows how the SNR varies
based on asteroid radius and distance.
Finally in Fig. 8 we show the True positive vs False positive
curves for various stages of the pipeline. This figure portrays
how the false detections reduce as we move through the
pipeline.
B. False alarm mitigation using Machine Learning
In order to reduce the number of false trajectories, we
experimented using Support Vector Machines(SVM) that dif-
ferentiate between “good” vs. “bad” image sequences. The
feature vector for classification was generated by taking a
5 × 5 window around each detection in the images to form
a vector of length 100 for a quadruplet(75 for a triplet). The
asteroid ground truth was used to generate a 0/1 label for each
sequence. Tables I and II show the 10-fold cross-validation
results on 180 image sequences generated with integration
time of 90 seconds and aperture size of 0.5 m.
Kernel Overall True positive False
accuracy(%) accuracy(%) positive
accuracy(%)
linear 99.5 87.07 99.88
chi-squared 99.77 95.87 99.89
rbf 99.25 100 99.23
TABLE I
SVM RESULTS 180 IMAGE QUADRUPLETS WITH LINE THRESHOLD 3.
 Kernel Overall True positive False
accuracy(%) accuracy(%) positive
accuracy(%)
linear 99.87 99.41 100
chi-squared 99.84 99.28 100
rbf 100 100 100
TABLE II
SVM RESULTS 180 IMAGE QUADRUPLETS WITH LINE THRESHOLD 4.
C. Real Imagery
It is important when developing an image processing
pipeline on simulated images to verify that the overall system
performs as expected on real imagery.
1) NEAT: Near Earth Asteroid Tracking (NEAT) [3] is
an Earth-based program run by NASA from 1995-2007 to
discover NEOs. Fig. 9 and Fig. 10 show the results at all
stages of the pipeline for one triplet of images of the 2002-
CY46 asteroid obtained from the NEAT system archive.
2) CATALINA: The Catalina Sky Survey (CSS) [20], [21] is
intended to discover NEOs, specifically potentially hazardous
asteroids that pose risk to Earth. Fig. 10 shows the final result
for one set of Earth-based images from CSS.
D. Implementation
Prototype and test algorithms were coded in MATLAB.
The algorithms that show promise were then implemented
in C++ for benchmarking on a MCP750, a good proxy for
the RAD750. The MCP750, with clock speed of 367 MHz,
is completely dedicated to the algorithm and not servicing
spacecraft or instrument commands during benchmarking. Run
time for Algorithm 1 on a quadruplet is between 1 and 10
seconds. Preliminary tests indicate that 16 MB of RAM is used
and the current executable is 1.3 MB. Any pre-processing such
as step 1 would be implemented on an FPGA. While further
optimization is under way, current results indicate that this
algorithm is a viable candidate for a spacecraft processor.
IV. CONCLUSION
We have developed a pipeline for space-based asteroid de-
tection. Its performance and run time make it a good candidate
for deployment on a RAD750. We demonstrated that the use
of quadruplets of image, instead of triplets is determinant
in detecting faint asteroids. Current run time also supports
a mission concept that aims for image acquisition at a 10
minute cadence with plenty of time for additional processing.
Further experiments and refinements of the pipeline will allow
for better understanding the trade-offs of a space-based system
designed for asteroids that cannot otherwise be detected with
an Earth-based telescope.
ACKNOWLEDGMENT
This work was supported by an Early Stage Innovations
grant from NASA’s Space Technology Research Grants Pro-
gram NNX14AB04G. The authors would like to thank E.
Christensen for providing the CSS data for this study.
R EFERENCES
[1] G. H. Stokes, J. B. Evans, and S. M. Larson, “Near-earth asteroid search
programs,” Asteroids, vol. 3, pp. 45–54, 2002.
[2] J. B. Evans, F. C. Shelly, and G. H. Stokes, “Detection and discovery of
near-earth asteroids by the linear program,” Lincoln Laboratory Journal,
vol. 14, no. 2, pp. 199–215, 2003.
[3] NASA. (2015) The NEAT survey. [Online]. Available:
http://neat.jpl.nasa.gov/
[4] ——. (2014) The Catalina Sky Survey. [Online]. Available:
http://www.lpl.arizona.edu/css/
[5] L. Denneau, R. Jedicke, T. Grav, M. Granvik, J. Kubica, A. Milani,
P. Vereš, R. Wainscoat, D. Chang, F. Pierfederici et al., “The pan-starrs
moving object processing system,” Pan, vol. 125, no. 926, pp. 357–395,
2013.
[6] NASA. (2014) The NEOWISE survey. [Online]. Available:
http://neo.jpl.nasa.gov/programs/neowise.html
[7] ——. (2015) A Near Earth Asteroid Census. [Online]. Available:
http://tinyurl.com/bypjz4d
[8] M. Shao, B. Nemati, C. Zhai, S. G. Turyshev, J. Sandhu, G. Hallinan, and
L. K. Harding, “Finding very small near-earth asteroids using synthetic
tracking,” The Astrophysical Journal, vol. 782, no. 1, p. 1, 2014.
[9] T. Yanagisawa, A. Nakajima, K.-i. Kadota, H. Kurosaki, T. Nakamura,
F. Yoshida, B. Dermawan, and Y. Sato, “Automatic detection algorithm
for small moving objects,” Publications of the Astronomical Society of
Japan, vol. 57, no. 2, pp. 399–408, 2005.
[10] P. S. Gural, J. A. Larsen, and A. E. Gleason, “Matched filter processing
for asteroid detection,” The Astronomical Journal, vol. 130, no. 4, p.
1951, 2005.
[11] J. Kubica, J. Masiero, A. W. Moore, R. Jedicke, and A. Connolly,
“Variable kd-tree algorithms for spatial pattern search and discovery,”
Robotics Institute, p. 253, 2005.
[12] J. Kubica, A. Moore, A. Connolly, and R. Jedicke, “A multiple tree
algorithm for the efficient association of asteroid observations,” in
Proceedings of the eleventh ACM SIGKDD international conference on
Knowledge discovery in data mining. ACM, 2005, pp. 138–146.
[13] J. Kubica, L. Denneau, T. Grav, J. Heasley, R. Jedicke, J. Masiero,
A. Milani, A. Moore, D. Tholen, and R. J. Wainscoat, “Efficient intra-
and inter-night linking of asteroid detections using kd-trees,” Icarus, vol.
189, no. 1, pp. 151–168, 2007.
[14] R. O. Duda and P. E. Hart, “Use of the hough transformation
to detect lines and curves in pictures,” Commun. ACM,
vol. 15, no. 1, pp. 11–15, Jan. 1972. [Online]. Available:
http://doi.acm.org/10.1145/361237.361242
[15] M. A. Fischler and R. C. Bolles, “Random sample
consensus: A paradigm for model fitting with applications to
image analysis and automated cartography,” Commun. ACM,
vol. 24, no. 6, pp. 381–395, Jun. 1981. [Online]. Available:
http://doi.acm.org/10.1145/358669.358692
[16] D. L. Donoho, X. Huo, I. Jermyn, P. Jones, G. Lerman, O. Levi, and
F. Natterer, “Beamlets and multiscale image analysis,” in in Multiscale
and Multiresolution Methods. Springer, 2001, pp. 149–196.
[17] C. H. Lampert, M. Blaschko, and T. Hofmann, “Efficient subwindow
search: A branch and bound framework for object localization,” Pattern
Analysis and Machine Intelligence, IEEE Transactions on, vol. 31,
no. 12, pp. 2129–2142, Dec 2009.
[18] P. Viola and W. M. Wells III, “Alignment by maximization of mutual
information,” International journal of computer vision, vol. 24, no. 2,
pp. 137–154, 1997.
[19] M. F. Skrutskie et al., “The two micron all sky survey (2mass),” The
Astronomical Journal, vol. 131, no. 2, pp. 1163–1183, 2006.
[20] S. Larson, J. Brownlee, C. Hergenrother, and T. Spahr, “The catalina
sky survey for neos,” in Bulletin of the American Astronomical Society,
vol. 30, 1998, p. 1037.
[21] A. Drake, S. Djorgovski, A. Mahabal, E. Beshore, S. Larson, M. Gra-
ham, R. Williams, E. Christensen, M. Catelan, A. Boattini et al., “First
results from the catalina real-time transient survey,” The Astrophysical
Journal, vol. 696, no. 1, p. 870, 2009.
 Performance metrics for simulated triplets. Performance metrics for simulated quadruplets.
6 6
Threshold = 80 Threshold = 90
Threshold = 100 Threshold = 100
5 Threshold = 120 5 Threshold = 120
Log10(Number of detections)

Log (Number of detections)

Threshold = 130 Threshold = 130
4 Threshold = 140 4 Threshold = 140

3 3

2 2

10
1 1

0 0
A B C D A B C D
Stages of the pipeline Stages of the pipeline

Fig. 3. (Left) The number of detections at various stages of the pipeline for triplets of images. (Right) The number of detections at various stages of the
pipeline for quadruplets of images. Following are the stages of the pipeline : (A) Initial (B) Pre-Processing+Registration+Thresholding (C) Image Differencing
+ Connected Components + Filter (D) Trajectory Detection.

Fig. 4. True positive vs False positive curve at various threshold values for asteroid detections at the final stage of the pipeline for simulated image sequences
generated with integration time of 90 seconds and aperture size of 0.5 m. The plots showing the detections for the complete set of asteroids whose size ranges
from 30 m to 1000 m. (Left) 4 images per sequence with line threshold 3.(Right) 5 images per sequence with line threshold 4.
Fig. 5. True positive vs False positive curve at various threshold values for asteroid detections at the final stage of the pipeline for simulated image sequences
generated with integration time of 90 seconds and aperture size of 0.5 m. (Left) The curves are stratified based on the asteroid radius.(Right) The curves are
stratified based on the asteroid SNR.

Fig. 6. True positive vs False positive curve at various threshold values for asteroid detections at the final stage of the pipeline for simulated image sequences
generated with integration time of 90 seconds and aperture size of 0.5 m. The curves are stratified based on the asteroid distance. (Left) Asteroid of radius
50 m (Right) Asteroid of radius 30 m.
Fig. 7. (Left) Detection histogram for SN R ≤ 1.3 for the entire simulated image set of 2100 sequences. We detect all asteroids with SN R > 1.3 (Right)
Plot showing how SNR varies depending on asteroid size and distance.

Fig. 8. True positive vs False positive curve at various threshold values for asteroid detections at various stages of the pipeline for simulated image sequences
generated with integration time of 90 seconds and aperture size of 0.5 m.(Left) 4 images per sequence with line threshold 3.(Right) 5 images per sequence
with line threshold 4.
Fig. 9. Image Processing Pipeline results shown on 2002 NEAT data. (row 1): input image triplet (CY46) taken approx. 10 minutes apart. (row 2): Image
Registration and Thresholding. Left: Image-1 registered to Image-2. Middle: Image-2 after thresholding Right: Image-3 registered to Image-2. (row 3): Image
Differencing: Artifacts such as crater-like formations are seen in the difference images above. This is the result of some celestial bodies being over-exposed.)
(row 4): Image Differencing: Filtered centroids shown in each image of the sequence.)
Fig. 10. Results shown on real imagery from NEAT and CSS. Asteroid trajectory detected is shown in green, true location of the asteroid is in red. The
registered images are super-imposed to visualize the trajectory in a single image. (Left) Trajectory Detection for the NEAT CY46 Triplet. (Center) Trajectories
detected on a CSS triplet. (Right) Trajectory detected on a CSS quadruplet.