Академический Документы
Профессиональный Документы
Культура Документы
# 2005. The American Astronomical Society. All rights reserved. Printed in U.S.A.
and
1
ABSTRACT
Matched filter (MF) processing has been shown to provide significant performance gains when processing stellar
imagery used for asteroid detection, recovery, and tracking. This includes extending detection ranges to fainter
magnitudes at the noise limit of the imagery and operating in dense cluttered star fields as encountered at low Galactic
latitudes. The MF software has been shown to detect 40% more asteroids in high-quality Spacewatch imagery relative
to the currently implemented approaches, which are based on moving target indicator (MTI) algorithms. In addition,
MF detections were made in dense star fields and in situations in which the asteroid was collocated with a star in an
image frame, cases in which the MTI algorithms failed. Thus, using legacy sensors and optics, improved detection
sensitivity is achievable by simply upgrading the image-processing stream. This in turn permits surveys of the nearEarth asteroid (NEA) population farther from opposition, for smaller sizes, and in directions previously inaccessible
to current NEA search programs. A software package has been developed and made available on the NASA data
services Web site that can be used for asteroid detection and recovery operations utilizing the enhanced performance
capabilities of MF processing.
Key words: methods: data analysis minor planets, asteroids techniques: image processing
1. INTRODUCTION
1
Current address: Department of Physics, United States Naval Academy,
Annapolis, MD 21401.
2
Current address: Lawrence Berkeley National Laboratory and Department
of Earth and Planetary Sciences, University of California, Berkeley, CA 94720.
1951
1952
Vol. 130
To form the MF, a spacetime template T(k) of the signal signature must be hypothesized. This can include the effects of motion
within a frames collection period, motion between frames, and
the signal intensity fluctuations. For an unresolved, constantvelocity, uniform-intensity target this can be formulated as a
single pixel set to a signal value in a given frame with a linear
spatial displacement occurring temporally across frames. This is
the simplest model that is applicable to asteroid searches and one
that works well for NEA detection. The signal template must also
be whitened by the same noise covariance used previously, thus
producing R1/2T(k). Dropping the temporal index k, the generalized MF output can then be simply written with the full
covariance inverse as
M T H R1 V hV i;
No. 4, 2005
M
> :
T H R1 T
M 2
> ;
Var(M )
1953
1954
Vol. 130
or in a space-based search system, which induces parallax effects due to the satellites motion around the Earth. SALTAD is
currently implemented with only a linear motion hypothesis
but could be upgraded to include more complicated motion
effects.
Given the linear motion template and assuming an unresolved
point target, the MF formulation simply reduces to a shift and
stack operation in the spacetime domain. That is, each frame is
simply shifted in the two spatial dimensions relative to the first
frame by an amount corresponding to the hypothesized motion
shift for that frame. Each shifted frame is then added to a running
sum frame that produces the MF output of equation (2). Note that
the frames used are the whitened-PSF-convolved data set. To
actually shift the frames, another interpolation is required to remap the imagery at each time step. For this stage the algorithms
used were nearest neighbor and bilinear interpolation. Half-pixel
step sizes for the lowest speed hypothesis were found to produce
the greatest detection performance, as seen in Figure 1, but increasing this to full pixel steps resulted in a loss of only 10% in
the numbers of asteroids found in the imagery. Full-pixel shifts
are desirable from a computational standpoint since the shifting
and stacking operation involves no interpolation (simple address
shift to align frames) and is thus a very significant run-time
advantage. Note that subpixel shifting with bilinear interpolation
is an option in the SALTAD package if run-time performance is
not a limiting factor.
An alternative to spatial shifting and stacking is to do spatial Fourier processing on the sequence of images (take twodimensional fast Fourier transforms of each image). The linear
motion of an object in Fourier space appears as a plane of energy
at a different angle relative to the background clutter. Applying
the motion hypothesis template in the Fourier domain is a simple matter of taking the point-by-point product of the Fouriertransformed hypothesis template with the Fourier-transformed
images and inverse transforming, getting the integrated signal
energy for all asteroid starting positions. For a limited hypothesis
set and large dimensional spatial imagery, this can be a computational advantage. At this time, however, SALTAD does not
support this mode of processing.
Given the MF output, a first-pass test for detection is made
using the maximum likelihood detector form of equation (5). In
this case a CFAR detector is used, where a donut region around
the pixel under test is used to form a variance estimate locally.
The MF output of the pixel squared is then tested for whether it
exceeds a user-defined factor applied to the variance. If this threshold is passed, the robust detection parameter of equation (3) is
No. 4, 2005
Fig. 2.Detections and false alarms for the two detection statistics TRT and
MLE on a sample Spacewatch data set. Crosses show false alarms, filled circles
show new MF detections, and open circles show MF detections that were also
MTI detections.
1955
1956
Vol. 130
high star clutter. The following subsections discuss the first results from that analysis.
4.1. Detection Enhancement in Limiting Magnitude
In both the first-year report to NASA by Gural (2003) and the
latest performance analysis results generated by the Spacewatch
project, the SALTAD MF algorithms have demonstrated an
ability to detect a larger number of asteroids at the fainter limits
of the telescopic collection system. In a composite result of several processing scans, comparing the MF with MTI algorithms
on identical data sets, the cumulative detection of asteroids as a
function of magnitude was generated. The software used in the
comparison was the latest incarnation of the Spacewatch MTI
algorithms embodied in the software package IMPACT. For the
MF results, one of the authors (A. G.) ran the SALTAD software,
and final detection lists were screened for false alarms by visual
verification. In Figure 3 can be seen the results of several such
comparisons. Note the dramatic rise in the number of asteroids
discovered in the imagery by the MF algorithms starting near
21 mag. In total, 44% more asteroids were located in this set of
scans by the SALTAD software at a 50% false-alarm rate. By
placing the detections in histogram bins of 0.5 mag, as in Figure 4, it can be seen that the MF extends the limiting magnitude
of detection.
An exact measure of the degree of improvement is still
forthcoming but has been estimated to be nearly 1 full mag. In
fact, the effective increase in detection is nearly 100% between
Fig. 5.Evolution of true detections and false alarms for a single data set.
Pushing the limits to fainter asteroids shows an increasing false-alarm rate for
each new detection (flattening of the curve).
No. 4, 2005
1957
Fig. 8.Detection signal strength vs. the combined (eq. [6]) detection statistic, where black circles show detections and gray circles show false alarms.
Fig. 6.Detection signal strength vs. the MLE detection statistic, where black
circles show detections and gray circles show false alarms.
Fig. 7.Detection signal strength vs. the TRT detection statistic, where black
circles show detections and gray circles show false alarms.
and 8 respectively show how the asteroid signal strength for detections is distributed among the false alarms for the MLE-only,
TRT-only, and combined figure of merit for equation (6), where
27,000 candidates from one scan are combined into a single display. The goal is to maximize separation of the false alarms (gray
symbols) from the true detections (black symbols), since a threshold must be drawn across this distribution. Without good separation, accepting more detections by moving a threshold line results
in higher numbers of false alarms. For example, in the MLE-only
and TRT-only cases, setting a threshold (vertical line) such that all
true detections fall to the right of the threshold inadvertently also
accepts a large number of false alarms. However, greater separation of detections from false alarms is possible when one uses
some combined cost function as seen in Figure 8, in which the
thresholding vertical line avoids a large portion of the false-alarm
region of the plot. Choosing an improved detection statistic is an
area that needs to be explored further, using perhaps other outputs
from the processing. Figure 9 is similar to Figure 2 in showing
a clearer separation between true asteroids and false alarms by
working in a combined MLE and TRT space rather than in either
parameter plotted separately against signal strength.
It should be noted that the NEAT project has also collected and
archived a large sample of asteroid imagery and had supplied
Fig. 9.Separation of true detections (black circles) from false alarms (gray
circles) in TRT/MLE space.
1958
Fig. 10.Early Spacewatch detection in a cluttered star field using the SALTAD MF processing. The asteroid is located in the white box.
SAIC with an image set just prior to the completion of this project. The data set processed was collected on 2001 August 1 and
consisted of three frames of 4K ; 4K imagery plus the detection
list of asteroids found by the NEAT processing system software.
After processing was completed, the SALTAD software identified 26 of the 26 previously known asteroids with the addition of
five new detections, yielding a 20% improvement in detection
performance. The new detections were typically of asteroids in
the fainter range of magnitudes previously discovered by NEAT,
and the processing yielded very few false alarms. Once again, it
was found that image artifacts, in this case an image quadrant
filled with numerous hot pixels in the first collection frame, were
the cause for reduced detection sensitivity in the quarter of the
image space where they existed. However, in good-quality regions of the collected imagery, the software demonstrates improved performance on a Spacewatch-independent collection
system and also emphasizes the preprocessing removal of image
artifacts as a requirement for reliable MF processing.
4.2. Performance in Cluttered Star Fields
One claim for the MF image-processing procedure is that it
allows one to detect moving objects even in heavily cluttered
backgrounds. The gain is primarily due to the star removal procedure in the clutter suppression step prior to matched filtering.
For example, the LINEAR stream follows a similar approach for
clutter suppression that has permitted them to make observations
near the Milky Way. For an asteroid search in a stellar field, the
term heavily cluttered corresponds to dense star regions. An
MTI technique that searches for stars that appear in common
across multiple image frames will exclude an extensive portion
of the dense star field and eliminate a large percentage of asteroid
tracks due to their close proximity to the myriads of stars. The
MF clutter-suppression procedure outlined here attempts to remove the stationary background objects before searching for
moving objects of interest, followed by an enhancement of signal energy through temporal integration. The combined effects
allow for a significant performance improvement in cluttered
fields. First results of asteroid discoveries for which the asteroid
was either near or superposed on stars in the field are shown in
Figures 10 and 11. In each case the MTI technique would have
failed to detect the asteroid motion, and in many cases even very
bright asteroids would have been missed.
A few dense star-field regions were also processed by the
Spacewatch analysts using both the IMPACT and SALTAD software, although this was difficult, as archived Spacewatch imagery typically avoids low Galactic latitudes. In one analyzed
scan, the MF detector identified nine asteroids, whereas the MTI
Fig. 11.Examples of asteroid detections missed by the MTI technique (due to common star/galaxy exclusion regions) yet detected by the MF. Images are displayed
using a three-color multiframe display technique (i.e., frame 1 is red, frame 2 is green, and frame 3 is blue; thus, stationary objects appear gray).
Fig. 12.Example of an asteroid recovery reacquisition of Apollo 1998 VD35 (within the white box).
1960
REFERENCES
Auerbach, S. P., Hauser, L. E., Boynton, F. P., Janda, R. S., & Sofianos, D. J.
Porat, B., & Friedlander, B. 1990, IEEE Trans. Pattern Anal. Machine Intell.,
1996, Proc. SPIE, 2759, 25
12, 398
Barniv, Y. 1985, IEEE Trans. Aerosp. Electron. Syst., 21, 144
Rabinowitz, D. L. 1991, AJ, 101, 1518
Barniv, Y., & Kella, O. 1987, IEEE Trans. Aerosp. Electron. Syst., 23, 776
Ralston, K., et al. 1996, Airborne InfraRed Measurement System (AIRMS)
Chen, Y. 1989, IEEE Trans. Aerosp. Electron. Syst., 25, 343
Final Report (Arlington: DARPA)
Gural, P. S. 2003, Asteroid Search with Advanced Detection Algorithms
Reed, I. S., Gagliardi, R. M., & Stotts, L. B. 1988, IEEE Trans. Aerosp.
( NASA Tech. Rep. 20040021361; Hanover: NASA)
Electron. Syst., 24, 327
Jedicke, R., Morbidelli, A., Spahr, T., Petit, J., & Bottke, W. F. 2003, Icarus,
Sanders-Reed, J. N. 1998, IEEE Trans. Aerosp. Electron. Syst., 34, 844
161, 17
Stokes, G. H., Evans, J. B., Viggh, H. E. M., Shelly, F. C., & Pearce, E. C.
Kelly, E. J. 1986, IEEE Trans. Aerosp. Electron. Syst., 22, 115
2000, Icarus, 148, 21
Lampropoulos, G. A., & Boulter, J. E. 1987, Soc. Photo-Opt. Instrum. Eng.,
Viggh, H. E. M., Stokes, G. H., Shelly, F. C., Blythe, M. S., & Stuart, J. S. 1998, in
3163, 138
Proc. Sixth Int. Conf. and Exposition on Engineering, Construction, and
Mohanty, N. C. 1981, IEEE Trans. Pattern Anal. Machine Intell., 3, 606
Operations in Space, ed. R. G. Galloway & S. Lokaj ( New York: Society), 373
Pohlig, S. C. 1989, IEEE Trans. Aerosp. Electron. Syst., 25, 56
Watson, G. H., & Watson, S. K. 1997, Proc. SPIE, 3163, 45
. 1992, Maximum Likelihood Detection of Electro-optic Moving Targets
( Tech. Rep. 940; Lexington: MIT), 37