satellite detection algorithms developed forcradpdf.drdc-rddc.gc.ca/pdfs/unc108/p534603_a1b.pdf ·...

47
Satellite detection algorithms Satellite detection algorithms Satellite detection algorithms Satellite detection algorithms developed for developed for SofS applications SofS applications Warning: This presentation contains several animations that cannot Martin P. Lévesque RDDC Valcartier P d be executed with the PDF format. Some slide may look like confused because of this Presented to: NATO SET147- Microsatellites and Surveillance of Space July 15th 2010 confused because of this. Defence Research and Development Canada Recherche et développement pour la défense Canada Canada

Upload: nguyentruc

Post on 30-May-2018

222 views

Category:

Documents


0 download

TRANSCRIPT

Satellite detection algorithmsSatellite detection algorithmsSatellite detection algorithmsSatellite detection algorithmsdeveloped for developed for

SofS applicationsSofS applicationsWarning:

This presentation contains several animations that cannot

Martin P. LévesqueRDDC Valcartier

P d

be executed with the PDF format.

Some slide may look like confused because of this Presented to:

NATO SET147- Microsatellites and Surveillance of SpaceJuly 15th 2010

confused because of this.

Defence Research andDevelopment Canada

Recherche et développementpour la défense Canada Canada

PLAN

- SofS context- Acquisition modesAcquisition modes- Processing in SSM- Processing in TRM

-ProcessingIt ti b k d l l ith- Iterative background removal algorithm

- Star detection- Streak detection- Streak detection- Iterative matched filter- Alarm extraction and false alarm rejection

Defence R&D Canada • R & D pour la défense Canada

j

MOLNIYA(40000km)(40000km)

GPS(5000km)

LEO(800km)

GEO(36000km)

Defence R&D Canada • R & D pour la défense Canada

( )

Search Area vs Delay of Observation

(TLE) Prediction accuracy after a too long period:

Position estimation accuracy

after a too long period:... lost object

after a certain period of time:... requires a new acquisition

pointingaccuracy

Defence R&D Canada • R & D pour la défense Canada

TLED t bDatabase

Extractedobsolete TLE

Update:detectedsatellites

Observation cycle Observationplanning and

tasking

Data reduction(image

processing)

satellites

gp g)

Observationschedule

Astronomicalimage

Observation

Defence R&D Canada • R & D pour la défense Canada

Acquisition modes

• SSM: Sidereal stare mode

• TRM: Track rate mode

• TRM sequence: multiple TRM

– Perfect tracking

– With tracking error

Defence R&D Canada • R & D pour la défense Canada

Acquisition in SSM : (Sidereal Stare Mode)

Pointing

Sidereal motion

Sidereal trackingSidereal tracking

Expected satellite

Acquisition

DownloadDownload

Data analysis

Defence R&D Canada • R & D pour la défense CanadaDefence R&D Canada R & D pour la défense Canada

Acquisition in TRM : (Track rate Mode)

Sidereal tracking

Set track rate

Expected satellite

Acquisition

Download

Data analysis

Star detectionStar detection

Satellite detection

Defence R&D Canada • R & D pour la défense CanadaDefence R&D Canada R & D pour la défense Canada

Acquisition of a TRM sequence

Case 1:Perfect tracking

Set track rate

Perfect tracking

1Acquisition123

Image summation

VetoVe o(logical and)

Defence R&D Canada • R & D pour la défense Canada

Acquisition of a TRM sequence

Case 2:With tracking

Set track rate

error

1Acquisition123

Image summation

Defence R&D Canada • R & D pour la défense Canada

Processing of images acquired in SSMBad pixel correction

Noise estimation

Background estimation and removal

Star detection (but not streak)

Star removal

Streak detection(matched filter)(matched filter)

Alarm extraction

Defence R&D Canada • R & D pour la défense Canada

False alarm rejection

Background estimation and removal

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

BackgroundOriginal image

Method 1: iterative polynomial fit(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

False alarm rejection

Image - backgroundProfile of an horizontal line

starsbackgroundg+ noise

Defence R&D Canada • R & D pour la défense Canada

Background estimation and removal

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Original imageI

Method 2: iterative local means estimation and clipping(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

Calculation of the local mean

Firstiteration

Noiseestimation

n 8400

8500 original star profile (PSF)

fi t (15 15)

first clipping

False alarm rejection

Image clipping

Iterations2 to 5

n : noise standard deviation Estimated background

(Smoothed image)

<Bi>

n

8200

8300

8400 first mean (15x15)

second mean (15x15)after first clip

third mean (15x15)

second clipping

clipping

Clipped imageIbi

8000

8100

( )

fourth mean (15x15)fifth mean (35x35)

375 380 385 390 395 400 405 410

7800

7900

final background = minimum fourth and fifth means

Defence R&D Canada • R & D pour la défense Canada

Background estimation

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Method 2: iterative local means estimation and clipping (but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

Original image First meanFirst clipping Second mean... after five clipping False alarm rejection

Defence R&D Canada • R & D pour la défense Canada

Background residue and noise estimation

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Original image; acquiredin sidereal stare mode.

Estimated image background(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

F l l j i

Residualbackground:

False alarm rejection

Residue: Residual background + noiseBackground free image Exclusion zones

< n / 5

Defence R&D Canada • R & D pour la défense Canada

Residual noise

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Profile of the image residue

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

F l l j i

20

40

60Profile of the image residue

2000

2500Local noise distribution

False alarm rejection

-20

0

20

coun

t

500

1000

1500

coun

t

0 100 200 300 400 500 600 700 800 900 1000-60

-40

pixel

-80 -60 -40 -20 0 20 40 60 800

500

gray level

Gaussian distributionp

Residual background < n / 5

Gaussian distribution

Defence R&D Canada • R & D pour la défense Canada

Star detection … but not the streak

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

- Star detection and removingBackground-free image Detected stars

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

F l l j iFalse alarm rejection

S t d d bl t filt f Remaining objects

1 2 3 4

In(7x7)

out

Segmented double gate filter for star detection ... but not streak

6

512

11

In(7x7)

In(3x3)

Defence R&D Canada • R & D pour la défense Canada

8 710 9

Star removal

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Zeroing intensity above threshold ...... inappropriate,

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

create artefacts. False alarm rejection

Central disk set to zeroZone of subtracted profileUnaffected faint star

A B C

MedianMeanStandard deviation

S l i h fil

0 2 4 6 8 10 12024681012

Solution: measure the star profiles

... and subtract it

Defence R&D Canada • R & D pour la défense Canada

Streak detection

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Remaining objects 1st convolution

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

False alarm rejection

Matched filtre

Solution: iterative matched filter

Several false alarms

l

Defence R&D Canada • R & D pour la défense Canada

Matched filter for streak detection

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Star profile bStreak profilea

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

Convolvedn/2-n/2

C l dn/2-n/2

False alarm rejection

Convolvedby

1/nRectangle function

1/n

Convolvedby

n/2-n/2n/2-n/2

Almost a rectangle function

2b/n

Triangle functiona

Defence R&D Canada • R & D pour la défense Canadan-n

n-n

Iterative matched filter

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Objects(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

New Objects

False alarm rejection

convolution

22x

clippingpp g

Defence R&D Canada • R & D pour la défense Canada

Iterative matched filtert l i

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Single stars detected and removed:

Matched filter 1st convolution

3rdeconvolution

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

Remaining objects l False alarm rejection

iterations

Detected streak

Defence R&D Canada • R & D pour la défense Canada

Alarm extraction

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection (but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

A: Streak with remainingstars (boosted contrast)

B: Convolution peaks (after several iterations)

D: Binary masks generatedwith the clipped peaks

False alarm rejection

56

7

81

23

5

4 9

Noise still present Noise severely attenuatedin convolved image

Segmentation and labelling

Defence R&D Canada • R & D pour la défense Canada

Examples of extracted alarms

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

A pair of stars

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

A pair of stars False alarm rejection

The streak

A false alarm cause by a noise pattern

Defence R&D Canada • R & D pour la défense Canada

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Alarm rejection parameters(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

Moment ratio >10False alarm rejection

Length ratio > 1

<ISNR> (estimation of the local SNR)<ISNR> (estimation of the local SNR)> 0.5 for long streak> 1.0 for short streak

Defence R&D Canada • R & D pour la défense Canada

Moment ratio criteria

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

pMyypMxx

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

pMxx

pM

False alarm rejection

Myy

pMyy / pMxx >> 100 for a streak (typically) pMyy / pMxx 1 for a starpMyy / pMxx < 10 for a pair of starsp yy p f p f

Valid alarm: pMyy / pMxx >> 10

Defence R&D Canada • R & D pour la défense Canada

Length ratio criteria

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Ratio between the length of the length of the detection peak and

Length ratio = length of extraction mask (or convolution peak) > 1length of expected streak

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extractionRatio between the length of the length of the detection peak and

the length of the expected streak (129 pixels)

3

3,5

False alarm rejection

1 5

2

2,5

engt

h ra

tio

L

Mask length Streak length

0,5

1

1,5Le

Lm

sL

00 5 10 15 20 25 30 35

ISNR

Defence R&D Canada • R & D pour la défense Canada

Normalized ISNR Measurement

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Extraction mask(from the convolution

peaks)

Extracted object Standard selection mask (from the expected streak characteristics)

Selected pixels (for intensity calculation)

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

False alarm rejection

Defence R&D Canada • R & D pour la défense Canada

Valid detection threshold: determined by probability of detection in next slide

Performance evaluation

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection (but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction- The matched filter has very good performances- Consequence: Detection thresholds are set to lower value False alarm rejectionConsequence: Detection thresholds are set to lower value

and new false alarms are generated. - False alarm rejection performs very well, so:

what are the ultimate threshold detection limits?

• Calibrated synthetic image for detection testing

streak + PSF + photon noise

+ sensor noise

Defence R&D Canada • R & D pour la défense Canada

Simulated image with streak and pairs of stars

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection pairs of stars (but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

False alarm rejection

Pairs of stars represent th t bthe worst case becausethey are not eliminatedby star detection and removal process.

Two stars with the streakTwo stars with the streakalignment are like a short streak.

Defence R&D Canada • R & D pour la défense Canada

Probability of detection without false alarm (in presence of pairs of stars)

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection false alarm (in presence of pairs of stars) (but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

0 7

0,8

0,9

1 False alarm rejection

0 4

0,5

0,6

0,7

babi

lity

of d

etec

tion

50 pixels78 pixels129 pixels

Streak length

0 1

0,2

0,3

0,4

Prob

0

0,1

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1

ISNR

Defence R&D Canada • R & D pour la défense Canada

Detection probability without false alarm

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection false alarm

ISNR = 2.0 ISNR = 0.5 ISNR = 0.2

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

False alarm rejection

100% 100% 50%

Raw probability of detection:The algorithms always detect the streak with ISNR > 0.1, but with false alarms

b bili f d iNet probability of detection:With ISNR > 0.5, always 100% of detection completely without false alarm.

Defence R&D Canada • R & D pour la défense Canada

Detection results

Bad pixel correction

Noise estimation

Background estimation and removal

Star detection

Detections and Alarms

(but not streak)

Star removal

Streak detection(matched filter)

Alarm extraction

-----------------------------------------------------------------------image report: 0144 Expected length (pixels): 68.2495Expected orientation (degrees): 51.0725

False alarm rejection

p ( g )Alarm <ISNR> Center of intensity Probability of FAindex (counts) (Line) (Sample) (percent)-----------------------------------------------------------------------CONFIRMED DETECTIONS:

7 1.398 649.5 567.7 0.00-----------------------------------------------------------------------POSSIBLE DETECTIONS:

6 0.270 858.8 401.0 1.005 0.261 924.9 301.8 1.00

-----------------------------------------------------------------------OTHER REJECTED ALARMS: 10 1.239 10.5 1001.6 1.001 0.437 34.4 7.8 1.0011 0.297 40.1 1014.5 1.003 0.281 986.0 223.4 1.002 0.190 944.1 77.8 1.008 0.167 912.3 963.7 1.004 0.120 928.9 251.5 1.009 0.102 1013.0 975.1 1.00

-----------------------------------------------------------------------

Defence R&D Canada • R & D pour la défense Canada

Processing of images acquired in TRMBad pixel correction

Noise estimation

Bad pixel correction

Noise estimationrecycled

Background estimation and removal

Background estimation and removal

recycledStar detection (but not streak)SSM TRMStar streak detection

(matched filter)

Streak feature extractionrecycled

Star removal

Streak detection(matched filter)

Star streak removal

( )

S t llit d t ti S d i(matched filter)

Alarm extraction Alarm extraction

F l l j ti

Satellite detection Star detection

Streak unmerging

Defence R&D Canada • R & D pour la défense Canada

False alarm rejection False alarm rejection

Automatic extraction of streak features:length and orientation

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extractionlength and orientation

Fourier

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

transform

Radontransform

Profile at 145degrees

Alarm extraction

False alarm rejection

g g

145 degreesProfile at 145degrees

transform

degrees

M i

The zeros of the sin(x)/x functionindicate the length ‘l’ of the

Rect(x/l) function Maximumvalue

Rect(x/l) function,i.e. the streak length

Defence R&D Canada • R & D pour la défense Canada

Star detection-1

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Matched filter

clipping

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Original gray shade image

clipping

Filtered gray shade image Convolved streaks

Alarm extraction

False alarm rejection

g g

Defence R&D Canada • R & D pour la défense Canada

Binary mask: original > 2 n Binary mask: filtered image > 2 n Mask : convolved streaks > 1 n

Star detection-2

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Alarm extraction

False alarm rejection

g gBright star Faint star

Mean value Thresholdset at 1 n

centroid

Red line: streak profileBlue line: convolution peak profile

Defence R&D Canada • R & D pour la défense Canada

Star detection-3

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Binary mask: filtered image > 2 n

Bright star removed New binary mask:convolution peak > n

Alarm extraction

False alarm rejection

g g

Matched filter

Detected bright starsl k

Defence R&D Canada • R & D pour la défense Canada

New convolution peaksDetected bright and faint stars

Star unmerging

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Extraction maskExtracted objectsConvolution peaksAlarm extraction

False alarm rejection

g g

Modeled peakPeak subtractione sub c o

Geometric centerPeak maximum Peak subtraction

Case of merged stars detected

Second peak detected

Example of unmerging

Defence R&D Canada • R & D pour la défense Canada

p

Star detection and unmerging

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Alarm extraction

False alarm rejection

g g

Defence R&D Canada • R & D pour la défense Canada

Star removal and satellite detection

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Condition 1: Satellite intensity > noiseDetection threshold

depends on the measured PSF

Alarm extraction

False alarm rejection

g gmeasured PSFBackground-free image - Stars (i.e. the filtered streaks) = image difference

Insertedtest pattern

for sensitivityevaluation

Condition 2:and

Satellite brighter than the overlapping streakCondition 2: Satellite brighter than the overlapping streak(with enough contrast)

Defence R&D Canada • R & D pour la défense Canada

Detection and alarm rejection

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

3- Rejection:Brighter than close

Alarm extraction

False alarm rejection

g g

PSF area

gstreak? (again)- connect to closest object

Closest streakmask2- Rejection:

- Object-to-object comparison: Brighter than mean brightness

Alarm does not overlap close streak mask

of overlapping streak?

Streak in the1- Alarm:

Single-pixel alarm Can be caused by

Streak in thePSF range

- Pixel-to-pixel comparison: Brighter than aligned pixel

Defence R&D Canada • R & D pour la défense Canada

ysignal pixelisation

Satellite detection

Bad pixel correction

Noise estimation

Background estimation and removal

Streak feature extraction

Star streak removal

Star streak detection(matched filter)

Satellitedetection

Stardetection

Streakunmerging

Alarm extraction

False alarm rejection

g g

DetectionRejected false alarms

Defence R&D Canada • R & D pour la défense Canada

Sensitivity of SSM vs TRM acquisition

SSM: PSF=3

MODE Detection threshold

Required objecti t it

SSM PSF 3

SSM: PSF=3,Streak = 130 pixels

210 nintensity

TRM, PSF<1 9 n 9 n

TRM, PSF=3 3 n 45 nSSM: PSF=3,

Streak = 50 pixels 160 nSSM, 50 pixels n 160 n

SSM, 130 pixels n/2 210 n

TRM: PSF=3(3x3) local average

45 n

TRM: PSF<1Single pixel

9 n3 n /2

Defence R&D Canada • R & D pour la défense Canada

n n n/2

Detection algorithms for TRM image sequence

Is in development, it includes:- Star detection and astrometry done for each imagesy g

- Provides knowledge of real pointing (drift estimation)- Image registration

- Uses satellite predicted position and- track rate correction (drift correction) which uses the real sensor

pointing (astrometry) - Summation of registered images

- More sensitive detection (better SNR) with the summed image.- Differentiation of registered images

- Reject star streaks and cosmic ray hits.- Detection in individual frame:

- cued by the global detection,- uses a lower detection threshold;

Defence R&D Canada • R & D pour la défense Canada

* false alarm elsewhere are ignored.

Conclusions

- Detection algorithms developed for both SSM and TRM.- Star and satellite (point or streak) detection is f ll t tifully automatic.

- Iterative background estimation and removal more reliable and accurate than the classical dark frame method.a d accu ate t a t e c ass ca da a e et od.

- Iterative matched filter very efficient for streak detection.- Sensitivity limited by noise level:y y

- streak (very long) > 1 n (> 0.5 n),- single point object (in TRM) > 9 n,- adaptable to the PSF.

- Detection for TRM sequence in development.- TRM acquisition more sensitive than SSM.

Defence R&D Canada • R & D pour la défense Canada