Fast Radar Image ChangeDetection by Matlab Embedded

Routines

Hans HellstenSaab EDS, April 2016

Overview

q Saab has developed a novel type of low frequency airborne radar – CARABAS -which detects objects even if camouflaged, e.g. by trees screening.

q The radar images are temporally stable (i.e. do not change from oneregistration to another) so detection can be achieved by “change”. For instancea vehicle being parked and camouflaged in between flights will be detected.

q This talk addresses the specific probability method for change detection: Itassesses the probability of the small natural changes and compares with achange caused by an appearing object. Thereby the probability that a changemeans that an object has appeared in any particular position is obtained.Targets become found and listed in order of decreasing probability.

q A challenge is the computation time in obtaining the statistics for naturalchanges since images are large (millions of pixels) and must be compared witheach other. Nevertheless using Matlab embedded routines a method werefound to reduce the computation time by one to two orders of magnitude, thusmaking the approach practical.

q It is suggested that the method practiced has a wider application speeding upany routine which requires manipulating functions defined over large data sets –typically images.

High band – 220 MHz - 0.75 m resolution

Low band – 55 MHz - 3 m resolution

Change detection imageryfrom two flights – appearingobjects are red anddisappearing green – thoseunchanged are yellow.

Upper image is for a higherfrequency radar band. Thelower band at the lowerimage evidently works muchbetter

High band – 220 MHz - 0.75 m resolution

Low band – 55 MHz - 3 m resolution

-500 -400 -300 -200 -100 0 100 200 300 400 500

-250

-200

-150

-100

-50

0

50

100

150

200

250

Lower band image from first flight

-500 -400 -300 -200 -100 0 100 200 300 400 500

-250

-200

-150

-100

-50

0

50

100

150

200

250

Lower band image from second flight

1. Asses change statics by counting number of cases when for a given amplitude in Athere is a given amplitude change between A and B

2. Divide by the calculated probability that a target causes such a change thus obtainingprobability that for ach image point there is a target

3. Find the most probable target candidate and exclude4. Redo process until there are no more probable targets

The computational burden in step 1 can be very large since it typically means testing eachimage pixel whether it has certain image amplitude and change amplitude (within sometolerance). It thus has to assess all image pixels (millions) several hundred times. Accordingto 3 this must be redone for every target in the image.

Statistical Method

0 0.1 0.2 0.3 0.4 0.5 0.6

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 -7

-6

-5

-4

-3

-2

-1

0

1

10-logprobability

density

Magnitude

referenceim

age

Change magnitude

NOTE: The stronger the intensity in thereference image the more uncertain theamplitude in the update

Change statistics with no appearing ordisappearing objects

Changes in this area will be detectable

Ra

Ua0 0.2 0.4 0.6 0.8 1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

( ) ( ) ( )xuxuxu terrainobjecttotal +=Target model

Magnitude

referenceim

age

Change magnitude

Cum

ulativeprobability

Fast method of obtaining statistics

Define logical masks: DSUPPOA=AMP>=BILD;Statistics: sum(DSUPPOA)

AM

P

CO

ND

CH

AN

GE

AM

P

GENERALIZATION OF FAST METHOD

q Consider any function f(x) defined on any set X, e.g. the pixels in an image.

q Consider any operation which for each value f=const of the function performsome task specific to that value. Examples can be counting the number ofelements in X holding this value, or it can be to change the value according tosome rule (e.g. multiplication), or to sum the value for the subset of pointsholding the value (integration).

q Common for all these situations is that it is not required to loop through the setbut that the set can segmented as f -1(const). The operation can be performedas one single computational step for all elements in f -1(const) while the overallsolution is attained by recombining the segments across all possible functionalvalues.

q The gain in computational effort is that functional values (within a pre-definedtolerance) are common to many elements.

GENERALIZATION OF FAST METHODOperate here to accessmany elementscollectively by fewoperations

Many Elements(e.g. image)

Few

Ele

men

tsre

pres

entin

gth

em

any

elem

ents

Performance of MethodExperiment with artificial insertion of three additional weak targets

Artificial: 0.75 x pers car

Artificial:0.5 x pers car

Artificial:pers car

Land Rover

pers car

200 400 600 800 1000 1200 1400 1600 1800

100

200

300

400

500

600

700

800

900

-7

-6

-5

-4

-3

-2

-1

12.25 %

Probability image – 0 target hypothesis10-log

probability

200 400 600 800 1000 1200 1400 1600 1800

100

200

300

400

500

600

700

800

900

-7

-6

-5

-4

-3

-2

-1

44.06 %

2.35 %

10-logprobability

Probability image – 1 target hypothesis

200 400 600 800 1000 1200 1400 1600 1800

100

200

300

400

500

600

700

800

900

-6

-5

-4

-3

-2

-1

6.57 %

99.80 %5.28 %

10-logprobability

Probability image – 2 target hypothesis

200 400 600 800 1000 1200 1400 1600 1800

100

200

300

400

500

600

700

800

900 -6

-5

-4

-3

-2

-1

25.25 %

99.88 %18.69 %

10.38 %

10-logprobability

Probability image – 3 target hypothesis

200 400 600 800 1000 1200 1400 1600 1800

100

200

300

400

500

600

700

800

900 -6

-5

-4

-3

-2

-1

99.90 %

99.91 %99.89 %

99.85 %

4.87 %

10-logprobability

Probability image – 4 target hypothesis

200 400 600 800 1000 1200 1400 1600 1800

100

200

300

400

500

600

700

800

900-6

-5

-4

-3

-2

-1

99.92 %

99.93 %99.91 %

99.88 %

6.66 %

Conclusion: 4 targets found- the weakest not !

10-logprobability

Probability image – 5 target hypothesis