A Low Cost Gun Launched Seeker Concept Design for Naval ... › Content › techdigest › pdf › V22-N04 › 22-04-Maurer.pdfForward Air Support-Marine-Gun Launched Loitering Munition,

634 JOHNSHOPKINSAPLTECHNICALDIGEST,VOLUME22,NUMBER4(2001)

D. E. MAURER et al.

T

ALowCostGunLaunchedSeekerConceptDesignforNavalFireSupport

Donald E. Maurer, Eric W. Rogala, Isaac N. Bankman, Bradley G. Boone, Kathryn K. Vogel, Christopher Parris

heLowCostGunLaunchedSeekerProgramisacomponentoftheNavy’sefforttodevelopeffectiveweaponsforsurfacefiresupportmissions.Theobjectiveistoimprovetheperformanceofprecision-guidedmunitionsagainsthardenedtargetsbydevelopingalow-costterminalseeker,includingdetectionandtargetselectionalgorithms,toautono-mouslyacquireatargetandprovidecommandstoguidetheprojectile.Thistechnologywillhavewidespreadapplicationtothetime-criticalengagementofbothground-andsea-basedthreatsbyprovidingasmart-weaponscapabilityunderthecontrolofthefieldcom-mander.Thisarticleprovidesanoverviewoftherequiredsystemcapabilities,discussesthedesignrequirementsimposedonsuchasystembythesecapabilities,andsummarizesAPL’sefforttodevelopasystemthatmeetsmissionobjectives.

INTRODUCTIONThis article describes a proposed target detection

system, consisting of an infrared (IR) seeker withtargetdetection,classification,andselectionalgorithms,designed to meet the mission objectives of the LowCost Gun Launched Seeker (LCGLS) Program. Thissystemisintendedtoimprovetheperformanceofpro-jectiles like the Extended Range Guided Munitions(ERGM)againsthardenedtargetsbydevelopingalow-cost,uncooled IR staring focalplanearray (FPA) ter-minal seeker.The seekerwill autonomously acquire atarget and provide commands to guide the projectile.Aground-basedforwardobserver,aircontroller,orair-borne asset locates and classifies a target and issues a“callforfire.”TargetlocationandtargettemplatesareloadedintotheLCGLSandtheprojectileislaunched.

The seeker must survive the initial 15,000-g launchacceleration.ItnavigatestothetargetusingtheGlobalPositioningSystem-AidedInertialNavigationSystem.At approximately 5 km from the threat, the seekerenters search mode. It acquires the target at approx-imately 2–3 km, begins tracking, and provides com-mandstoguidetheprojectile.Biasedproportionalnav-igation begins to create a lofted trajectory and steepapproach at a terminal velocity of approximately 150m/s.Attheoptimumpoint,aninexpensivestandofffuzeinitiatesthewarhead.

The seeker has been proposed to track fixed, relo-catable,andmobilelandandseatargetswithsufficientaccuracy to achieve a small circular error probability(CEP).Iftheseekerisactivatedbelowthecloudlevel,

JOHNSHOPKINSAPLTECHNICALDIGEST,VOLUME22,NUMBER4(2001) 635

LOW COST GUN LAUNCHED SEEKER CONCEPT

nominallyat300to900mbasedoncloud-freeline-of-sightstatistics,ithasabout2–6stofindthetargetand,ifnecessary,completeasingledivert.Ifthecloudceil-ingishigher,theseekercouldbeactivatedearlierandmoretimewouldbeavailabletofindthetarget.Thus,some of the major constraints include the ability tosurvive a high-g launch, limited space and weight fortheseeker,environmental factorsaffectingseekerper-formance, and terminal engagement geometry, whichdeterminestheseekerfieldofview(FOV)andthetimeavailable for target detection. These constraints limitthefeasibleopticaldesignsandthecomplexityandcom-putationalrequirementsofthedetectionalgorithms.Anadditional programmatic constraint is cost (preferablylessthan$10,000perroundfortheseeker).

Thetechnologywillhavewidespreadapplicationtothetime-criticalengagementofbothground-andsea-based threatsbyprovidinga smart-weaponscapabilityunder the control of the field commander. Many ofthesescenariossharethesamecompressedtimelineinthe terminal engagement process, which requires sub-stantial real-time image processing functionality sup-porting highly responsive terminal guidance, targetseparation from clutter, and aimpoint selection. Theabilitytoaccuratelydelivermunitionsinsupportofpre-cisionattackusingtacticaltargetingsourceswillgivethecommandergreaterflexibilityandautonomy,thusreduc-ing the time required to respond in a dynamic threatsituation.

AlthoughoriginallydesignedasanenhancementofERGM,theLCGLStechnologyisapplicabletoawiderange of weapons programs where a sensor or seekersystem is needed. Examples include the Rapid Ord-nanceDefenseSystem,JointDirectAttackMunitions,ForwardAirSupport-Marine-GunLaunchedLoiteringMunition, as well as Affordable Weapon, PrecisionGuidedMortar,andothersmartsubmunitionsconcepts.Unmanned aerial vehicles, precision targeting, battledamageassessment,andidentificationfriendorfoe(IFF)arealsopotentialapplicationareas.

The focusof thisarticle isonthealgorithmsbeingdevelopedforthesystem.However,toprovidecontext,the desired functional capabilities of the LCGLS aresummarizedaswell as theproject’s requirements.Ear-lier work focused on conceptual designs and prelimi-nary analyses to determine if they met requirements.Inviewof thepromisingoutcomeof theseefforts,webegantodevelopasetofalgorithmsthatwillautono-mouslydetectandtrackthedesiredtargets.Groundtar-getsposeespeciallychallengingautomatictargetrecog-nition problems and continue to be the focus of ourwork.Apreliminary studyof thealgorithmexecutiontime is discussed here, followed by a demonstrationof thealgorithm’s typicalperformance foravarietyofgroundtargetimages,includingbuildings,bunkers,andairfields.

FUNCTIONAL CAPABILITIESIntheconceptualdesignweconsideredabroadmis-

sionencompassingincreasinglychallengingtargetsce-narios.Subsequentlywefocusedonthetwomostcriticalfunctionalcapabilities:targetsegmentationandtermi-nalaimpointselection,i.e.,theacquisitionofhardtar-gets against complex natural backgrounds. Additionalcapabilityagainstmovingtargets,althoughnotneces-sarilymorechallenging,willlikelyinvolvein-flighttar-getingupdates,which require a data link.More chal-lengingscenariosincludeselectinghigh-prioritytargetsinaclutterbackground,especiallyneartheforwardedgeof thebattle; controlling salvodistributions formaxi-mizinglethalityoverlargeareas;andavoidingcounter-measuresandjamming.

The FOV needed to support these capabilities isdeterminedbytheterminalengagementgeometryillus-trated in Fig. 1. The diameter of the footprint thatmust be contained in the FOV at initial acquisitionby the seeker consists of a combination of 10 m forthe ERGM baseline terminal guidance accuracy forhand-overand10m for targeting(root sumsquared).Ordinarily this diameter corresponds to a 50% targetcontainmentrequirement.However,weusea90%con-tainmentprobabilitycorrespondingto26minorderto

h

aM cos z

vM cos

aM sin

aM

g

vM sin

vM

FOVFOV2

2

rmin

rmax

r

xCEP

xFP

x

y

Figure 1. Terminal engagement scenario (aM = acceleration per-pendicular to the projectile boresight, vM = terminal velocity, xFP = diameter of the FOV footprint, xCEP = circular error probability, and = seeker angle).


D. E. MAURER et al.

establish a conservative FOV. For simplicity, assumetheseekerispointedstraightdown(=0).Thenitcanbe shown that the FOV must be greater than 10° foracquisitionrangesaslowas300m.Ofcourse,iftargetacquisitionandtrackoccursatashorterrange,theFOVmust be correspondingly larger. Therefore, we nomi-nally assume a minimal 14° FOV. This should allowtargetcontainmentinthe90%CEPdowntoarangeofapproximately275m.

Amajorconsiderationistheairframeonwhichtheseeker must operate. ERGM, like most guided projec-tiles, has a relatively low divert capability. The air-framehasbeendesignedtoremovesmallheadingerrorsthroughout flight as it is guided to a predeterminedpoint. The ERGM divert capability is a function ofvehicle speedbut is typicallyabout±2g.Astudywasconductedearly in theprogramtodeterminewhetherERGM or a similar platform could provide enoughdivert capabilityduring the terminalphase to removethe IR-measured guidance errors given a few secondsofhomingtime.Theresultsof thestudywereusedtoassessterminalguidancepolicyoptions(terminalveloc-ity, acquisition range, approach angle, and hit-to-killversusshoot-to-kill).

Atwo-degree-of-freedom(2-DOF)simulation,con-tainingarepresentationoftheERGMkinematics,wasdevelopedtoparametricallystudytheterminalhomingphase.Thissimulationwasusedtoconductatradestudyofinterceptgeometryandseekerlookangleasafunc-tionof projectile divert capability, projectile velocity,andpositionandattitudeerrorsatacquisition.Asampleoutput in Fig. 2 shows miss distance as a function of

thegcapabilityoftheairframe.Inthiscasetheinitialpositionerror(deliveryplustargeting)was30m,initialheadingerrorwaszero,andtheprojectilevelocitywas150 m/s. The 2-DOF results showed that a ±2 g air-frameismarginalforahit-to-killscenarioifthetargetisacquiredatapproximately300m(suchasifthecloudceilingislow).Betterperformanceisachievedforear-lier acquisitionandhence longerhoming times; thus,thetargetmustbedetectedandclassifiedassoonafterseeker activation as possible in order to optimize thelimiteddivertcapability.This,inturn,limitstheavail-ableprocessingtime.

SEEKER DESIGNInitialassumptionsmadetominimizecost,complex-

ity,andform-factorincludebodymountingtheseekerand optics; using lightweight optical designs; z-planechippackaging;anduncooledFPAtechnologytoreduceseekerform-factorandobviatecryogeniccooling.Theuseofbody-mountedoptics (comparedwithgimbaledoptics) simplifies the design, reduces cost and weight,andincreasessurvivabilityduringlaunch.Z-planechippackaging is a state-of-the-art technique that reducesFPA and post-processor form-factors. It consists ofstacked thinned-silicon integrated circuits coupled tothe IR FPA. Northrop Grumman has fabricated sucha chip under the DARPA-sponsored Flexible Manu-facturingProgram.1AnuncooledbolometricFPAdoesnot requirecryo-cooling,which iscostlyand takesupspace(althoughtheuncooledFPAmayrequireather-moelectric cooler for temperature stabilization). Non-uniformitycompensationmayberequired,however,toimprovetargetselectionandclutterrejection.

The DetectorAlthoughtheexperienceofprocessingmeasuredIR

image data may ultimately help determine the exactspectral band, long-wave IR (LWIR) is preferred foranumberof reasons. For instance,LWIRattenuationis generally less than mid-wave IR (MWIR) at shortranges (about 2.5 km), although, for ranges less than1km,thesensorshouldnotbesignificantlysensitivity-limited ineitherbandundermostatmosphericcondi-tions.Althoughmostenergy is intheLWIRbandforexpected targets near equilibrium, differential cluttercontrastisgenerallygreaterintheMWIRband.Further-more,eventhoughMWIRhasbetterdiffraction-limitedspatial resolution, a well-designed seeker modulationtransferfunctionshouldnotbedominatedbytheoptics,but by detector size. Finally, and most importantly,uncooled array technology is currently available onlyforLWIR.

Target-background contrast, T, in the LWIR atthe FPA yields an output signal-to-noise ratio (SNR)givenby

300

270

240

210

180

150

120

90

60

30

0

Alti

tude

(m

)

–30 –27 –24 –21 –18 –15 –12 –9 –6 –3 0

Downrange (m)

1-g limit

2-g limit

3-g limit4-g limit

5-g limit

Figure 2. A 2-DOF simulation showing miss distance as a func-tion of g capability of the airframe.



SNRNETD

,atm≅ 0 T

where 0 is the mean optics transmissivity over theprescribedspectralband,atmisthemeanatmospherictransmissivity over the same band, and NETD is thenoiseequivalenttemperaturedifference.

The mean T, based on the contrast temperaturerangeforatypicalmilitaryvehicle(T-72tank)againsta grassy background, over various azimuths, is about1.0 K for measured data and 1.5 K for modeled data(personalcommunication,S.E.Motz,U.S.ArmyMis-sile Command, 24 Apr 1998). Significant factors fordetermining theactual scenecontrast (andhence thelocalscenecontrastSNR),however,arediurnalvaria-tionsinT,whichcansometimesbezero,principallyas a result of the combined effect of differential solarloadingandtheconstituentmaterials,theirorientation,andtheirassociatedemissivity.Thiseffect,calledther-mal crossover, is particularly noticeable in “pillbox”-typebuildingstructures.

The extinction coefficients, , for LWIR versusMWIR for various atmospheric obscurants (includinggases, haze, fog, rain, snow, and dust) are not signifi-cantly different overall. At an altitude of 900 m (thecloud-free lineof sightabout90%of thetime,world-wide) with =0.7 km1, corresponding to weatherthatis“poor”orbetter80%ofthetime,andNETD=0.1K, a typicaluncooledbolometerFPA2hasaSNRvariation between 8 and 12. The larger the entranceapertureoftheseeker,thebettertheSNR.Theentranceaperturesize,however,willbelimitedbytheairframetoapproximately10cm.

Now,toensureoptimalresolutionforidentificationofavehicle-sizedtargetatthegreatestexpectedacqui-sitionrange,weassumethat8pixelsarerequired fortarget identification,3 and that the minimum targetdimensionis5.18m(≈10%CEP).ThenthenumberofpixelsrequiredintheFOVrangesbetween85foraverticaldescentto176forarelativelyextremedevia-tionfromvertical.Therefore,a256256arrayshouldbesufficient.

OpticsThe most stringent design requirements include the

abilitytosurviveahigh-gshockatlaunch;minimalsize,weight,andvolume;andlowcost.Thefavoreddesign,af/1.54systemwithaFOVof22°,aclearapertureof20mm,and79.9%ensquaredenergyonthecenter50-mpixel,hasthelargestFOVandthemostcompactsizeofthecandidatedesignsstudied.Thelengthofthesystemis49.8mmanditcanfitinacylinderofvolume35.2cm3.

Requiring both LWIR operation and shock surviv-abilityseverelylimitsthechoiceofviablelensmateri-als. First, thematerialmustbe adequately transparent

throughouttheentireLWIRspectralband.Ofthefewmaterials that are chemically stable enough to with-standtheintendedstorageanduseenvironment,onlythosethatcansurvivethemechanicalstressoflaunchshockareacceptable.Apreliminarystructuralanalysisusing various materials, including polyethylene, zincsulfide,zincselenide,andgermanium,determinedthatall these shouldsurvive. It is important tonote,how-ever,thatthelenseswerenotdesignedtoimageduringtheshock,andthat,intheanalyses,the mechanical stress was not applied as an impulse.Neitherdid theanalysesaccount for material or surface characteristics; thesecouldsubstantiallylowertheallowablestress.

All the materials being investigated would requirea thin-film antireflective coating on both surfaces toincreasetransmissiontoanacceptablelevelandtopro-videforaspecificwavelengthpassbandoptimizedtotherequired LWIR camera response. For the front opticalelement,thiscoatingmustsurvivethestressesofbeingexposed to avarietyof atmospheric conditions at rela-tivelyhighspeeds.Also,ifthereisexcessiveaerodynamicheatingofthefrontoptic,theimpactonboththeopticalandmechanicalintegrityofthelenshastobetakenintoaccount.Forinstance,asgermaniumisheatedtogreaterthan50°C, the transmission through thebulkmaterialdrops,andtheLWIRemissivityincreases.Whengerma-niumreaches250°C,itisessentiallyopaqueintheLWIRand thusno longer aviable lensmaterial.Webelieve,however,thatifopticsarenotexposeduntiltheterminalphaseoftrajectory,theywillnotbesubjectedtoatmo-sphericflowatsubsonicspeedsformorethanafewsec-onds,andhencetheseconsiderationswerenotamajorconcern.

DETECTION ALGORITHMSTheanalyses summarized in thepreceding sections

provideapositiveinitialassessmentofthefeasibilityoffielding a gun-launched imaging IR system. The nextcriticalstepistodemonstratealgorithmsthataccuratelydetectdesiredtargetswithahighprobabilityofsuccess.Limitedprocessingtimeis,perhaps,themostsignificantconstraint.Theprojectileapproachesimpactatahighrateof speed; the timebetween seekeractivationandimpactisonlyafewseconds.Inordertohavesufficientdivertcapability,thetargetmustbeselectedassoonaspossible. Moreover, the frame rate is on the order of60Hztoavoidscalechangeduringtracking;therefore,iftheframesareprocessedmuchslowerthanrealtime,itwillbedifficult tocorrelateanobject found inoneimagewiththesameobjectseveralframeslater.Sincethere is no uplink, the algorithms must be capableof autonomous operation. Such a system cannot relyon detailed targeting information to guide its search;thus,standardtechniquesrequiringterrainmappingordetailedtargetlibrariesarenotfeasible.


D. E. MAURER et al.

Thefundamentalphilosophyguidingouralgorithmdesignistousedifferentprocessingmodalitiesoperat-inginparalleltoprovidedetectioncapabilityunderagreater range of image conditions than would other-wisebepossible.Figure3showsthebaselinedetectionarchitecture implementing this approach. The pro-posedsystemincludestwocomplementaryparallelpro-cessesconsistingofregion-basedandedge-baseddetec-tionandsegmentation.Byfusingtheoutputsofthesedetectionalgorithms,wehopetoenhancerobustnessand attain greater discriminatory power. The fusedoutputwillreducefalsealarmsandprovideregionsofinterest (ROIs) for further analysis.A set of featuresforanalyzingtheseROIswillevolveasthealgorithmsdevelop.Ideally,theonlyROIspassedtotheclassifierwould be those containing target objects. Since theseeker has a restricted FOV—it is boresighted andtheprojectilehaslimitedmaneuverability—itcannotsearchanextendedfootprint.If itdoesnotdetectitsintendedtarget, the seekermustdefault toanyman-madeobjectitcanfind.Ourfirstobjective,therefore,is todemonstratethecapabilityofthissystemtodif-ferentiatebetweenman-madeobjectsandnaturalfea-tures. If this is successful, more refined classificationalgorithms and techniques for detecting moving tar-getswillbedeveloped.

Region-Based Processing

Histogram EnhancementTheregion-baseddetectorusesintensitytosegment

theimage.Ideallythehistogramofanimageconsistsoftwomodes,onecorrespondingto thebackgroundandtheothertothetarget.Thusathreshold,chosentobethegray levelbetween themodesatwhich thehisto-gram is minimal, can be used to partition the imagesinto target andbackgroundclasses.However, thehis-togram may not exhibit this simple form; thereforehistogram enhancement techniques have been devel-oped to accentuate the desirable features and thusextend the range of images to which this techniquecanbeapplied.RoundsandSutty4describeahistogramenhancementtechniquebasedonco-occurrencematri-ces. In general, a co-occurrence matrix measures thefrequency with which pairs of nearby pixels havesimilar intensity. An example is the matrix whose(m, n)th entry is the number of pixel pairs (i, j) and(i 1, j + 1)forwhichthefirstpixelhasgraylevelmandthe second has gray level n. Similar directional co-occurrence matrices are obtained for pairs consistingof a pixel and any one of its neighbors. Since pixelpairswithinatargetregiontendtohavesimilarinten-sity,theseregionscontributetolargerentriesnearthemaindiagonal,andtheco-occurrencematrixhasamoreprominent modal structure along the main diagonalthantheordinaryhistogram.ThisisillustratedinFigs.4aand4b.

Thealgorithmthencalculatesanenhancedhistogramby averaging the co-occurrence matrix entries withinanarrowbandcontaining themaindiagonal.Tofindthethresholdbetweentargetandbackground,thetwomodesarefirstdetectedandthenthethresholdcanbechosentobetheintensityvaluebetweenthetwopeaksatwhichthehistogramassumesitsminimalvalue.

ThesimplestpeakdetectorcomparesthehistogramvalueH(u) at a test celluwith itsvalues at adjacentcells and declares a peak if H(u1)<H(u) andH(u+1)<H(u). If the histogram is rough, however,thismethodcanproducemanyfalsealarms.Therefore,normally,awindowiscenteredatthetestcellu,andH(u)iscomparedwiththeaveragevalueofthehisto-gramoverthiswindow.Thelengthofthewindowcanbe adapted to the data at hand. For peak detectionintheLCGLSapplication,thistechniqueisextendedby centering a window of length 2N+1 at the testcell and computing two histogram averages m and M(assume m≤M), one for each half-window on eitherside of the test cell. The peak detection threshold isT=RM+(1R)m,whereRisanon-negativeparame-terthatdetermineshowmuchlargerH(u)mustbethantheadjacenthistogramvalues inorder tobeclassifiedasapeak.This reducestosimplewindowaveraging ifR=0.5.When the test cell is locatedat thecenterof

Image

Detectcontrasting

intensity levelsEnhance/detect

edges

Fuse output

Extractfeatures

Detect target

Classify

Target label

Figure 3. Baseline target detection architecture.



a relatively symmetric peak, M≈m, and the responseis,again,essentiallyequivalenttowindowaveraging.Afurther smoothing process removes smaller peaks thatmay otherwise cause false alarms. The improvementprovidedbythistechniqueisevidentwhencomparingtheenhancedhistogram(redcurve)withthestandardintensityhistogram(bluecurve)showninFig.4c.

QuantizationAnalysesofhistogramsobtainedfromgroundtarget

imagery collected at the Naval Air Warfare Centershowthat,ingeneral,usefulinformationiscontainedinallsignificantpeaks.Therefore,thetechniquedescribedabovehasbeengeneralizedtoamultithresholdapproachbydefiningasetofintensitylevelsT0,T1,…,Tncorre-spondingtothelocationofpeaksexceedingthedetectorthreshold.Therawimageisquantizedbyreplacingthevalueofeachpixelwhoseintensityliesinthehalf-openinterval[Ti,Ti+1)byi.Thequantizedimageretainsthe

salienttargetfeatures,butthereareonlyaboutadozenintensitylevels.Thisrepresentsasignificantdatareduc-tionand,hence,fasterprocessing.Figure5isatypicalexample.Figure5ashowsarawimage(color-enhanced)of aircraft of various sizes, buildings, and roads in thedesert. Figure 5b is a plot of the enhanced histogram(blue curve) showing several significant peaks. Themagentacurveisthethreshold,determinedasdescribedabove. There are at least four peaks exceeding thethreshold (the scaleof thefigure is too gross to showdetailsofthestructureonthefarleftofthehistogram).ThequantizedimageisshowninFig.5c.

FilteringAlthoughquantizationreducesthedynamicrange,

objectsmaystillbecomposedofseveralintensitylayers.An attribute signature comprising the geometric andintensitycharacteristicsoftheselayersprovidesapow-erful feature that can be used to filter the quantized

Figure 4. Calculating the enhanced histogram: (a) the original image, (b) matrix obtained by averaging the directional co-occurrence matrices, and (c) the enhanced histogram compared with the standard intensity histogram.

40

Num

ber

of p

ixel

s

30

20

10

0

–10

–20

0

20

40

60

80

100

Gra

y le

vel

Threshold = 45

0 50 100Gray level

Intensity histogram

Enhancedhistogram

0 20 40 60 80 100Gray level

(a) (c)(b)

Figure 5. Quantization: (a) raw image and (b) enhanced histogram (blue curve is the co-histogram obtained from the raw image data; magenta curve is the threshold). Each peak exceeding the threshold corresponds to a gray level in the quantized image (c).

(a) (c)(b)

Hot

Cold


D. E. MAURER et al.

imagetoremovenontargetregionssuchasvegetationandterrain.Theattributesignature,forexample,mayspecifyacceptablerangesinsize,shape,andintensity;anditcouldalsoincludeimportantinformationlinkingthesefeatures.Acompactdatastructurecalledacom-ponenttree5hasbeendevelopedtoimplementfilteringbasedonattributesignatures(seetheboxedinsert).

Filteringan image reduces toadecisionprocessonthe associated component tree that determineswhichnodes are removed. This process provides a degree ofrobustnesssincethetestcriteriacanbefuzzyorexpressedin linguistic rather than quantitative terms. The onlyrestriction is imposedbycomputational limitations. Iftheprocessingisappliedtoarawimagewith256graylevels, for example, thecomponent tree isquitecom-plex.If,however,theimageisfirstquantizedwiththe

multithreshold algorithm, the number of gray levelsis reduced to a dozen or so, greatly simplifying thecomponent tree; consequently, more complex filtercriteriacanbeimplementedwithoutdegradingthefil-teringspeed.Researchiscontinuingtodetermineopti-malfeaturesfordefiningfiltercriteriaandhowbesttoperformthefilteringandsegmentationsteptoproduceas robust an algorithm as possible within the LCGLSconstraints.

Edge-Based ProcessingToprovidegreatersegmentationcapability,analgo-

rithm that extracts edge information has been imple-mentedtoworkinparallelwith,andcomplement,theregion-processingalgorithm.Thisedge-basedtechnique

FILTERING COMPONENT TREESAconnectedcomponentofanimageisaconnectedregion,

allofwhosepixelsaregreaterthanorequaltoagiveninten-sity level. Two components of different levels are linked iftheyareconnectedtoeachotherandthereisnocomponentofanintermediatelevelconnectedtoboth.Thecomponenttree derived from an image consists of nodes correspondingtoeachconnectedcomponent,andtwonodesareconnectedbyanedgeifthecorrespondingcomponentsarelinked.Theleavescorrespondtoregionalmaxima,i.e.,connectedregions,allofwhosepixelshavethesameintensityandsuchthatallsurroundingpixelshavelowerintensity.Therootnodecorre-spondstotheentireimage.Thesedefinitionsareillustratedinthefigure.Theheightofthecomponenttreedependsonthenumberofgraylevels,andthebranchingstructuredependsonthedefinitionofconnectivity.

Standard definitions include four- or eight-connectivity,althoughtheconcepthasbeengeneralizedintheimagepro-cessingliterature.6Twopixelsarefour-connectedifandonlyiftheyhaveacommonedge,andareeight-connectediftheyhave a common edge or vertex. Since four-connectivity isthemorestringentcondition,thefour-connectedcomponenttreeofagivenimageisusuallymorecomplexthantheeight-connectedcomponenttreeofthesameimage.Eachnodehasanassociatedgraylevelofitscorrespondingcomponent,thelocationofarepresentativepixel,andasetofattributeschar-acterizing the component. The representative pixel locatesthe component in the image, and the gray level value is asimplefeaturedescribingthecomponent.Ingeneral,thenodedescriptionmayconsistofavectorf=(f1,…,fn)ofattributes,orfeatures,thatcapturesalientinformationabouttheassoci-atedcomponent.Besidesarea,typicalfeaturessuchasperim-eter,compactness,aspectratio,andconvexitymaybeuseful.

Thecomponent tree supportsaclassofnonlinearfilters,calledconnectedfilters,thatpreserveonlythosecomponentssatisfying a given criterion based on the component’s attri-butes.Thecriterionmay,forexample,requiretheseattributestoliewithinspecifiedlimits.Aconnectedfilterneverintro-ducesnewcomponentstotheimage,andifanimagecompo-nentispreservedbyoneapplicationofthefilteritispreservedbyanynumberofrepeatedapplications.Inmathematicalmor-phology,afiltersatisfyingthesepropertiesiscalledathinning.

If,inaddition,thefilterisincreasing(i.e.,wheneveracompo-nentCsatisfiesthecriterion,sodoesanysuperset),thefilterisanopening.Connectedcomponentfilterseitherpreserveacomponentwithoutalterationorremoveitentirely,sothereis no distortion as occurs with openings and closings usingstructureelements7—adesirablepropertyforsegmentation.

Componenttreefilteringisadecisionprocessthatclassi-fiesnodesasactiveorinactive.Anodeisactiveifandonlyifthefilterpreservesitsassociatedcomponent.Aconnectedfilter that ignores the links is called flat and can be imple-mentedbyclassifyingeachnodeasactiveorinactive.Ifthefiltercriterionisincreasing,everynodeaboveanactivenodeisalsoactive;inthiscase,therefore,thefiltercanbeimple-mentedbystartingattherootandmovingupthetree,branchbybranch.Onceaninactivenodehasbeenfound,itisunnec-essarytocheckhighernodes.Thisincreasingpropertyfacili-tatesrapidimplementationofthistypeoffilter.Anexampleofaflatfiltercriterion is the following:The area of the node is greater than 12. A second type of connected filter, calleda signaturefilter, applies to an entirebranch.Theattributesignatureofabranchisasequenceofnodeattributes.Iftheattributesignaturesatisfiesaspecifiedcriterion,thebranchisdeclaredactive.Forexample,thefollowingisasignaturefiltercriterion: In order to be active, a branch must contain a node whose corresponding connected component has area between 10 and 40 pixels and eccentricity between 0.4 and 0.6.

Segmentinganimageconsistsofsubtractingthefilteredimage from the original. Thus, the filter criterion must bechosensoastoinactivatenodesassociatedwithtargetimagecomponents.Oneway to implement thisprocess is tofirstapply a signature filter to activate all branches containingno target-like components. For instance, a branch wouldbeactivatedifitcontainsnonodeswhoseconnectedcom-ponent is convex and has an area within expected targetbounds.Iftheobjectiveistolocatethemovingtankshowninthefigure,anybranchwhosegray-levelstructuredoesnotindicate a hot engine would also be activated. This may,of course, require somegenerala priori knowledge, suchastargettypeandanestimateoftherangetogo.

Next, a flat filter can be applied to determine whichnodesintheremainingbranchestoactivate.Inthefigure,



candetectavarietyofpotentialtargetsindiverseback-grounds. The algorithm requires no user interventionor inputduring itsexecutionanddoesnotdependonalibraryofpredefinedtargets.Itisdesignedtoidentifyman-made objects, structures, and vehicles in a fullyautomated manner using a single frame of passive IRimagery.

To achieve this objective, the algorithm is com-posedofninestepsorsubalgorithms,wheretheoutputofoneservesas inputtothenext.Theprocessstartswithanoise-reductionfilter(Step1)andthenStep2enhancestheedgesforthedetectionprocess(Step3).The adaptive edge segmentation procedure (Step 4)binarizestheedgedetectorresult,whilealsoperform-ingasizediscriminationprocess(Step5)whereobjectsthat are too small or too large to be the target are

filtered out. In Step 6, the remaining objects arethinnedtoproduceacontourof1pixelwidth.Achaincode algorithm (Step 7) then strings contour pixelstogetherforanalysis.InStep8,theshapediscrimina-tionalgorithmidentifiesobjectsbydesiredshapeattri-butes.Finally,Step9appliesatargetlikelihoodindextorankthepotentialtargetsaccordingtotheirshapeattributes. A more detailed discussion of the edge-basedprocessingstepsfollows.

Thenoisereductionalgorithm(Step1)isasimplemedianfilterthatreducesrandomnoiseandeliminatesdead pixels. The advantage of the median filter overmanyothernoiseremovalfiltersisthatitpreservestheedgecontrastwhileeliminatingnoiseconsiderably.Theedgeenhancementalgorithm(Step2)utilizesgradientandvarianceinformation.Ateachpixeltheenhanced

Part (a) is a schematic quantized image of a tank with a hot engine. The image in (b) shows the connected components. Each color corresponds to an intensity level from cold (deep blue) to hot (red). The leaves in the component tree (c) correspond to the regional maxima; each leaf determines a branch terminating at the root node (the image plane shown in dark blue). Part (d) shows the result of filtering the image as discussed in the text.

forexample, thetankisonaroadthat isclearlynotpartof the tank. All nodes up to and including the one cor-responding to the road can be activated by filtering withthe criterion that activates nodes whose area is greater

than the expected area of the tank. In the figure, thegunbarreliscomposedofseveralregionalmaxima,which,althoughpartofthetank,wouldbeinactivatedbythisfil-teringscheme.

6

5

4

3

2

1

0

Leve

ls

(a) (b)

(c) (d)

f = (f1, . . ., fn)


D. E. MAURER et al.

valueiscomputedusingthelocalintensitygradientandlocalvarianceinhorizontalandverticaldirections,pro-ducingtwoimagesthathighlightedgesintwoorthogo-naldirections.

Step3,edgedetection,processestheedge-enhancedimageandyieldsabinaryimagewhereonlysignificantedgepixelsare retained.This stepfirst removespixelswhere the intensity variations are negligible, such asthoseassociatedwithbackgroundnoise.Thentheedgedetectionalgorithmidentifieslocalmaximumpixelsontheedgesofeachimagebycomparingeachpixeltoitsimmediateneighborhoodandsavingonly thosepixelswhosevaluesarelargerthantheirneighbors.Thefinalresultisabinaryimage,B,obtainedasthelogicalORofthehorizontalandverticaldirectionresults,whereedgepixelsaresettothevalue1.

Whilethefirstthreestepsidentifymostedgepixels,they do not guarantee a closed contour around eachobject. Therefore, Step 4 is an adaptive edge-fillingalgorithm that uses both the binary image of edgedetectionandtheoriginalnoise-filteredimagefromthefirststep.Theedgediscontinuitiesinthebinaryedgeimagearefilledbyinspectingeachpixelwithavalueof0andconvertingittoa1ifitfulfillsdesiredneigh-borhoodcriteria.Thefinalresultisanedge-thickenedversionofB inwhichthecontoursareconnected.Alabelingalgorithmassignsauniquevaluetoallpixelsthatformthecontourofanobject.Thisstepprovidesa means to distinguish distinct objects that will sub-sequentlybeanalyzed separately todetermine if theyare likelytobetargets.Step5, sizediscrimination, isimplementedusingthenumberofpixelsineachcon-tour.Atthisstepweassumesomea prioriknowledgeof the target type and that an estimate of the rangetogoisavailable(e.g.,fromGPSorpassiveranging).Smallobjectsthathavefewerpixelsthanafirstthresh-old and very large objects that have more pixelsthanasecondthresholdareremoved.Inthismannermanybackgroundobjectsarediscardedbeforefeatureextraction.

TheremainingobjectsarethinnedinStep6usingastandardthinningalgorithmthatretainsthemedialaxisof thecontoursbysettingmostpixels tozerodepend-ingonneighborhoodcriteria.Thefinalresultisasetofcontours,1pixelinwidth.Toanalyzetheshapeoftheresultingcontours,eachoneisrepresentedwithaone-dimensionalarrayusingthechaincodealgorithm(Step7).Thisencodesthetransitionbetweenadjacentpixelsofthecontourwithauniquecodeforeachoftheeightpossibledirections.

Theshapediscriminationalgorithm(Step8)removesanyremainingnontargetobjectsusing thechaincodeinformation.Theprincipaldiscriminator inourinitialstudy was straight edges. Man-made objects such astanks, bunkers, bridges, and planes are characterizedbystraightedges,whereasmostnaturalobjectstendto

have a more arbitrary appearance. Furthermore, man-madeobjectstendtohaveadjacentstraightedges.Thechain codes were analyzed to find sequences wherethedirectionalvalueremainedconstantoveraprede-termined number of pixels. This analysis eliminatedbackground objects that may have contained a fewpixelsinaline.Moreover,sincewecannotexpectper-fectlystraightedges,weallowforsomevariationfroma straight line. Typically, a sequence of pixels havingthe same direction, except for one or two values, isassumedtobeastraightedge.Thealgorithmalsoiden-tifieslinearsequencesthatadjoinoneanothertoformstraightedges.Thisinformationisusedtocalculatethetarget likelihood index (Step 9). The index is higherwhenthecandidatecontourcontainsmoreadjoiningorconnected straight edges. Potential targets are rankedwithrespecttothisindex.

ALGORITHM TIMING STUDIESThe ability to implement these algorithms is as

important as developing them. To accomplish thisgoal,twodifferent—yetcomplementary—technologiesare being investigated: microprocessors and field-programmable gate arrays (FPGA). These are fun-damentally different devices that require differentprogrammingparadigmstogetgoodresults.Themicro-processor has a fixed architecture (or configuration)thatperformsasetofoperationsasdeterminedbythemanufacturer(e.g.,Intel,Motorola)duringthedesignprocess.Moderndesignsusesuper-scalararchitecturesand instruction pipelines to increase performance bytaking advantage of fine-grain parallelism in a shortsequenceofoperations.TheFPGAhasanarchitecturethat is determined at programming time by the enduser.Thisprocessallowspipelinelengthsandcompu-tationunitstobedeterminedbythealgorithmsbeingimplemented.

To determine which technology is truly best fora particular algorithm, the algorithm must be imple-mentedanditsexecutiontimescompared.Thisisnotalways practical, however. A rule of thumb is thatoperations that do not lend themselves to pipelineoperations or that are not data-independent tend toworkbetterinamicroprocessor.Thereasonisthedispa-rateclockspeedsbetweenFPGAandmicroprocessors.FPGAmanufacturersarenowboastingclockspeedsashighas300MHz.However,inpracticethebestspeedsachievedareabouthalfthemaximumofmanufacturerclaims.TheFPGAfunctionswellwithalgorithmsthatcanbedecomposedintomanysmallindependentcom-putationsbecauseoftheconfigurablearchitecture.

Thetwomainchallengesinimplementingthealgo-rithms so far have been avoiding repeated memoryscans and adapting some nonlinear operations tovector-style processing. Because of the size of the



imagesbeingprocessed,therawimagedataandtem-porarymemoryareasdonotstayintheL1orL2cacheofamicroprocessor.Consequently,repeateddatascanscan require relatively long timesbecauseof the timerequired to access main memory. Also, pieces of thealgorithms with strong data interdependence cannotbemadetotakeadvantageofspecialpropertiesofanFPGAormodernmicroprocessors.

ThemicroprocessorcurrentlybeingusedisaMotor-olaPowerPC7400withtheAltiVec™processingunit(aspecialvectorprocessoraddedbyMotorolatoaccel-eratedigitalsignalprocessor[DSP]andimageprocess-ing applications). The PowerPC 7400 is on a PowerMacintoshG4 AGP motherboardwith a 1-Mb half-speed L2 cache and runs at 400 MHz. Early bench-marksshowedthatthePowerPC7400hadmoreprom-iseofbeingadequatethanothermicroprocessorsandDSPs.Infact,itwas4to5timesfasterwithoutusingthe AltiVec unit than the DSP initially being used.TheAltiVecunithasincreasedtheprocessingspeedsof several algorithmsbya factorof2ormore—evenbeforeoptimization.

Ascanbe seen inFig.6, thealgorithms thathavebeenbenchmarkedexecuteinveryshortperiodsoftime,withtheexceptionofthemultiblob.Themostimpor-tant part of reducing execution time has been smartimplementation. Algorithms need to be structured sothat repeated memory accesses are not necessary andsothatcomputationscanbeeasilypipelined.Todate,however, the timing studies are incomplete and agreatdealofworkstillneedstobedone.Althoughthefindingspresentedherearepreliminary,theydorepre-senttheresultofoptimalimplementationsonthePow-erPC7400.

RADAR SITE SCENARIOAninitialdemonstrationofthedetectionalgorithms

was carried out using a sequence of 120 frames simu-lating action against an air defense unit, as shown inFig.7a.Theradarantennawastheprimarytarget.Afteracquisition, a correlation tracker was initiated to seeif the antenna could be tracked until the end of thesequence.The seekerbegan functioningatanominalrange-to-targetof600m.Figure7showsthestages inapplying region-basedandedge-basedprocessing.Themultithreshold applied to the enhanced image histo-gramproduces thequantized image shown inFig. 7b,havingsevengraylevels.

Thefollowingcriterionwasusedtofiltertheassoci-atedcomponenttree:anodeisactiveifandonlyifitsareaislessthan20pixels,greaterthan1,000pixels,oritsgraylevelislessthan4.Componentsremainingafterthisfilteringstepwereconsideredtarget-like.Thiscri-terionwassufficienttoeliminateallnaturalbackgroundfeatures in the image. The remaining objects includethe radar antenna (ROI 1) and some very hot reflec-tions from buildings (ROIs 2 and 3). The buildingsthemselveswerenotdetectedbytheregion-basedpro-cessing because they were connected to a large com-ponentthatalsocontainedbackgroundandwaselimi-natedbythefilter’ssizecriterion.

Figures7dthrough7f showsteps intheedge-basedprocessing. Many of the edges and contours that canbeseeninFig.7eareeliminatedwhenthebackgroundthreshold is applied. Figure 7g shows the final resultof combining the region- and edge-based techniques.Additional filtering of the segmented ROI in Fig.7cusing criteria involving commongeometric image-processingfeatures,suchasconvexityandeccentricity,eliminatedthehotreflectionsfromthebuildings.

Table1comparesthevaluesofseveralofthesefea-turesforeachofthethreeROIs(1,2,and3,fromlefttoright).Inallcasesthevaluesfortheantenna(object1)aremarkedlydifferentthanthosefortheremainingtwoobjects.Thus,iftheradarantennaistheintendedtarget in this image, a criterionbasedon its intensityandgeometricstructurewouldbeabletodiscriminatebetweenitandthebackground.Fortheedge-basedpro-cessing,thinningandsizeandshapediscriminationareabletoremoveallobjectsexceptthetwodarkbuildings.The clutter background has been eliminated, and weareleftwiththedesiredtargets.Thiswasaccomplishedwithnodirectuserinputandnodependenceontargetstructureapartfromsomemildassumptionsconcerningtargetsizeandintensity,andthatthetargetiscomposedofstraightedgesandcorners.

TheimportanceofthefusionstepcanbeseeninFig.7g.Theedge-processingalgorithmdetectedthemajorbuildings, but not the radar antenna. By combiningthecomplementary informationfrombothregionand

Co-histogram

Median filter

Line gradient

Line threshold

Local areathreshold

Multiblob

0 10 20 30 40 50 620

613.4

Execution time (s)

PowerPC 7400

FPGA

Figure 6. Preliminary algorithm execution times.


D. E. MAURER et al.

(a) (b) (c)

(d) (e) (f)

(g)

Region-based processing

Edge-based processing

Combined

Figure 7. Detection processing steps: (a) original image, (b) quantized image, (c) seg-mented image (enlarged), (d) median filtered image, (e) edge-enhanced image, (f) edge-segmented image, (g) fused object set.

Theradarantennawasdetectedinasingle frame,and two additional frames were used to verify thedetection.Once this validationwas complete, track-ing began. As shown in Figs. 8a and 8b, the imageframe inwhich the targethasbeendetected isbina-rized, and a small region centered at the estimatedtargetcentroidisusedasatemplate.Eachsubsequentimageframeisbinarizedandcorrelated,withthetem-plateconstructedfromthepreviousimageinordertoupdatethetargetlocation(Fig.8cisatypicalcorrela-tionsurface).Thenthetemplateisupdatedtoaccountforchangesintheimageduetotheseeker’smotion.Inthiswaytheantennawastrackeduntilthelastframeinthesequence(Fig.8d).

Although the algorithms have not been tested onan extensive set of images, this example is typical oftheperformanceobtainedontherealimagerywehaveexaminedsofar.Thegroundclutter,whilenotextreme,

Table 1. Feature characteristics for ROIs in the segmented image.

ObjectFeature 1 2 3

Area(pixels) 26 49 40Perimeter(pixels) 15 29 25Compactness 8.65 20.02 15.63Aspectratio 0.69 1.78 1.82Convexity 1.93 0.76 0.86

edgeprocessing,amorecomplete setof targetobjectsisobtained.Thus,byfusingtheoutputsofthesecom-plementary processors, the discrimination power ofmoresophisticatedimageprocessingtechniquescanbeattainedbysimplehigh-throughputprocessorsworkinginparallel.



istypicalofadesertenvironmentandconsistsprimarilyofcreosotebushesthatcanhaveintenseIRsignatures.

CONCLUSIONSThisarticleprovidesanoverviewofthedesiredsystemcapabilitiesofthe

LCGLSandsummarizesworkdoneattheLaboratorytoaddressdesigncon-straintsandassesssystemfeasibility.Thepreliminarydesignincludes

• Body-mountedopticsandlensmaterialsthatcansurvivethehighgforceatlaunch

• AnuncooledbolometricFPAandlightweightopticstoreducespaceandweightrequirements

• UsingtheLWIRspectralbandtominimizeatmosphericdegradation

Inademonstrationusingasequenceofrealimagerytosimulateclosingonaradarsite,thetargetdetectionsystemwasabletoacquireandtrackthedesignated target.Moreover, algorithms thathavebeenbenchmarkedaregenerallyfast.

Template(15 20 pixels)

Antenna Building

(a) (b)

(c) (d)

Figure 8. Correlation and tracking: (a) template of a region of the binarized raw image centered on the estimated target centroid, (b) subsequent binarized and correlated image frame, (c) typical correlation surface, (d) updated template and image resulting from the seeker’s motion.

Whiletheseresultsarepromising,morein-depthanalysesarerequired.Thealgorithmsneedfurtherdevelop-menttooptimizedetection,segmen-tation, and classification for speedandcomputationalefficiency.Asetof features that are robust against awidevarietyoftargetandimagecon-ditionsandbackgroundsmuststillbedefined.Moreover,anautomaticpro-cedure must be developed for com-bining the region- and edge-basedalgorithms, or fusing their outputs.Finally,theperformanceofthesystemneeds tobevalidated inend-to-endflightsimulationstudies.Ifsuccessful,suchatechnologycouldbe insertedintoawiderangeofweapons.

REFERENCES 1Carson, J.C.,Shostak,D.R.,andLudwig,

D. E., “Smart Sensors Integrate DetectorsandProcessors,”Laser Focus World,113–115(Jul1996).

2Herring,R. J., andHoward,P.E., “DesignandPerformanceof theULTRA3203240Uncooled Focal Plane Array and Sensor,”SPIE2746,2–12(1996).

3Holst, G. C., “Target Discrimination,”Chap. 2, in Electro-Optical Imaging System Performance,JCDPublishing,WinterPark,FL,pp.412–440(1995).

4Rounds, E. M., and Sutty, G., “Segmen-tation Based on Second-Order Statistics,”SPIE205,126–132(1979).

5Jones,R.,“ConnectedFilteringandSegmen-tation Using Component Trees,” Comput. Vis. Image Underst.75(3),215–228(1999).

6Goutsias, J., et al. (eds.), “ComputationalImagingandVision,”Vol.18ofMathemati-cal Morphology and Its Applications to Image and Signal Processing,Kluwar,TheNether-lands(2000).

7Weeks,A.R.,“FundamentalsofElectronicImage Processing,” SPIE/IEEE Series on Imaging Science & Engineering,pp.294–386(1996).

ACKNOWLEDGMENTS: The authors wish tothankRogerHorman,JohnFraysse,andDeanZabelof the NSWC/DD for their sponsorship, guidance,andsupportofthiseffort.Wealsowishtoacknowl-edgethecontributionsofMikeElko,KevinBaldwin,CraigMitchell,LindaHowser,andBillGreen.


D. E. MAURER et al.

THE AUTHORS

DONALDE.MAURERisamemberoftheSeniorProfessionalStaffinAPL’sAirDefenseSystemsDepartment.HeholdsaB.A.fromtheUniversityofColoradoandaPh.D.fromtheCaliforniaInstituteofTechnology.BeforejoiningAPLin1982,heheldpositionsattheInstitute forDefenseAnalysesandtheCenter forNavalAnalyses.AtAPL,Dr.Maurerhasbeeninvolvedinmodeling,simulation,andalgo-rithmdevelopmentinsupportofvariousprograms.HewasthePrincipalInvestiga-torforanIR&Dprogramindistributeddatafusion,wherehedevelopedametrictomeasurethesimilaritybetweenoceansurveillancetrackpictures.AstechnicalleadfortheHigh-SpeedAnti-radiationMissileReceiverSensitivityImprovementProgram,hedevelopedacquisitionandtargetdetectionalgorithms.Currently,heisinvolvedindevelopingandevaluatingradar-to-IRhandoveralgorithmsinsup-portoftheNavyTheater-Wideprogram.Dr.MaurerisareviewerfortheMath-ematical ReviewsandisamemberoftheAmericanMathematicalSociety,theNewYorkAcademyofSciences,andtheSocietyoftheSigmaXi.Hise-mailaddressisdonald.maurer@jhuapl.edu.

ERIC W. ROGALA, a member of APL’s Senior Professional Staff, is with theElectro-Optical Systems Group in ADSD. He received B.S. and M.S. degrees inopticalsciencesfromtheUniversityofRochesterin1990and1992,respectively.In1999hereceivedaPh.D.inopticalengineeringfromtheUniversityofArizo-na’sOpticalSciencesCenterforhistheoreticalandexperimentalworkonanovelphase-shiftinginterferometercapableofmappingtheprofileandcomplexindexofrefractionofasurface.InMarch2000Dr.RogalajoinedADSD’sElectro-OpticalSystemsGroupandhasbeeninvolvedindesigningandanalyzingopticalsystemsaswellasdevelopingpatternrecognitionsoftware.Hiscurrentinterestsincludeauto-mated target recognition in passive IR imagery, automated analysis of solar flareimages,boresighterroranalysis insapphiredomeforSM-2BlockIVA,modelingof3-Dlaserradarimages,andphase-shiftinginterferometryofbiologicalcells.Hise-mailaddressiseric.rogala@jhuapl.edu.

ISAAC N. BANKMAN, a member of the APL Principal Professional Staff, isSupervisorof the ImagingandLaserSystemsSection inADSD’sElectro-OpticalSystemsGroup.HereceivedaB.S.inelectricalengineeringfromBosphorusUniver-sity,Turkey,in1977,anM.S.inelectronicsfromtheUniversityofWales,Britain,in1979,andaPh.D.inbiomedicalengineeringfromTechnionUniversity,Israel,in1985.He joinedAPL in1990andworkedon signalprocessingalgorithms fortransientdetection,imageprocessingalgorithmsforobjectrecognition,andimageregistrationalgorithms.Inthepast6yearshehasdevelopedanalyticalmodelsofladarandradarsignaturesofballisticmissilepartsandrelateddiscriminationalgorithms.His recent work also includes ladar sensing and laser communication in marineapplications,imageanalysisformissileguidance,andradarandpassiveIRfirecon-trolsystemsforshipself-defense.Hise-mailaddressisisaac.bankman@jhuapl.edu.



BRADLEYG.BOONEisaPrincipalProfessionalStaffphysicistintheRFEngineeringGroupoftheAPLSpaceDepartment.AfterearninghisPh.D.inphysicsfromtheUniversityofVirginiain1977,hejoinedAPLandbecameinvolvedinavarietyofadvancedmissileguidanceprojectsandwasPrincipalInvestigatoronnumerousIR&Dprojectsinopticalsignalprocessing,superconductingelectronics,andpatternrecognition.Afterservingasasectionsupervisor(1983–1996),hewasgroupsupervisorfortheElectro-OpticalSystemsGroup from1997 to2000.Hehaspublishedover30 technicalpapers andonebook, andholds fourU.S.patents.HehastaughtextensivelyfortheG.W.C.WhitingSchoolofEngineeringandwasvisitingprofessorintheElectricalandComputerEngineeringDepartmentin1990–1991.Heiscurrentlyworkingonopticalcommunicationsfordeepspaceapplications.Hise-mailaddressisbrad.boone@jhuapl.edu.

KATHRYNK.VOGELisaSeniorProfessionalStaffengineerintheElectro-Opti-calSystemsGroupof theAirDefenseSystemsDepartment.She received aB.S.degreefromDrexelUniversityin1983andanM.S.degreefromTheJohnsHop-kinsUniversityin1988,bothinelectricalengineering.SincejoiningAPLin1983,Ms. Vogel has worked on the test, evaluation, and simulation of infrared (IR)andinertial systemsprimarily formissileguidance.Herexperienceincludescoor-dinatingIRseekertestactivities,analysisoflaboratoryandflightdata,seekerper-formanceevaluation,inertialcomponenttesting,andmodelingandsimulationofIR seeker dynamics and inertial systems. Currently, she is working on the testandevaluationof theStandardMissile-3KineticWarhead.Here-mailaddress [email protected].

CHRISTOPHERPARRISreceivedaBachelor’sdegreeinelectricalengineeringfromTennesseeTechnologi-calUniversity in1995.AfterearninghisMaster’sdegree in1997,hewasemployedby theNavalSurfaceWarfareCenterinDahlgren,Virginia,workingintheAdvancedSignalProcessorsGroupintheAreaSystemsBranch.Hismaininterestishigh-speeddigitalsignalprocessorsforradarandimageprocessing.Hise-mailaddressisparriscp@nswc.navy.mil.

Documents

A Low Cost Gun Launched Seeker Concept Design for Naval ... › Content › techdigest › pdf › V22-N04 › 22-04-Maurer.pdfForward Air Support-Marine-Gun Launched Loitering Munition,