A Networked Mobile Sensor Test-bed for Collaborative Multi-target Tracking Applications

Transcription

1 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. A Networed Moble Sensor Test-bed for Collaboratve Mult-target Tracng Applcatons Subr Bswas, Sonn Gupta, Fan Yu, and Tao Wu Electrcal and Computer Engneerng Department Mchgan State Unverst Abstract Ths paper presents the desgn, archtecture, mplementaton, and epermental results from a networed moble sensor test-bed developed for collaboratve sensor tracng applcatons. The test-bed comprses a fleet of networed moble sensors, an ndoor localzaton sstem, a control, debuggng and management nfrastructure, and a tered wreless ad hoc networ for seamless ntegraton of the above three components and the estng wreless nfrastructure. Frst, the software and hardware archtectural detals of a Swarm Capable Autonomous Vehcle SCAV sstem for our collaboratve applcatons are presented. Second, the detals of an ndoor self-localzaton and Kalman flter based navgaton sstem desgn for the SCAV platform are presented. Thrd, as an eample multsensor applcaton, a collaboratve mult-target tracng problem and a heurstcs-based networed soluton are formulated. Fnall, the performance of the collaboratve tracng framewor s evaluated on the laborator test-bed for characterzng the mpacts of localzaton and navgaton errors on the dstrbuted tracng performance. The epermental stud also characterzes the tradeoff between the tracng performance and the consumed wreless bandwdth. The epermental results demonstrate a number of counterntutve results due to varous errors n sensor localzaton and navgaton.. Kewords Moble Sensors, Sensor Networ, Collaboratve Applcatons, Agent Swarmng, Mult-target Tracng, Indoor Localzaton, Sensor Navgaton, Kalman Flter. Introducton Collaboratve moble sensng s ganng ncreasng populart as an enablng technolog for a wde range of sensng applcatons ncludng envronmental montorng, securt survellance, and target tracng. Target tracng s a prevalent mltar requrement for survellance and reconnassance scenaros Correspondng author sbswas@egr.msu.edu; Ths wor was partall supported b grants from Natonal Scence Foundaton CMMI , Ar Force Research Laborator, and Mchgan State Unverst s Hgh Assurance Computng Intatve. Ths author was wth the Networed Embedded and Wreless Sstems NeEWS laborator at MSU durng ths wor.

2 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. ncludng those n urban battlefelds and nsde large buldngs and publc places such as arports and tran statons. In recent ears, the technolog of unmanned sensng and tracng [] has eperenced a notceable shft from usng few epensve and feature-rch autonomous sensors, to deplong swarmng fleets of large number of relatvel nepensve and smaller autonomous sensors that are wrelessl networed. The advantages of the latter approach are as follows. Frst, a networed fleet provdes a more robust soluton n whch the loss of few nepensve sensors does not sgnfcantl degrade the overall sensng and tracng capabltes. Second, as outlned n [] and [3], collaboratve tracng can be desgned to be more effcent than sngle-sensor tracng, especall n the presence of multple moble targets. The need for such swarmng approach to enable collaboratve sensng can be further motvated b observng how a large number of organsms n the nature use swarm dnamcs for chasng and evadng pras and predators. Schoolng fshes, for eample, swarm awa from a predator usng group coordnaton that reles on smple localzed communcaton. A pac of cheetah uses group coordnaton to collaboratvel trac and tacle a pre. Smlarl, a swarm of bees uses collaboratve mechansms to localze, trac, and chase ntruders usng group swarmng. Successful deploment of a collaboratve moble sensor sstem to cater to such applcatons wll requre the followng crtcal sstem components. Frst, a sensor platform wth robust sensng, moblt, and networng capabltes needs to be developed. Second, self-localzaton abltes wll be needed for target tracng. Thrd, smooth sensor navgaton mechansms wll have to be developed n the presence of localzaton errors and naccuraces at hgh platform speeds. Fourth, self-healng moble ad hoc wreless networ protocols wll be needed for nter-sensor data dssemnaton. Fnall, mult-sensor collaboratve applcatons wll have to be developed b leveragng the underlng localzaton, navgaton, networng and sensng servces. Each of these four areas has seen recent research actvtes [4-] and progress towards applcaton-drven sstem ntegraton. Ths paper contrbutes towards these ntegraton trends b developng a sstem that comprses all four subsstems as well as ther functonal ntegraton n the contet of a collaboratve mult-target tracng applcaton n ndoor settngs. The contrbutons of the paper are as follows. Frst, our eperence on desgn and development of the above four subsstems are reported n the contet of a laborator prototpe sstem. Second, a dstrbuted mult-target tracng framewor usng the above subsstems have been developed and etensvel

3 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. characterzed n the presence of varous measurement errors. Thrd, specfc algorthmc nsghts are provded as to how adaptng tracng can be mplemented to cope wth non-deal localzaton and networng condtons. It s demonstrated that such non-deal condtons can cause theoretcall sound tracng algorthms to generate unepected results, and therefore specal epermental consderatons wll be necessar whle mplementng such moble sensor sstems under non-deal condtons. The rest of the paper s organzed as follows. Secton descrbes the related wor on collaboratve mult-target tracng applcatons. The sstem archtecture s presented n Secton 3. The hardware and software sstem components of the moble sensor platform are presented n Secton 4. Secton 5 detals an ndoor localzaton framewor used b the moble sensors for self localzaton, and a fltered navgaton framewor s presented n Secton 6. Desgn of a collaboratve tracng algorthm and ts mplementaton and performance usng the proposed test-bed are presented n Secton 7. Fnall, the paper s concluded n Secton 8 wth a summar and a lst of ongong wor on ths topc.. Related Wor Indoor sensor localzaton [0, 3] for movng devces can be performed n an actve or n a passve mode. In the actve approach, each moble devce has an actve transmtter, whch perodcall broadcasts beacons to a number of eternal fed unts. The estmated dstance of the moble devce from each such fed unt s fed bac wrelessl to the moble devce, whch then computes ts own coordnate usng the locatons of the fed unts and ndvdual dstances. In the passve approach, t s the eternal fed unts that broadcast the beacons and the moble devce computes ts own coordnate based on the estmates of the ndvdual dstances. The mpacts of the denst and number of eternal unts have been analzed n []. In ths paper we use the passve approach wth a Kalman Flter [4] based estmaton for compensatng the localzaton errors caused due to movng sensors. Ths approach s smlar to what has been taen n [3] to our nowledge, that s the onl other publshed ndoor localzaton and sensor navgaton wor ecept that, unle n [3] whch uses an Etended Kalman Flter EKF, we have used a regular Kalman Flter wth an assumpton of lneart for the sensor dnamcs. Epermentall t s demonstrated that the lnear KF wth PVA poston, veloct, acceleraton state abstracton was suffcent for accurate 3

4 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. estmatons n the presence of localzaton errors. Target tracng usng sensor networs has been etensvel eplored [3, 5, 6, 7, 30] n the recent lterature. Both sngle- [9, 30] and mult-target [5, 8, 3, 3] tracng applcatons were nvestgated. The mechansm n [5] eplores sensor networ based mult-target tracng solutons and ther performance n the presence of non-deal condtons due to pacet loss, communcaton dela varaton, and false postve sensor detecton. These and the algorthms presented n [8, 3, 3] are desgned for sstems operatng n outdoor felds that are larger than the vehcle dmensons b several orders of magntude. The feld dmensons are also few order of magntudes larger compared to the GPS based locaton estmaton errors, whch could be up to few meters [6]. In contrast, n ths paper we nvestgate the mpacts of locaton and navgaton errors on mult-target tracng performance n a ver tght ndoor envronment.e. an area of few square meters that s not as large compared to the vehcle dmensons.e. appromatel 5cm for the moble sensor used n our test-bed and the ndoor localzaton errors, whch have an upper bound of 0cm. Such tghter feld dmensons e.g. an area of 3m 3m has been used for our epermentatons gve rse to unque target tracng ssues whch are dealt wth n ths paper. 3. Moble Sensor Test-bed and Developed Servces The developed moble sensor test-bed and ts ntegraton wth the estng wreless networ nfrastructure are shown n Fgure. The test-bed contans the followng man components: a fleet of ndoor moble sensors, an ndoor localzaton sstem for enablng sensor self-localzaton n GPSdened envronments, 3 a control, debuggng and management nfrastructure, and 4 a tered wreless ad hoc networ for seamless ntegraton of the above three components and the estng research nfrastructure n varous wreless networ laboratores n Mchgan State Unverst. As for the networed moble sensors, we have developed a Swarm Capable Autonomous Vehcle SCAV platform as descrbed n Secton 4. 4

5 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. Sensor Networ and Mddleware Testbed n MSU NeEWS Lab Control Debug and Management Sstem CDMS Sensor Networ n Kellogg Bologcal Staton, MSU Localzaton Infrastructure n GPS-dened Indoor Laboratores Tered Wreless Bacbone Indoor Moble Sensor Testbed Fg. : Components of the moble sensor testbed and ther ntegraton Developed moble sensor servces and ther dependences are depcted n Fgure. The SCAV moble sensor sstem comprses the followng four servces: a wheel or trac based locomoton, b multmodal onboard sensng, c both ad hoc and access pont orented wreless networng, and d nfrastructure asssted ndoor localzaton usng Rado Frequenc RF and Ultra Sound US sgnals. Self localzaton for the moble sensor nodes wthn an n-buldng coordnate sstem have been acheved usng a Tme Dfference of Arrval TDOA based rangng technque [], coupled wth collaboratve mult-lateraton [3]. A passve localzaton scheme has been adopted so that a SCAV can self-localze b usng the receved RF and US sgnals from prenstalled localzaton nfrastructure. A pont-to-pont navgaton servce s developed usng the localzaton and the SCAV locomoton servces. A framewor of Baesan flterng has been ncorporated to tacle the ntrnsc localzaton errors contrbuted b noses such as ambent RF and US nterference, vehcle vbraton, and naccuraces ntroduced b the localzaton hardware components. Addtonal naccuraces are caused when the localzaton ntervals are too large to capture accurate locatons of the movng sensors. A Poston- 5

6 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. Veloct-Acceleraton PVA lnear Kalman Flter [4] has been used for smooth navgaton n the presence of those localzaton errors. Collaboratve Moble Sensor Applcatons e.g. Networed Mult-target Tracng Fltered Navgaton Pont-to-Pont Navgaton usng Baesan Flter Framewor Networ & Sensng Locomoton Localzaton Self Localzaton RF and Ultrasonc based Localzaton SCAV Sstem Archtecture Networng, Sensng, Locomoton and Localzaton Indoor Localzaton Infrastructure Fg. : Sstem components of the collaboratve tracng usng SCAVs Fnall, a set of collaboratve sensor applcatons have been developed usng the ad hoc networng and sensng servces from the SCAV archtecture, and the smooth navgaton servce s mplemented usng the Kalman flter. As a representatve applcaton, a dstrbuted mult-target tracng sstem and ts varatons nvolvng dfferent numbers of traced and tracng sensors have been presented n ths paper. Throughout the rest of the paper, these moble sensor servces components wll be further elaborated. 4. Swarm Capable Autonomous Vehcle SCAV Archtecture The pcture of a moble SCAV sensor unt s shown n Fgure 3. The mddle pcture shows the frontvew of a fleet of four moble sensors, and the left-most pcture shows the bac-vew of a sngle sensor wthout ts collson bumper. The followng features are currentl avalable [5] n our moble SCAV fleet. Self-localzaton wth centmeter level resoluton, Baesan flter based pont-to-pont navgaton, 3 Mult-modal onboard sensng, 4 Navgatonal collson avodance usng nfra-red obstructon sensng, 5 Both Ad Hoc and nfrastructure based rado communcaton usng 900 MHz RF, and 6 Control, Debuggng and Management Sstem CDMS. 6

7 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. The CDMS software sstem s desgned for managng the SCAV nfrastructure locall or through a networ as shown n Fgure. The CDMS sstem runs on a Wndows sstem and supports the followng essental operatons for the SCAV test-bed: compled mage and software download usng wreless lns, SCAV sstem debuggng b allowng remote prnt console, 3 command lne and graphcal nterfaces for remote command dspatch to targeted moble sensors, 4 supportng a varet of sensor and moble ad hoc networ protocols for seamless connectvt wth the SCAV test-bed and the bacbone mesh networ, 5 real-tme locaton tracng for ndvdual sensors and ther epermental postprocessng, 6 command drven SCAV navgaton, 7 sensor specfc moblt plannng through a graphcal user nterface, and 8 data upload and telemetr from the SCAVs usng wreless lns. Real-tme locaton tracng screenshot from an eample -SCAV leader-followng eperment s shown n Fgure 3. Phscal appearance Swarm of four sensors CDMS Screenshot Fg. 3: SCAV moble sensor platform and the Command Debug & Management Sstem CDMS The nternal subsstem level hardware and software components of the SCAV archtecture are shown n Fgs. 4. A SCAV sstem contans the followng subsstem modules. An Atmel ATmega Localzaton, Navgaton and Tracng LNT processor that forms the central processng platform, A Htach H 8/39 based locomoton controller, 3 An Ultrasonc/RF localzaton card CRICKET [6], from Crossbow Technologes [7], 4 A sensor card contanng onboard sensors ncludng acceleraton, temperature, magnetc feld, lght, and sound, and 5 An nfrared promt sensor for navgatonal collson avodance, and 6 A 900MHz rado nterface for networ mplementaton. Pacet routng protocols are mplemented over ths 900MHz rado nterface for ad hoc networ operatons among the SCAV unts, the wreless bacbone, and the CDMS modules as shown n Fgure 4. The raw data-rate of each ln s appromatel 0Kbps, and CSMA and AODV are the MAC and the routng protocols that run on these nterfaces. Both pont-to-pont and all-to-all pacet communcatons 7

8 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. are supported usng AODV and floodng respectvel. No ln laer protocol s used for hop-level error recover, and t s left up to the applcaton laer e.g. the mult-target tracng as presented n Secton 7 to deal wth pacet losses. As for software, the Htach locomoton controller runs an embedded mcro-ernel BrcOS [8], a publc doman operatng sstem, used for mplementng SCAV s locomoton control usng geared motors and navgatonal touch and nfrared sensors. A seres of motor control and debuggng APIs have been mplemented between the LNT processor and the locomoton processor usng a custom bult seral nterface. The SCAV moble sensors are capable of movng wth a mamum speed of 0cm per second. Self-localzaton n a SCAV s performed wthn the LNT processor b computng coordnates usng the sensor unt s dstance from a number of pre-nstalled localzaton beacons n nown locatons. Ths dstance nformaton s suppled b the CRICKET RF/US locaton sensor cards [6]. Note that the beacons are asnchronousl broadcast b four ndependent CRICKETS at an nterval of 0.8 seconds. No cloc snchronzaton was necessar between the celng-mounted CRICKET unts. In LNT processor, the embedded mcro-ernel TnOS [9] s run for supportng coordnate computaton, Kalman Flter based navgaton and collaboratve tracng servces as shown n Fgure. The LNT processor s also used for mplementng a number of remote commands to and from a Command, Debug and Management Sstem CDMS as shown n Fgure 4. The CDMS s mplemented usng a 900MHz RF nterface and an Atmel ATmega processor, whch s connected to a Wndows PC console over RS-3 seral nterface. Usng ths arrangement, commands can be sent to specfc SCAVs from a user nterface n the Wndows PC. Smlarl, measurement and sstem performance parameters from specfc SCAVs can be uploaded through the CDMS usng the same Wndows based user nterface. 8

9 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. Rado Lns to Other SCAVs Lght, temperature, sound, and acceleraton sensors MTS-30 Sensor Card 5-pn Parallel Interface Htach H 8/39 Locomoton Controller Navgatonal Touch and Lght Sensors Swarm Capable Autonomous Vehcle SCAV RS-3 Interface Wndows PC 900MHz RF Ln Seral nterface Local Debug Wheel/Trac Motors Localzaton, Navgaton and Tracng LNT Processor MICA MOTE Vehcle Control Seral nterface MICA Vehcle Localzaton LCD Panel LED Arra Dmenson: Command, Debug and Management Sstem CDMS Ultrasonc and 400MHz RF Transcevers Crcet Localzaton Sensor Vehcle Level Debug Output Fg. 4: Sstem components and software modules n SCAV platform Sstem Software Components Command User Interface and Performance Analss Module Wndows XP Commands Performance RS-3 Interface APIs for Commands and Performance Upload from SCAVs TnOS APIs to CDMS for Command and Performance Upload Collaboratve Tracng SCAV-to-SCAV Kalman Flter Ad Hoc based Navgaton Networng Localzaton TnOS Seral nterface RF/US based Range Sensng wth Respect to Fed Beacons TnOS Seral nterface SCAV Locomoton Control and Local Debug Operatons BrcOS The prmar remote commands nvoed from the CDMS to SCAVs LNT processors are: SCAN: for scannng the locaton of specfc or all SCAVs present n the epermental feld, CHECK_BATT: for checng the remanng batter on specfc SCAVs, 3 GO: for sendng a SCAV from ts current locaton to a new destnaton wth a specfed speed, and 4 STOP: to force a SCAV to stop navgaton. And the prmar remote commands nvoed from a SCAV s LNT processor to CDMS are: 5 RPRINT: for remote prntng from a SCAV to the console PC s screen, and 6 PERF_UPLOAD: used for remote upload of collected performance data from SCAV sensors to the remote PC s hard drve. We have constructed a test-bed of three dentcal CDMS unts and fve SCAV unts whch are deploed n a 3m 3m epermental feld for conductng eperments nvolvng localzaton, navgaton and networed multtarget tracng. Multple CDMS unts are used for prntng debuggng nformaton from dfferent SCAVs to dfferent CDMS screens. 9

10 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, Moble Sensor Self Localzaton The components of the localzaton sstem are depcted n Fgure 5. When a self-localzaton s needed, a SCAV computes ts absolute dstances from a number of statc localzaton beacons pre-nstalled at nown coordnates. Dstance from a beacon s computed b measurng the tme dfference of arrval TDOA between an ultrasonc and a 433MHz RF sgnal smultaneousl transmtted b the beacon. Once dstances to a suffcentl large number of locaton beacons are computed, the SCAV computes ts own coordnates usng the dstance values and the nown coordnates of the relevant beacons. The MIT CRICKET localzaton hardware [6] has been used as the statc beacons shown n Fgure 5, and the localzaton sensor n the SCAV, as shown n Fgure 4. Fed Locaton Beacons on Laborator Celng d d RF/US T d 3 d MHz RF Moble Ad Hoc Networ 900 MHz RF Debug Ln CDMS Fg. 5: In-laborator SCAV localzaton sstem 5. Dstance Estmaton The followng method s used for dstance computaton. If V RF and V US are the nown speed of RF and ultrasonc sgnals through ar V RF >> V US, and T represents the tme dfference of arrval at the SCAV, then the dstance between the SCAV and the beacon can be wrtten as [0]: d = T V RF V. US If the dstance to the th locaton beacon, nstalled at coordnate,, z s computed as d, then the 0

11 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. = d d d d d d N N N N N measurement error ε can be wrtten as: d z z = ε, where,, z s the actual locaton of the SCAV. Wth dstance measurement from N beacons, the localzaton problem becomes equvalent to estmatng the actual coordnate,, z, so that the cumulatve squared error quantt = N ε s mnmzed. Ths constraned optmzaton problem s nown as Mnmum Mean Square Estmate MSME [] n the lterature. 5. Onboard Coordnate Computaton at Real-tme The smplest wa to solve ths sstem of N nd order equatons N d z z,...,, ; = = s to convert t to N- lnear equatons b par-wse subtractons. For eample, the quadratc equatons for = and can be combned 3 nto the lnear equaton d d = for N,..., =. Ths provdes a lnear sstem of equatons: If ths N- th order sstem s represented as V U A =, then the soluton for MMSE can be computed b the followng matr operaton: [ ] V A A A U T T =., as long as 3 N and the matr A s non-sngular. We have mplemented ths MSME based self-localzaton mechansm on the onboard LNT processor as shown n Fgure 4. A 3m 3m SCAV navgaton test-bed has been created wthn a laborator, at the celng of whch four statc beacons have been nstalled. The beacons have been postoned wthn a 0.5m 0.5m area rght above the center regon of the 3m 3m test-bed. 3 If all locaton beacons are placed at the same heght, the z-coordnate cancels out from the equatons.

12 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, Localzaton Error Cm meter 3 meter Test-bed Fg. 6: Performance of sensor self-localzaton Performance of SCAV localzaton s presented n Fgure 6, whch shows the absolute error represented as the dfference between a SCAV s actual coordnate and ts locaton estmated b the localzaton mechansm descrbed above. The prmar source of ths error s from fault dstance computaton caused b the naccuraces of the CRICKET hardware used as the beacons as well as lsteners wthn the SCAV unts. Other sources of errors were dentfed as varous ambent RF and US nterferences, and hgh frequenc mechancal vbraton contrbuted b the SCAV motors. The measurements were taen b statcall postonng a SCAV at dfferent locatons n the test-bed as shown n Fgure 5. For each locaton, 50 dfferent locaton computatons were used to report an average error. From Fgure 6, observe that the mamum localzaton error s about 7cm, whch s onl about.7% of the dagonal of the entre test-bed area. Comparng wth the localzaton errors reported n [] and [], we consder the mplemented ndoor locaton servce to be suffcentl accurate for effectve navgaton and target tracng. Also t can be observed that the errors are generall larger awa from the center of the test-bed. Ths s because the beacons are placed n an area rght above the center regon, and as the SCAV s moved awa from ths regon the beacon dstances ncrease. The dstance computaton errors were found to be postvel correlated wth the dstance tself [0]. Ths eplans hgher errors near the perpher of the navgaton test-bed. Also, the locaton estmaton performance was found to be

13 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. unbased. 6. Fltered Sensor Navgaton The most prmtve component of the pont-to-pont SCAV navgaton servce s to be able to navgate a moble sensor from ts current locaton to a new destnaton locaton n the mnmum possble tme. Ths functon has been mplemented as a remote command to the SCAVs LNT processor. An remote entt such as the CDMS or another SCAV can ssue ths command to ntate navgaton for a SCAV to a specfc destnaton. A SCAV can also self-nvoe the command locall. In both cases, once the command s sent to ts LNT processor, a SCAV s requred to autonomousl navgate to the specfed destnaton. The navgaton latenc depends on the accurac of the underlng localzaton servce as descrbed n Secton 4. Although the localzaton errors n the cases of statc SCAV placements, as reported n Fgure 6, are ver small, the error numbers were found to be sgnfcantl larger wth movng sensors. Whle ultrasound nterferences due to mechancal vbraton and motor noses partall responsble for ths naccurac, the prmar reason for such hgh errors s as follows. Due to the asnchronous nature of the RF and US transmssons from multple localzaton beacons four n our nstallaton, the smultanet condton of dstance measurement does not hold [3]. In other words, the dstances computed to dfferent localzaton beacons correspond to dfferent SCAV locatons, whereas the MSME mechansm for coordnate computaton mplctl assumes that all d values are measured from the same SCAV locaton. Ths contrbutes to heav coordnate errors that can translate nto slow pont-to-pont navgaton due to frequent erroneous headng changes. 6. Error Reducton usng Kalman Flter Snce the locomoton dnamcs of a SCAV can be determnstcall modeled, and s nown a pror, t s possble to desgn a Baesan flter [4] that can partall compensate for the localzaton errors b predctng the sensor s locaton at successve localzaton nstances accordng to ts underlng locomoton dnamcs. The predcton value can then be ncrementall corrected usng the locaton computed b the SCAV s LNT processor. Consderng the lneart of SCAV s movement dnamcs, we mplement a lnear Kalman Flter [4], whch mplements a specal case of the Baesan flter b assumng that the sstem and measurement noses are strctl Gaussan. Note than unle n the ndoor 3

14 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. fltered navgaton approach n [3], whch uses an Etended Kalman Flter EKF, we use a regular Kalman Flter wth the assumpton of lneart for the SCAV s movement dnamcs. Although an EKF can provde better flterng performance for a wheeled non-holonomc sstem such as the SCAV, the processng requrement for EKF can be prohbtve for Atmel Atmega 8L 4 MHz whch s used as SCAV s LNT processor platform. Ths processng constrant s partcularl relevant when consderng the other sstem operatons such as networ protocol processng, sensor navgaton, and mult-target tracng as descrbed n later parts of the paper. Consderng ths resource constrant, whch s tpcal for an embedded sensng platform such as the SCAV, we have chosen a lnear Kalman Flter nstead of an Etended Kalman Flter for reducng the localzaton errors. We use two lnear one dmensonal PVA poston, veloct, acceleraton Kalman Flters KFs, one for the dmenson wth state vectors[ ] T & &, and the other for dmenson wth state vector[ & & ] T. These two flters wor dentcall, but ndependentl, and here we present the Kalman equatons onl for the dmenson. Eact same equatons appl for the dmenson as well. We have eplored both the second order PV and the thrd order PVA flters, and based on the epermental accuraces, the thrd order mplementaton s chosen. Careful nvestgatons revealed that frequent stop-and-go, and drecton changes, and the subsequent changes of SCAV veloct could not be accommodated well b a PV flter n the 3m 3m epermental envronment. A thrd order flter provded better error estmatons. Also, we have tred two approaches, one wth coupled and states, and the other wth decoupled states as specfed above. Epermentall we found that the estmaton accuraces from both the models were prett comparable. The decoupled approach was adopted for ts relatve computatonal lghtness. Assumng a constant & &, and gnorng & & and hgher order dervatves, the dscrete state equatons for SCAV locomoton at the th teraton can be wrtten as: = & & Δt Δt w & & = & && Δt w && & && w =, where Δt s the dscrete tme step, and w represents the sstem nose. Assumng tme nvarant sstem nose, the above sstem of equaton can be wrtten more concsel n vector format as: S = A S W The measured value of the through localzaton at the th step can be wrtten as: v m =, where v s the measurement error and s the actual value of. Assumng tme nvarant measurement error, t 4

15 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. can be wrtten as: Z = H S V. The constant matrces A and H n Equatons and are: Δt Δt 0 Δt 0 0 and [ 0 0]. Consderng ths sstem and measurement formulaton, the Kalman predcton equatons can be organzed as: T S and P = A P A Q, 3 = A S where - ndcates a predcton and ndcates a Kalman correcton. The matr P represents the sstem state covarance, and Q represents the covarance matr of the sstem nose parameters. The Kalman correcton equatons are: K S T T = P H H P H R = S K Z H S = P K H P P 4 The matr K s the Kalman gan, and R represents the covarance matr of the measurement error parameters. The LNT processor n SCAV teratvel solves the sstem of Equatons 3 and 4 wth Δt set to be 500ms. The covarance matrces Q and R were epermentall tuned at dagonal element values of and 0. respectvel for the best flter performance. An outler rejecton mechansm was also embedded wthn the flter so that whenever a measured coordnate n Z s found to be too far from the last predcated value n S, the Kalman gan s forcbl reduced to deemphasze ths new measurement, and thus reducng the effects of overl erroneous locaton measurements. When the absolute dfference between Z and S s found to be more than 50% of the estmated value S, the measured Z s consdered to be an outler. The mplemented gan reducton was 90%, whch s epermentall optmzed. In other words, the Kalman gan s reduced to one tenth of the orgnal gan n the events of outlers. 6. Sensor Navgaton Both the flters for the and dmensons are cloced snchronousl at 500ms ntervals and the corrected and coordnates from S are used for navgatng a SCAV. In case of a mssed beacon, the 5

16 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. most recentl receved beacon s used for flter processng at the clocng tme ponts. Such beacon losses were found to be present and the do contrbute to the overall navgaton error reported n Secton 6.3. Fgure 7 depcts an ntegrated localzaton and fltered navgaton approach that s mplemented n the LNT processor for pont-to-pont SCAV navgaton. Note that after the Kalman correcton step s performed, the desred sensor headng s computed b usng the corrected coordnates S and S from the current and the prevous steps. At ths stage, the SCAV maes slow turns f t determnes that a headng change s necessar to realgn ts movement towards the current destnaton. A constant speed dfferental between the left motor and the rght motor s used for mplementng the turns. When a turn needs to be eecuted, the Atmega LNT processor nstructs the Htach locomoton controller about the speeds of the ndvdual motors and the duraton for whch the speed dfferental should be mantaned. These three parameters together determne the amount and the speed of the turn. Ths ntegrated localzaton, flterng, and navgaton decsons run contnuousl tll the SCAV reaches ts destnaton specfed n the GO command. GO Command Receved b the LNT Processor Reached es Destnaton? no S Kalman Predcton MSME based Coordnate Measurement Z Stop Navgaton Keep Movng for Δt tme Turn to Change Headng based on S, current headng, and destnaton S ABS - Z > Outler Threshold? no Compute Headng usng and S Kalman Correcton S S es Outler rejecton: Force Outler Detected a Low Kalman Gan Fg. 7: Integrated localzaton and navgaton algorthm 6

17 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. It s possble to mplement the Kalman process and the navgaton process as two asnchronous eecuton threads, so that the flter produces corrected coordnates drven b a separate cloc, and the navgaton and other possble applcaton threads can access to those coordnates asnchronousl. That wa, the cloc frequences of those processes can be set ndependentl. We have also epermented wth navgaton mplemented n a stop-and-go manner so that between the moves, a SCAV wats for some tme to compute coordnates usng statc localzaton. Snce the statc localzaton produces more accurate coordnates, the stop-and-go polc s able to offer more accurate navgaton than the contnuous mode of operaton presented n Fgure 7. Due to the stops however, the stop-and-go was observed to be consstentl slower than the contnuous mode. Note that the ntegrated navgaton mechansm n Fgure 7 mplctl uses certant equvalence prncple to separate navgaton from locaton estmaton n that the navgaton module assumes the SCAV s at the mean poston estmate. These mean poston values, n turn, are used for perodc headng correcton durng navgaton. 6.3 Epermental Sensor Navgaton Performance Performance of the SCAV navgaton sstem s reported n Fgures 8 and 9. Fgure 8:a shows an epermental scenaro n whch frst a GO command was sent to a SCAV to move t from locaton 350,350 to 50,50, and then another GO was gven from 50,50 to 50,350. Note that the 300cm 300cm area n the graph represents the SCAV test-bed used for the followng eperments. After the navgaton was completed, the PERF_UPLOA command was remotel nvoed from the CDMS see Fgures -4 for wrelessl uploadng the performance numbers to a PC for data post processng. 7

18 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. Y Coordnate cm Change of Destnaton Fltered Locaton Measured Locaton X Coordnate cm Dstance Between Succesve Navgaton Ponts cm Fg. 8: Navgaton performance: a absolute locatons, and b dstance between successve navgaton ponts Fltered Average: 4.7 cm Measured Average: 9.5 cm Change of Destnaton Tme seconds The ponts n Fgure 8:a represents the measured coordnates from the vector Z and the Kalman corrected coordnates from S. These ponts clearl demonstrate how the presented flter s able to smooth out the navgaton of a SCAV n the presence of measurement errors. The effects are further quantfed n Fgure 8:b, n whch the dstances between successve measured locatons Z, and the Z fltered locatons S, are plotted wth tme. The average of that dstance for the measured locatons S s appromatel 0cm, although the ndvdual measurement goes as hgh as 40cm. In the deal case, the average should be around 4.5cm, whch s the dstance traveled n 500ms Δ t at a SCAV speed of 9cm/sec used n these eperments. Although there are some hgh values, the fltered dstances average to a prett close value of 4.7cm. Ths further confrms the effectveness of the flter. Note the spe that corresponds to the change of destnaton caused b the second GO command. Also observe as to how the measurement errors are sgnfcantl larger compared to the statc localzaton errors presented n Fgure 6. Ths dfference can also eplan wh a stop-and-go navgaton, that uses statc localzaton, can delver more accurate navgaton compared to ts contnuous counterpart used n the eperments here. 8

19 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. Headng Change degrees Change of Destnaton Measured Fltered Fltered Average: 30 degree Measured Average: 56 degree Tme seconds Frequenc percentage Fltered Error Dstrbutaon n Vehcle Headng Measured Fltered Measured Error n Headng degrees Fg. 9: Navgaton performance: a headng changes, and b headng error dstrbuton The flter s effectveness n allevatng navgaton naccuraces n terms of sensor headng are presented n Fgure 9. In Fgure 9:a, the absolute dfference between the computed headngs n successve navgaton nstances are plotted wth tme. The measured values correspond to headng computed usng Z, and the fltered values correspond to headng computed usng S. In deal stuaton, the headng should reman constant at an angle of 45 o for the frst leg, and 90 o for the second leg of the navgaton segments shown n Fgure 8:a. The dfferences n headng, therefore, should reman zero ecept at the pont of destnaton change at the end of the frst leg. The results ndcate sgnfcantl hgher errors and varaton n the measured headng changes compared to those computed through flterng. To quantf these errors further we plot the dstrbutons of the absolute headng errors dfference from 45 o and 90 o n the frst and the second legs n Fgure 9:b. Observe how the fltered errors are clustered mostl wthn an error of 30 o, whereas the raw measurement errors are spread over a much wder range. 9

20 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. These results valdate our ntegrated sstem n terms of: a the SCAV platform archtecture ncludng ts hardware and software components, b the CDMS based debuggng nfrastructure, c ndoor self localzaton sstem, d dnamc navgaton wth a Kalman flter model, and e ntegrated operaton of all the above subsstems. As shown n Fgure, all these developed servces are leveraged b the collaboratve sensor applcatons ncludng the mult-target tracng framewor as presented n the net secton. 7. Collaboratve Mult-target Tracng The tracng problem s formulated as follows. There are M-number of Moble Target Agents MTAs plng n a reconnassance area, and the goal s to effcentl trac and follow the MTAs usng A-number of Autonomous Reconnassance Vehcles ARVs. The ARVs are able to sense MTAs and navgate collaboratvel for mamzng ther tracng coverage. The problem s to develop dstrbuted algorthms for such collaboratve tracng. Successful tracng for an MTA s defned b the stuaton when at least one ARV follows the MTA whle beng wthn a pre-defned range, termed as the Tracng Range. For a suffcentl large tme horzon, f D represents the cumulatve tracng duraton for MTA, then the overall tracng performance for the entre tracng sstem can be represented b a Cumulatve Tracng Inde CTI, defned as M = D, where M s the total number of MTAs n the sstem. Whle mamzng the CTI wll ensure the best utlzaton of the ARV fleet, eepng the Coeffcent of Varaton COV of D =,, M wll ensure that certan MTAs wll not reman untraced whle the others are aggressvel traced. The COV s computed as the rato of standard devaton of the D values and ther average. The objectve of the tracng algorthms wll be to mamze CTI, whle eepng the COV small. 7. Networed Tracng Algorthm We propose the followng heurstc based collaboratve tracng algorthm. The e concept of the algorthm s that the ARVs wrelessl share ther ndvdual tracng performance, and ndvdual ARVs swtch to dfferent MTAs when that s deemed approprate for meetng the CTI and COV objectves stated above. A pseudo-code for dfferent components of the algorthm s presented n Fgure 0. 0

21 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. /* Collaboratve Tracng Algorthm eecuted b ARV j */ Sensng: ARV j perodcall senses to chec f t s wthn the tracng range of an MTA Stored Informaton: A D-table wth the currentl stored D values for all MTA =,, M; all D values are ntalzed to be zero; Informaton Echange: ARV j Perodcall sends ts D-table to the rest of the ARVs through an nter ARV ad hoc networ Tracng: Whle for ever { IF not loced wth an target MTA { Reman statonar and wat tll an MTA s sensed; // MTA s sensed b the ARV j ARV j locs to MTA and starts tracng t D n ARV j s table s ncremented b one after each constant duraton of tracng b ARV j } ELSE { // ARV j s currentl loced to an MTA r // D r n ARV j s table s ncremented b one after // each constant duraton of tracng b ARV j IF an MTA s sensed b the ARV j { // ARV j ma need to swtch from MTA r tomta Compute Dˆ, the average D over all avalable MTA entres n the local D-table IF D r > Dˆ && D r > D Swtch_Threshold { ARV j swtches to MTA and starts tracng t } } } } D-Table Update: When RAV j receves a D-table from another RAV, t merges the receved table wth ts estng table so that: new MTA entres are created as needed, and for MTA entres, estng n both the tables, update the D r feld b the mamum of the two. Fg. 0: Pseudo code for tracng algorthm The tracng algorthm s decoupled from the D-table dssemnaton model, and can wor wth a varet of mechansms such as controlled floodng, dnamc spannng trees, and smlar networ protocols. Also note that as long as an ARV remans loced to an MTA, the approprate D value s updated wthn the local D-table of the ARV. The D value n ARV j s table s ncreased b one after each constant duraton of

22 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. tracng of MTA b ARV j. Snce the D-table merger see Fg. 0 alwas use the mamum networ wde avalable value of D, ths quantt represents the cumulatve tracng tme of MTA b all the ARVs n the networ. Accordng to the presented algorthm, when there are more or equal ARVs than MTAs, the algorthm reaches a stead state when at least one ARV gets loced to an MTA. No further target swtchng s necessar n these scenaros. Even when the MTAs outnumber the ARVs, the Cumulatve Tracng Inde CTI can be mamzed just b ensurng that each ARV gets loced to a dfferent MTA. Ths however does not mnmze the Coeffcent of Varaton of CTI, snce t s then possble that few MTAs ma reman untraced for the entre tme horzon. To avod ths, the concept of swtchng has been ntroduced n the algorthm. An ARV, whch s currentl tracng MTA r, decdes to swtch to a newl sensed MTA onl f t concludes that MTA r had been better traced than the networ wde average tracng hstor of all the MTAs n the sstem, and ts MTA r s CTI s more than that of MTA b more than a threshold amount. A e assumpton for ths tracng framewor s that whle sensng an MTA, an ARV s able to detect the dentt of an MTA. Ths enables the ARVs to collaborate for decdng when and how to swtch across dfferent MTAs. Wthout ths assumpton the problem becomes more comple and that s a topc of our ongong wor. 7. Implementaton of Tracng on the SCAV Sensor Test-bed The tracng algorthm was evaluated for varng number of MTAs and ARVs usng up to fve SCAV unts and ther pont-to-pont navgaton servces as descrbed n Secton 5. The MTAs are realzed usng a set of SCAVs, programmed to move along preset trajectores, and a dfferent set of SCAVs were used as the ARVs wthn our 3m 3m navgaton test-bed. The perodc MTA sensng process b the ARVs, ndcated n Fgure 0, was emulated b programmng each MTA to send a perodc wreless LOC beacon contanng ts locaton nformaton. Snce an ARV s aware of ts own locaton, when t receves a LOC beacon from an MTA wthn a preset Tracng Range, a successful MTA sensng s assumed to be accomplshed. Also, an ARV navgates toward the locaton nformaton found n the most recentl receved LOC pacet from a loced MTA. Due to ths locaton based emulaton, the sensng process had smlar order of errors as found for the dnamc locaton errors n Fgures 8 and 9. Dfferent levels of

23 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. MTA sensng senstvt were emulated b smpl varng the preset Tracng Range. Controlled floodng was used for the D-table echange see Fgure 0 through a 900MHz moble ad hoc networ formed b the ARVs. The ARVs perodcall transmt SYNC beacons contanng ts local D-table nformaton for the controlled floodng. When a large number of movng SCAVS more than 3 are placed wthn the 3m 3m navgaton testbed, frequent sensor collsons were observed. To mtgate such collsons, the followng measures were mplemented. usng an Infra Red based promt detector for collson avodance, ensurng a mnmum separaton between all movng sensors, and 3 a networed shadow tracng to avod collsons between multple ARVs tracng a sngle MTA. Under such stuatons, onl one ARV drectl follows the MTA, and the other ARVs collaboratvel choose movng shadow ponts of the MTA whch are slghtl awa from the MTA tself. Ths prevents an undesrable convergence between the tracng ARVs, thus reducng the sensor collson possbltes. The shadow arbtraton among multple ARVs s accomplshed based on the unque dentfers of each SCAV unts. Also a speed dfferental has been ntroduced between the MTAs and the ARVs, so that a speedng ARV can alwas catch up wth a loced MTA b gettng closer to t. 7.3 Tracng Performance The graphs n Fgure :a show scenaros nvolvng one and two Autonomous Reconnassance Vehcles tracng a Moble Target Agent whch s programmed to travel along the trajector: 350,350 50,50 350,50 50, ,350 50,50 50, ,50 350,35 0. For each eperment, 60 LOC pacets were sent out b the MTA at the rate of one pacet ever s. The Cumulatve Tracng Inde CTI was measured b countng the number of LOC pacets receved b an ARV whle beng wthn the tracng range of the MTA. All reported results represent an average from three separate tracng runs wth the same ntal condtons. The ARVs are ntall placed such that the are wthn the sensng range of the MTAs. For the -MTA/-ARV scenaro, the tracng problem reduces to a smple leader-follower strateg wthout an need for MTA swtchng as ntroduced n the general algorthm n Fgure 0. In the absence of MTA swtchng, the algorthm s alwas epected to provde the best CTI of 60. However, the results n Fgure :a ndcate that the measured CTI can be sgnfcantl smaller, especall for short tracng 3

24 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. ranges. Ths s because of the navgatons errors, as shown n Fgure 8 and 9, whch occasonall force the ARV to erroneousl drft out the tracng range of the MTA, especall when the range s small. Wth ncreased tracng range, snce the navgaton errors become smaller n a relatve term, the CTI value mproves, although t fnall saturates at a value of 0, whch s lower than the best case of 60. Cumulatve Tracng Inde ARV -ARV -Moble Target Agent MTA Tracng Range cm Fg. : Tracng performance: effects of a tracng range, and b consumed wreless bandwdth Cumulatve Tracng Inde MTA -ARVs Tracng Range: 50cm : 00cm : 50cm : 00cm ARV Snc. Interval seconds Introducng two ARVs can help mprovng the stuaton smpl because when one ARV looses the MTA due to navgaton error, t s lel that the other ARV s stll loced to t. As shown n Fgure :a, ths collaboraton effect mproves the CTI for all tracng ranges, eventuall achevng the near-best case performance of 57, for a tracng range of 00cm. Snce two ARVs can smultaneousl trac one MTA, onl one assumes the role of the prmar tracer and the other remans as a shadow tracer as eplaned before. These roles however frequentl alter when the prmar tracer momentarl moves out of the tracng range due to navgaton errors. The effects of varable SYNC beacon frequences are shown on Fgure :b. Lower SYNC ntervals correspond to hgher D-table echange frequenc among the ARVs. Wth more frequent D-table echanges, an ARV s epected to tae better MTA swtchng decsons wth fresh D values receved from the other ARVs. The results ndcate that, although mld, ths effect s there for smaller SYNC ntervals s to 6s and for larger tracng ranges. For smaller transmsson ranges, however, the navgaton errors offset the benefts of fresher D values. Ths results generall ndcate that better tracng performance can be obtaned at the epense of hgher networ bandwdth n the form of frequent 4

25 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. D-table echanges over the nter ARV ad hoc networ. Cumulatve Tracng Inde per MTA ARV -ARV 3-Moble Target Agent MTA Tracng Range cm CTI: Coeffcent of Varaton Fg. : Tracng performance wth more MTAs than ARVs ARV -ARV 3-Moble Target Agent MTA Tracng Range cm Performance when the MTAs outnumber the ARVs s shown n Fgure. The trajector 350,350 50,350 50,50 350,50 350,350 50,50 350,50 50, ,35 0 s used b three MTAs wth suffcent phase dfference for avodng ecessve sensor collsons due to SCAV crowdng. As shown n Fgure :a, wth ncreasng tracng range, the CTI values mprove due to the same reasons as for the -MTA/-ARV scenaro n Fgure :a. Also, as shown n Fgure :b, Wth ncreasng tracng range, the coeffcent of varaton for CTI sgnfcantl reduces due to the fact that wth larger tracng ranges the ARV s able to sense the MTAs more frequentl. As a result, the ARV s able to perform MTA swtchng more effectvel, thus reducng the varance of tracng tme across the MTAs. CTI per MTA and the correspondng coeffcent of varaton, when two ARVs tracng three MTAs, are also reported n Fgure :a, b. It s evdent that ntroducng a second ARV enables collaboraton, and mproves the CTI values. Also, the CTI varance across the MTAs s reduced b the fact that an MTA s now less lel to reman untraced. As shown n Fgure 3, we have epermented wth dfferent Swtch_Threshod values see the algorthm n Fgure 0 n order to evaluate ts effects on the MTA swtchng process. Wth hgher threshold values, MTA swtchng s dscouraged, as a result of whch the coeffcent of varaton for CTI ncreases. Meanng, whle few MTAs are well traced, few others ma reman untraced due to the lac of prompt 5

26 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, 009. MTA swtchng. Although counterntutve, our eperments showed that the CTI values also degrade wth hgher Swtch_Threshod values. Close nvestgatons revealed that the SCAV navgaton errors see Fgures 8 and 9 are responsble for ths effect n the followng manner. Cumulatve Tracng Inde Tracng Range: 50cm Cumulatve Tracng Inde Coeffcent of Varaton Tracng Swtchng Threshold Fg. 3: Effects of varng swtchng thresholds CMI: Coeffcent of Varaton As seen n Fgure, certan amount of loss of tracng happens even n -MTA/-ARV scenaro smpl due to the navgaton errors even n the absence of MTA swtchng. Smlar effects are present here for a large Swtch_Threshod, when ver few MTA swtchng are happenng. Wth more frequent swtchng, ths loss of tracng gets compensated because an MTA now spends less tme not beng traced, thus mprovng the CTI per ARV values. In the presence of deal localzaton and navgaton wth no or ver lttle error, we epect the CTI per ARV to n fact ncrease up to a pont wth hgher Swtch_Threshod values. These target tracng results demonstrate that non-deal localzaton and navgaton condtons can cause theoretcall sound algorthms to generate unepected results, and therefore specal epermental consderatons wll be necessar whle mplementng such moble sensor sstems under non-deal condtons. 6

27 To Appear, ACM/Sprnger Journal of Wreless Networs WINET, Summar and Future Wor We have presented the desgn, archtecture, mplementaton eperence, and epermental results from a networed moble sensor test-bed developed for collaboratve sensor applcatons. The test-bed comprses a fleet of networed moble sensors, an ndoor localzaton sstem for enablng sensor self-localzaton n GPS-dened envronments, a control, debuggng and management nfrastructure, and a tered wreless ad hoc networ for seamless ntegraton of the above three components and the estng wreless networ nfrastructure. A mult-target sensor tracng framewor has been mplemented on the moble sensor testbed as a representatve collaboratve applcaton. Jont networng and tracng mechansms are developed for tracng multple moble targets usng a team of networed moble sensors. In the frst part of the paper, the archtectural and mplementaton detals about the software and hardware desgn of the developed moble sensors and ts varous supported servces are presented. In the second part, frst the problem of mult-target tracng usng collaboratve moble agents has been ntroduced. Then a networed dstrbuted tracng algorthm has been formall proposed. Fnall, we report the epermental performance of the proposed tracng framewor mplemented n our moble sensor test-bed. In addton to valuable mplementaton nsghts about the localzaton, navgaton, Kalman flterng, and ad hoc networng processes, the epermental results lead us to the followng conclusons about the overall sstem. Frst, localzaton and navgaton errors, whch are usuall present n real-world scenaros, can sgnfcantl affect the tracng performance, even for a smple -leader/- follower scenaro. Navgaton errors are found to have more comple performance mpacts n scenaros wth multple targets and tracng sensors. Second, there ests a tradeoff between the tracng performance and the consumed wreless bandwdth. It s found that n most scenaros, better tracng can be acheved at the epense of ncreased communcaton through the moble ad hoc networs formed b the tracng sensors. Fnall, the frequenc at whch a tracng sensor swtches ts target s an mportant desgn parameter n scenaros wth hgher number of targets than the tracng sensors. Careful tunng of target swtchng wll be needed for strng a balance between acceptable overall tracng and the varance of tracng across ndvdual targets. Future wor on ths topc ncludes developng a number of sensor swarmng algorthms for collaboratng attac localzaton and search strateges usng the baselne tracng mechansms developed 7