cheap passive approximate localization · 2020-02-12 · • vhf: 88mhz-108mhz, bw: 200khz • less...
TRANSCRIPT
ì
UsingFMRadio
Cheap Passive Approximate localization
Jointworkwith:
AndreasAdolfsson,IntelligentRoboticsIncTathagataMukherjee,IntelligentRoboticsIncEduardoPasiliao,AFRL
PiyushKumarWebpage:compgeom.com/~piyush
LargeScaleLocalizationusingjustRSS
ReceivingMusic:Agoodantennaorheightmakesadifference
>40dbU
Localization
ì Defn:Toconfineorrestricttoaparticularlocality
ì But,IhaveaCellPhone!ì UnavailableGPSì Whatifyouareindoors?ì Whatifpowerwasaconcern?
ì (ComparedtoWifi/GPS)
ì Importantformanydifferentapplicationsincludingcommunication&navigationinGPSDeniedenvironments
=500mofnoGPS
But why FM?
Others:ADS-B,TV,ATC,…Ours:Twolevellocalizationsystem
ìFM LocalizationFMBroadcastSignal
Ø FMbroadcastband:• Largecoverage• Reliable• VHF:88MHz-108Mhz,BW:200kHz• Lesssensitivetoweathercondition
andindoorlimitationthanGPS
0 50 100 150 200 250
Range of FM Towers in miles
0
200
400
600
800
1000
1200
1400
Num
ber
ofFM
Tow
ers
Ø KSJS(FM90.5)60dBupolygoninSanJose,CA
ìFM Localization Methods
Ø RFbasedlocalizationtechniques• Algorithm:Beaconbased,Anchor
based,TimeofArrival,TimeDifferenceofArrival,AngleofArrival,Doppler
• Fingerprinting• WejustuseRSStolocalize
RSS Based Localization
Ø UsesRTLSDRforprototyping.
Ø CapableofscalingtoentirePlanet/US
Ø Simpleandscalablealgorithm
Ø ImprovesLocalization,bothindoorsandoutdoors
Ø Easytomakedistributed
Ø Workswithoutlineofsight
Ø NoSyncrequired
Test Data
ì Drovemultiplecars:350+Miles,multipledays
ì MeasuredFMSignalsatapproximately1000locations
Data Acquisition
ì CheapestRTLSoftwareDefinedRadio
Clock
Laptop
Data Acquisition
ì CheapestRTLSoftwareDefinedRadio
ì PowervsFrequencyplots
Power
Frequency
ìFM LocalizationSystemOverview
Ø PreprocessingPhase• MapGeneration
Ø QueryPhaseØ PeakFindingØ SubsetFilteringØ NearestNeighbors
ìFM LocalizationPreprocessingPhase
Ø Goal• Predictstheestimatedpower
atapointbasedonthepriorknowledgeofnearbyFMstation.
Ø MapGeneration• 40dBucoverage• EntireUSwithapproximately
2.4milex2.4milegrid• Powerspectrumineachgrid
ìFM LocalizationPreprocessingPhase
Ø Goal• Predictstheestimatedpower
atapointbasedonthepriorknowledgeofnearbyFMstation.
Ø MapGeneration• 40dBucoverage• EntireUSwithapproximately
2.4milex2.4milegrid• Powerspectrumineachgrid
ìFM LocalizationPreprocessingPhase
ìFM LocalizationPeakFindingPhase
Ø Powerspectrumatonelocation• Lookingfor“spikes”
alongthespectrum• Compareadjacent
signalstrengthwithathreshold
• Returnthechannelfrequencieswithpeak
ìFM LocalizationPeakFindingPhase
Ø Powerspectrumatonelocation• Lookingfor“spikes”
alongthespectrum• Compareadjacent
signalstrengthwithathreshold
• Returnthechannelfrequencieswithpeak
ìFM LocalizationSubsetFilteringPhase
Ø SubsetFiltering:SearchSpaceReduction• Goal:reduceinitialsearch
areadowntofewhundredsquaremiles
• GivenasetV,foranyqueryvectorq∈ {0,1},detectsifanyvectorp∈ Vsuchthatqisasubsetofp
ìFM LocalizationQueryPhase
Ø QueryPhase:Actuallocalizationalgorithm• Acquiresthepower
spectrum• Findingthepeaksinthe
acquiredpowerspectrum• InvokeSubsetFilter• Getaminimumvalue,which
indicatethedistancebetweentwospectrums,restrictedonagivensetofpeaksP.
• ReturnpredictedlocationfromGeohash
ìFM LocalizationEuclidianMetric/Calibrationchallenges
Ø Variability(BothTx andRx)• Time• Temperature• Humidity• ExperimentalError
Received Signal Strength (dBm)-64 -62 -60 -58 -56 -54 -52
Prob
abilit
y
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7102.3
fitted curve
Why min distance?
D_i =i-th ResultofthesubsetqueryX=PeaksatthelocationofinterestM=NumberofmatchesfromsubsetqueryL=Mostprobablelocation
ìFM LocalizationWhytheEuclidianMetric?
ìFM LocalizationFriis Model
Ø Friis Model:• Directreceivesignalstrength
calculation• 1700measurementsforloss
factorinTallahassee,FL• Assumingisotropic
transmission,ignoringmultipath/terraineffect
• Trilaterationfittingforcircle
FM 88.9
FM 94.9
FM 97.9FM 96.1
ìFM LocalizationFriis Model
FM 88.9
FM 94.9
FM 97.9FM 96.1
ìFM LocalizationResults
Ø EuclidianalgorithmwithGaussianprobabilityhastheminimumerrorcomparestoFriis ModelandKendall-TauModel
Improving accuracy in the air
https://youtu.be/DYP22RmxbQ8
Autonomous Data Collection
• DJIS1000+Frame+motors+ESCPixhawk autopilot+PX4firmware
• RTKGPSmodule
• FMAntenna
• i7NUCcomputer
• RTL-SDR+EttusB210
BluetoothSpeaker
Logitechc920Camera
Autonomous Data Collection
Afterdroneisarmed:• ChecksGPSaccuracy
• CollectsRSSIreadingonground
• Liftsto120metersintheair
• RemainsstationaryinairwhilecollectinganotherRSSIreading
• Lands
Collected30Datapoints• Accuracyofresultsimproveasdatapointsincrease
Data Processing
•Wefirstusethepreviousalgorithmtogetourerrorto5miles.
•Ourmodellearnstoestimatethedistancetothetransmittersfromagivenlocationusing:
•thetransmittedpower
•thereceivedpoweratthelocation
•theheightofthereceiver
•theheightaboveaverageterrain(HAAT)ofthetransmitter
WechoseaRandomForestregressionmodel,usingsupervisedlearningtechniquestoestimatethisdistancetoeachtransmitter.
randomforest/NeuralNet/SupportVectorMachine
Aerial Results
MinError:172meters,AverageError:3000meters.
Acknowledgements
ì AFRL
ì CompGeomInc.
ì IntelligentRoboticsInc.
ì EttusResearch
ì FloridaStateUniversity
Questions
ìFuture Work
Ø ImproveFMlocalizationaccuracy:• TDoA andAoA withdirectionalantenna• SimulatedDatabaseimprovement:
Splat!SimulationorRadioMap (DARPA)• Alternatemodalities:ADS-B,Iridium
SatelliteConstellation
Ø ComputerVisionforlocalizationandcollisionavoidance
FMLocalizationandRobotics