whisper: a spread spectrum approach to occlusion in acoustic tracking

WWHISPERHISPER: A Spread Spectrum Approach : A Spread Spectrum Approach to Occlusion in Acoustic Trackingto Occlusion in Acoustic Tracking

Nick VallidisNick VallidisUnder the supervision of Gary BishopUnder the supervision of Gary Bishop

July 10, 2002July 10, 2002

2

OutlineOutline

•• MotivationMotivation•• Thesis StatementThesis Statement•• Previous Work Previous Work •• AlgorithmAlgorithm•• System PerformanceSystem Performance•• ConclusionConclusion

3

What is tracking?What is tracking?

•• Determination of the position and/or Determination of the position and/or orientation of a target objectorientation of a target object

•• There is a nearly ubiquitous tracking device There is a nearly ubiquitous tracking device you are all familiar with…you are all familiar with…

MotivationMotivation -- Thesis Statement Thesis Statement -- Previous Work Previous Work -- Algorithm Algorithm -- Performance Performance -- ConclusionConclusion

4

The Mouse!The Mouse!


5

Virtual EnvironmentsVirtual Environments

•• Need high quality tracking informationNeed high quality tracking information


6

Head TrackingHead Tracking

•• Focus of the majority of Focus of the majority of past VE tracking researchpast VE tracking research


7

Body TrackingBody Tracking

•• Growing interest due to quality head trackingGrowing interest due to quality head tracking•• Uses:Uses:

––AvatarsAvatars––Mine style interaction [Mine 1997]Mine style interaction [Mine 1997]

•• Goals:Goals:––OcclusionOcclusion--toleranttolerant––BodyBody--centeredcentered


8

Body Tracking ExampleBody Tracking Example


9

Thesis StatementThesis Statement

Spread spectrum technology applied to Spread spectrum technology applied to acoustic tracking produces a robust tracking acoustic tracking produces a robust tracking device with better performance than existing device with better performance than existing acoustic systems. Extending the frequency acoustic systems. Extending the frequency range of the signal down into the audible range of the signal down into the audible range enables tracking in the presence of range enables tracking in the presence of occlusions.occlusions.

Motivation Motivation -- Thesis StatementThesis Statement -- Previous Work Previous Work -- Algorithm Algorithm -- Performance Performance -- ConclusionConclusion

10

Spread Spectrum ConceptsSpread Spectrum Concepts

•• What is spread spectrum?What is spread spectrum?•• Multipath = echoesMultipath = echoes•• Multiple access techniquesMultiple access techniques

––Frequency Division (FDMA) Frequency Division (FDMA) –– Radio & TVRadio & TV––Time Division (TDMA) Time Division (TDMA) –– EthernetEthernet––Code Division (CDMA) Code Division (CDMA) –– Sprint PCSSprint PCS


Hedy Lamar

11

Spread Spectrum AdvantagesSpread Spectrum Advantages

•• When applied to acoustic tracking:When applied to acoustic tracking:––Continuous signal allows frequent updatesContinuous signal allows frequent updates––Robust to noise and multipathRobust to noise and multipath––CDMA allows multiple simultaneous rangesCDMA allows multiple simultaneous ranges


12

•• Using audible frequencies in the spread spectrum Using audible frequencies in the spread spectrum signal allows for more diffractionsignal allows for more diffraction

•• System can function while occluded!System can function while occluded!

Also…Diffraction is now usable!Also…Diffraction is now usable!


13

Previous Work (bodyPrevious Work (body--centered)centered)

•• Gypsy (& Gypsy (& ShapeTapeShapeTape))

Motivation Motivation -- Thesis Statement Thesis Statement -- Previous WorkPrevious Work -- Algorithm Algorithm -- Performance Performance -- ConclusionConclusion

14


•• Gypsy (&Gypsy (& ShapeTapeShapeTape))•• WearTrackWearTrack [[FoxlinFoxlin 2000]2000]


15


•• Gypsy (&Gypsy (& ShapeTapeShapeTape))•• WearTrackWearTrack [[Foxlin Foxlin 2000]2000]•• Magnetic [Magnetic [InskoInsko 2001]2001]


16

Previous Work (acoustic)Previous Work (acoustic)

•• Acoustic system featuresAcoustic system features––Sound speed is convenient Sound speed is convenient ––LowLow--cost, easily available transducerscost, easily available transducers

•• Poor performance of previous systems:Poor performance of previous systems:––Not robust in noisy environmentsNot robust in noisy environments––Slow update ratesSlow update rates


17

Previous Work (acoustic)Previous Work (acoustic)

•• Phase coherent [Sutherland 1968]Phase coherent [Sutherland 1968]•• Lincoln Wand [Roberts 1966]Lincoln Wand [Roberts 1966]•• Modern systems:Modern systems:

––Intersense Intersense Constellation [Constellation [FoxlinFoxlin 1998]1998]––Nintendo Nintendo Powerglove Powerglove [1989][1989]––Logitech 3D mouse [circa 1992]Logitech 3D mouse [circa 1992]


18

WWHISPERHISPER vs. Other Acoustic Systemsvs. Other Acoustic Systems

•• Pulsed trackersPulsed trackers––Faster update rateFaster update rate––Multiple simultaneous range measurementsMultiple simultaneous range measurements––More robust to echoes and noiseMore robust to echoes and noise

•• Phase coherent trackersPhase coherent trackers––Much more robust to echoes & interferenceMuch more robust to echoes & interference


19

Previous Work (spread spectrum)Previous Work (spread spectrum)

•• Global Positioning System (GPS)Global Positioning System (GPS)•• Aetherwire Aetherwire [Fleming 1995][Fleming 1995]•• Spatiotrack Spatiotrack [[Palovuori Palovuori 2000]2000]•• Acoustic/Optical system [Acoustic/Optical system [Girod Girod 2001]2001]•• VTT Automation [2000]VTT Automation [2000]


20

WWHISPERHISPER OverviewOverview

Motivation Motivation -- Thesis Statement Thesis Statement -- Previous Work Previous Work -- AlgorithmAlgorithm -- Performance Performance -- ConclusionConclusion

21

Pseudonoise Pseudonoise and Correlationand Correlation

•• PseudonoisePseudonoise –– known noiseknown noise--like signallike signal

•• Autocorrelation has large value at 0 delayAutocorrelation has large value at 0 delay


22

Correlation DemoCorrelation Demo


23

Correlation is Expensive!Correlation is Expensive!

•• Assuming 1000 sample window, 500 Hz Assuming 1000 sample window, 500 Hz computation rate:computation rate:––1000 multiply1000 multiply--adds per delayadds per delay––1000 possible delays1000 possible delays––500 times per second500 times per second

Total Cost = 1 billion operations per second!


24

WWHISPERHISPER’s’s AlgorithmAlgorithm

•• Limit delay search rangeLimit delay search range•• Limit computation cost per delayLimit computation cost per delay

Initialize with Correlation Kalman filter

PredictReduced

Correlation search around

prediction

+

interpolation

Kalman filter Correct


25

Kalman Filter BasicsKalman Filter Basics

•• Predictor/CorrectorPredictor/Corrector

•• Three components:Three components:––State estimate with variancesState estimate with variances––Process ModelProcess Model––Measurement ModelMeasurement Model

predictpredict correctcorrect


26

Correlation with Computation ReuseCorrelation with Computation Reuse

100 kHz Sampling rate

1 kHz update rate= 100 samples between Kalman filter iterations


27


Input signal(chunks of 100 samples):

…




28



…

Correlation window (1000 samples)

Iteration

k




29



…

Correlation window (1000 samples)

Iteration

k

Correlation window (1000 samples)k+1

100 kHz sampling rate



30

Calculating Position from RangeCalculating Position from Range

•• Easiest to look at 2D example:Easiest to look at 2D example:

KnownLocation 1

KnownLocation 2


31



KnownLocation 1

KnownLocation 2

Range 1


32



KnownLocation 1

KnownLocation 2

Range 1

Range 2


33

Prototype ConfigurationPrototype Configuration

Microphone 2

Microphone 1

Microphone 330.5 cm

30.5 cm

Speakeron target


34

Error SourcesError Sources

•• NoiseNoise•• Local environmental conditionsLocal environmental conditions•• Directionality of transducersDirectionality of transducers•• Geometric dilution of precisionGeometric dilution of precision


35

Geometric Dilution of PrecisionGeometric Dilution of Precision


36

3D, 2 Target Demonstration3D, 2 Target Demonstration

Motivation Motivation -- Thesis Statement Thesis Statement -- Previous Work Previous Work -- Algorithm Algorithm -- PerformancePerformance -- ConclusionConclusion

37

Overall 3D PerformanceOverall 3D Performance

•• 1000 updates per second1000 updates per second•• Static Performance:Static Performance:

–– 0.46 to 0.91 mm standard deviation 0.46 to 0.91 mm standard deviation •• LatencyLatency

––1818--49 ms depending on signal49 ms depending on signal--toto--noise noise ratioratio


38

Noise Performance (Range)Noise Performance (Range)


39

Static 3D PerformanceStatic 3D Performance

ArbitraryArbitrary

ArbitraryArbitrary

Z*Z*

YY

XX

Rail Rail OrientationOrientation

300.6 mm300.6 mm300 300 ±± 1 mm1 mm

300.2 mm300.2 mm300 300 ±± 1 mm1 mm

302.6 mm302.6 mm300 300 ±± 1 mm1 mm

300.7 mm300.7 mm300 300 ±± 1 mm1 mm

300.0 mm300.0 mm300 300 ±± 1 mm1 mm

WWHISPERHISPERDistanceDistance

Measured Rail Measured Rail DistanceDistance

* Rail was not rigidly supportedMotivation Motivation -- Thesis Statement Thesis Statement -- Previous Work Previous Work -- Algorithm Algorithm -- PerformancePerformance -- ConclusionConclusion

40

Dynamic PerformanceDynamic Performance

σ =2.0 mm

mean =202.7 mm


41

Occluded Range PerformanceOccluded Range Performance

676.8676.8675.2675.2642.8642.8101.6101.6663.0663.0661.0661.0642.8642.876.276.2652.0652.0650.8650.8642.8642.850.850.8645.3645.3644.8644.8642.8642.825.425.4

Measured Measured (b) mm(b) mm

Calculated Calculated (b) mm(b) mm

Direct Direct (a) mm(a) mm

Sphere Sphere Radius mmRadius mm


42

Occluded 3D PerformanceOccluded 3D Performance


43

Occluded 3D PerformanceOccluded 3D Performance


44

InnovationsInnovations

•• Interesting computational approach to a Interesting computational approach to a spread spectrum ranging systemspread spectrum ranging system

•• Compared to existing acoustic systems:Compared to existing acoustic systems:––More robust to noise and More robust to noise and multipathmultipath––Better performance (rate, simultaneous Better performance (rate, simultaneous

targets)targets)•• Diffraction allows tracking during occlusionsDiffraction allows tracking during occlusions

Motivation Motivation -- Thesis Statement Thesis Statement -- Previous Work Previous Work -- Algorithm Algorithm -- Performance Performance -- ConclusionConclusion

45

ApplicationsApplications

•• Body trackingBody tracking•• HCI device (like mouse)HCI device (like mouse)•• LabLab--mounted (ceiling) trackermounted (ceiling) tracker


46

Future WorkFuture Work

•• Filtering & Auto Gain Control (AGC) on inputFiltering & Auto Gain Control (AGC) on input•• Miniaturization and “Miniaturization and “wearabilitywearability” of system” of system•• Adaptations abound:Adaptations abound:

––Output volume to background noiseOutput volume to background noise––Output spectrum to occlusion (ultrasonic?)Output spectrum to occlusion (ultrasonic?)

•• Use Use kinematickinematic structure of human bodystructure of human body


47

AcknowledgementsAcknowledgements

THE END!

Personal:My familyMy wife

Committee Members:Gary Bishop (Advisor)Greg WelchLeandra VicciHenry FuchsJohn Poulton

Collaborators:Scott CooperJason StewartJohn ThomasStephen BrumbackKurtis Keller

Funding:Office of Naval ResearchUNC CS Alumni FellowshipNASA Graduate Student

Researcher Program


48

No, really. That was the end.

49

SpectrogramSpectrogram

50

Dual Target VelocitiesDual Target Velocities

51

Dual Target Velocity DifferencesDual Target Velocity Differences

whisper: a spread spectrum approach to occlusion in acoustic tracking

Documents