Automated Intruder Tracking using Particle Filtering and a Network of Binary Motion Sensors
description
Transcript of Automated Intruder Tracking using Particle Filtering and a Network of Binary Motion Sensors
Automated Intruder Tracking using Particle Filtering and a Network of
Binary Motion SensorsJeremy SchiffJeremy SchiffEECS DepartmentEECS DepartmentUniversity of California, BerkeleyUniversity of California, Berkeley
Ken GoldbergKen GoldbergIEOR and EECS DepartmentsIEOR and EECS DepartmentsUniversity of California, BerkeleyUniversity of California, Berkeley
http://www.cs.berkeley.edu/~jschiffSupported by NSF Grants: 0424422/0535218
OutlineOutline IntroductionIntroduction Related WorkRelated Work Problem FormulationProblem Formulation
Setup and AssumptionsSetup and Assumptions Particle FilteringParticle Filtering
ResultsResults SimulationSimulation ExperimentalExperimental
Conclusion/Future WorkConclusion/Future Work
MotivationMotivation New class of technologies due New class of technologies due
to 9/11to 9/11 Automated SecurityAutomated Security Wireless Sensor NetworksWireless Sensor Networks
X10 PIR sensors - $25 X10 PIR sensors - $25 Robotic WebcamsRobotic Webcams
Pan, Tilt, ZoomPan, Tilt, Zoom 500 Mpixels/Steradian500 Mpixels/Steradian
Increased computer Increased computer processing speedsprocessing speeds Enables Realtime ApplicationsEnables Realtime Applications
Goal and ApproachGoal and Approach Wish to secure an Wish to secure an
environmentenvironment Low Cost Binary Sensors Low Cost Binary Sensors
X10 ~ $25X10 ~ $25 Optical BeamOptical Beam Floor PadFloor Pad Manufactured in ChinaManufactured in China
Noisy triggering patternNoisy triggering pattern RefractionRefraction
Use sensor triggering Use sensor triggering patterns to accurately patterns to accurately localize an intruderlocalize an intruder
IntuitionIntuition Utilize Sensor Overlap InformationUtilize Sensor Overlap Information
IntuitionIntuition Utilize Sensor Overlap InformationUtilize Sensor Overlap Information
OutlineOutline IntroductionIntroduction Related WorkRelated Work Problem FormulationProblem Formulation
Setup and AssumptionsSetup and Assumptions Particle FilteringParticle Filtering
ExperimentsExperiments SimulationSimulation Real-worldReal-world
Conclusion/Future WorkConclusion/Future Work
Related WorkRelated Work Pursuer/Evader GamesPursuer/Evader Games
Using line-of sight Using line-of sight optical sensorsoptical sensors
[Isler, Kannan, Khanna [Isler, Kannan, Khanna 2004]2004]
Tracking Multiple Tracking Multiple IntrudersIntruders
[Oh, Sastry 2005][Oh, Sastry 2005] Tracking Worn DevicesTracking Worn Devices
Track Infrared BeaconTrack Infrared Beacon [Shen et al. 2004][Shen et al. 2004]
Dynamic Shipment Dynamic Shipment Planning using RFIDsPlanning using RFIDs
[Kim et al. 2005][Kim et al. 2005]
Related Work IIRelated Work II Video Tracking SystemsVideo Tracking Systems
[Micilotta and Bowden 2004][Micilotta and Bowden 2004] Multiple Classes of SensorsMultiple Classes of Sensors
Multiple exclusive modes Multiple exclusive modes [Cochran, Sinno, Clausen 1999][Cochran, Sinno, Clausen 1999]
Fuse data of multiple sensor Fuse data of multiple sensor typestypes
[Jeffery et al. 2005][Jeffery et al. 2005] Automated Camera ControlAutomated Camera Control
[Song et al. 2005][Song et al. 2005]
Physical Devices
Virtual Devices
Related Work IIIRelated Work III Probabilistic Tracking Probabilistic Tracking
ApproachesApproaches Kalman Filtering Kalman Filtering
[Kalman 1960][Kalman 1960] Extended Kalman FilteringExtended Kalman Filtering
[Lefebvre, Bruyninckx, De [Lefebvre, Bruyninckx, De Schutter 2004]Schutter 2004]
Particle FilteringParticle Filtering Book: [Thrun, Burgard, Fox Book: [Thrun, Burgard, Fox
2005]2005] [Arulampalam et al. 2002][Arulampalam et al. 2002]
Related Work IVRelated Work IV Multiple humans controlling a cameraMultiple humans controlling a camera
[Song and Goldberg 2003][Song and Goldberg 2003] [Song, Goldberg and Pashkevich 2003][Song, Goldberg and Pashkevich 2003]
Panorama GenerationPanorama Generation [Song et al. 2005][Song et al. 2005]
Art Gallery ProblemArt Gallery Problem [Shermer 1990][Shermer 1990] [Urrutia 2000][Urrutia 2000]
OutlineOutline IntroductionIntroduction Related WorkRelated Work Problem FormulationProblem Formulation
Setup and AssumptionsSetup and Assumptions Particle FilteringParticle Filtering
ExperimentsExperiments SimulationSimulation Real-worldReal-world
Conclusion/Future WorkConclusion/Future Work
Setup and AssumptionsSetup and Assumptions Room GeometryRoom Geometry
List of nodes and List of nodes and edgesedges
Discretize spaceDiscretize space Discretize timeDiscretize time
Setup and Assumptions Setup and Assumptions IIII
Intruder occupied Intruder occupied world-space cell j world-space cell j in iteration in iteration
Sensor i triggered Sensor i triggered during iterationduring iteration
Sensor i Sensor i experienced experienced refraction period refraction period in in iterationiteration
Setup and Assumptions Setup and Assumptions IIIIII
Three Conditional Three Conditional DistributionsDistributions Trigger while experiencing Trigger while experiencing
refractionrefraction
Trigger from intruderTrigger from intruder
Trigger from no intruderTrigger from no intruder
OutputOutput Estimated Estimated
intruder locationintruder location Objective:Objective:
Minimize error Minimize error between ground between ground truth and truth and estimation.estimation.
CharacterizationCharacterization Per sensor typePer sensor type Grid over sensor Grid over sensor
spacespace Determine Determine
Refraction period Refraction period False Negative RateFalse Negative Rate False Positive RateFalse Positive Rate
DeploymentDeployment Convert to Convert to
world-spaceworld-space Overlay Overlay
grid grid TransformeTransforme
d point to d point to CellsCells
Deployment IIDeployment II Determine potential non-zero Determine potential non-zero
characterization cells via convex characterization cells via convex hullhull
Inverse Distance Weighting Inverse Distance Weighting Interpolation according to distanceInterpolation according to distance Determines values for cells without Determines values for cells without
readings inside convex hullreadings inside convex hull
Particle filtersParticle filters Non-ParametricNon-Parametric
Sample Based Method (Particles)Sample Based Method (Particles) Particle Density ~ Likelihood Particle Density ~ Likelihood Tracking requires three distributionsTracking requires three distributions
Initialization Distribution Initialization Distribution
Transition Model (Intruder Model)Transition Model (Intruder Model)
Observation ModelObservation Model
Determines Determines
ExampleExample
ExampleExample
Intruder ModelIntruder Model StateState
Position, Orientation, Speed, Position, Orientation, Speed, and Refracting Sensorsand Refracting Sensors
Euler Integration for positionEuler Integration for position Gaussian Random Walk for Gaussian Random Walk for
new speed and orientationnew speed and orientation Orientation change inversely Orientation change inversely
proportional to speedproportional to speed Deterministic refraction Deterministic refraction
periodsperiods Rejection Sampling to Rejection Sampling to
enforce room geometryenforce room geometry
Intruder Model IIIntruder Model II Time between iterations:Time between iterations: Empirically determined constants:Empirically determined constants:
Intruder Model - Intruder Model - ExampleExample
Example state at iteration 0
Intruder Model - Intruder Model - ExampleExample
Accepted state for iteration 1
Intruder Model - Intruder Model - ExampleExample
Example state at iteration 1
Intruder Model - Intruder Model - ExampleExample
Accepted state for iteration 2
Intruder Model - Intruder Model - ExampleExample
Example state at iteration 2
Intruder Model - Intruder Model - ExampleExample
Rejected state for iteration 2
Intruder Model - Intruder Model - ExampleExample
Example state at iteration 2
Intruder Model - Intruder Model - ExampleExample
Rejected state for iteration 2
Intruder Model - Intruder Model - ExampleExample
Example state at iteration 2
Intruder Model - Intruder Model - ExampleExample
Accepted state for iteration 2
Sensor ModelSensor Model Evidence is vector of which sensors are Evidence is vector of which sensors are
triggeringtriggering Triggering of sensors independent Triggering of sensors independent
given intruder state impliesgiven intruder state implies
If sensor refractingIf sensor refracting
OtherwiseOtherwise
OutlineOutline IntroductionIntroduction Related WorkRelated Work Problem FormulationProblem Formulation
Setup and AssumptionsSetup and Assumptions Particle FilteringParticle Filtering
ExperimentsExperiments SimulationSimulation Real-worldReal-world
Conclusion/Future WorkConclusion/Future Work
Simulation SetupSimulation Setup 22 Optical Beams22 Optical Beams
Perfect Perfect Optimal Optimal
PerformancePerformance
14 Triangular 14 Triangular Motion SensorMotion Sensor Perfect & ImperfectPerfect & Imperfect
Simulation ResultsSimulation Results Example PathExample Path Ground Truth Ground Truth
Red CirclesRed Circles Estimations Estimations
Grey CirclesGrey Circles
Simulation Results IISimulation Results II Baseline EstimateBaseline Estimate
Perfect Optical-Beam SensorsPerfect Optical-Beam Sensors
P(E)
Error E
Error E
P(E
)
Perfect Triangular Motion SensorsPerfect Triangular Motion Sensors
Imperfect Triangular Motion SensorsImperfect Triangular Motion Sensors
Simulation Results IIISimulation Results IIIP(
E)
Error E
Error E
P(E)
Error over Time – 4 Sec. Refraction, Error over Time – 4 Sec. Refraction, Imperfect SensorsImperfect Sensors
Density - 8 Sec. Refraction, Imperfect Density - 8 Sec. Refraction, Imperfect SensorsSensors
Simulation Results IVSimulation Results IVEr
ror
E
Time (Seconds)
Error E
P(E)
In-Lab ResultsIn-Lab Results 8 Passive Infrared Sensors8 Passive Infrared Sensors
X10X10 8 second refraction time8 second refraction time
Room 8x6 metersRoom 8x6 meters .3 m /Cell dimension.3 m /Cell dimension Sampled every 2 secondsSampled every 2 seconds 1000 Particles1000 Particles
In-Lab Results IIIn-Lab Results II
OutlineOutline IntroductionIntroduction Related WorkRelated Work Problem FormulationProblem Formulation
Setup and AssumptionsSetup and Assumptions Particle FilteringParticle Filtering
ResultsResults SimulationSimulation ExperimentalExperimental
Conclusion/Future WorkConclusion/Future Work
ConclusionsConclusions Real-time Tracking SystemReal-time Tracking System Binary Sensors with Refraction PeriodBinary Sensors with Refraction Period Particle Filtering for Sensor FusionParticle Filtering for Sensor Fusion
Conditional Probability ModelsConditional Probability Models Models Models
Intruder Velocity Intruder Velocity Room GeometryRoom Geometry Sensor CharacterizationSensor Characterization
Future WorkFuture Work Effects of varying different componentsEffects of varying different components
Number ParticlesNumber Particles Types of sensorsTypes of sensors Spatial arrangements of sensorsSpatial arrangements of sensors
Multiple intrudersMultiple intruders DecentralizeDecentralize Vision ProcessingVision Processing Other applicationsOther applications
Warehouse TrackingWarehouse Tracking
Thank YouThank You Jeremy Schiff: Jeremy Schiff:
[email protected]@cs.berkeley.edu Ken Goldberg: Ken Goldberg:
[email protected]@ieor.berkeley.edu URL: URL: www.cs.berkeley.edu/~jschiffwww.cs.berkeley.edu/~jschiff