Building RF Sensor Networks

Transcription

1 Device-free localization in wireless networks 5th IEEE SenseApp Workshop Keynote Address

2 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

3 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

4 RF Sensor Networks Network measurements of the radio channel provide the means for enhanced network security, localization, and environmental awareness. The radio is the sensor.

5 Multipath Channel as a Problem Multipath Fading Shadowing Broadcast

6 This Talk: Multipath Networking Benefits Secret key establishment Location distinction Sensor localization Device-free localization

7 Secret Key Establishment EM reciprocity Not shared by 3rd party Meas t non-idealities ARUBE: 40 secret key bps a on NexusOne, TelosB a J. Croft, N. Patwari, S.K. Kasera, Robust RSS integer AtoB BtoA Uncorrelated Bit Extraction Methodologies for Wireless Sensors, IPSN counter

8 Location Distinction App: Detect TX position change Channel responses unique to location

9 Cooperative Localization Measured RSS Distributed weighted MDS a Implemented in Mica2 deployment: 0.55 cm RMSE Key: Don t queue. Gossip. a J. A. Costa, N. Patwari, A. O. Hero III, Achieving High-Accuracy Distributed Localization in Sensor Networks, ICASSP 2005

10 Large-Scale Indoor Localization A B A B CD AE F Infrastructure-based RSS Figure: Signal-distance map Key: No global RSS model. Kernel model.

11 Localization WSN: 1000s of Nodes (Really!) Feb. 2008: Awarepoint Deploys Largest RTLS in Healthcare Zigbee WSN across ft 2 : 91 floors, 17 buildings Locating 12,000 Zigbee radio tags, RSS, temp Installed in < six weeks

12 Device-free localization (DFL) Applications RFID identifies, locates people s tags How about people, objects not tagged? Apps: emergency response, smart homes, context-based authorization

13 DFL: Technologies Video cameras. Don t work in dark, through smoke or walls. Privacy concerns. Thermal imagers. Limited by walls. High cost. Motion detectors. Also limited by walls. High false alarms. Ultra wideband (UWB) radar. High cost. Received signal strength (RSS) in a wireless network

14 RSS-DFL: Measure many spatially distinct links Link RSS changes due to people in environment near link One person / object affects multiple links Mesh network of N nodes O ( N 2) RSS measurements

15 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

16 Channel Modeling Generic model for received power: P i,j = P(d i,j ) X i,j P i,j : measured RSS at node j transmitted by node i (dbm) P(d i,j ): model: Ensemble mean dbm at distance d i,j X i,j : fading error Question: Are {X i,j } independent?

17 Deployments for Channel Modeling Fifteen indoor and six outdoor measurement campaigns Results: close links have correlated X i,j 1 1 P. Agrawal and N. Patwari, Correlated Link Shadow Fading in Multi-hop Wireless Networks, IEEE Trans. Wireless Commun., August 2009.

18 What mechanism explains shadowing correlation? shadowing field px ( ) x i x k Spatially correlated shadowing field p(x) Assume X i,j, X k,l are integrals of p(x) link a link b X i,j = 1 xj x j x i 1/2 p(y)dy. (1) x i x j x l Mutual dependence on p(x) correlation of X i,j, X k,l

19 Experimental Correlation Results Link Geometry vs. Correlation Coefficient (Observed, Model) Correlation ρ Correlation ρ Geom- Meas- Prop. Geom- Meas- Prop. etry ured Model etry ured Model *** *** *** *** *** 0.26

20 Impact of correlation on sensor networking s d Multiple routes are not independent! Route diversity: links fade simultaneously more often Localization: RSS errors don t average out after many links. But correlation = spatial information. Correlation implies spatial field can be estimated N. Patwari, P. Agrawal, Effects of Correlated Shadowing: Connectivity, Localization, and RF Tomography, IPSN 2008, April, 2008.

21 Loss is Linear with Dynamic Object Shadowing Field Two shadowing fields: 1) Static, 2) Dynamic Let p(y) be the dynamic db shadowing loss field Let X a be the dynamic shadowing loss (change from empty condition) X a is a spatially filter of loss field p(y): X a w i,j h i,j (y)p(y)dy. (2) Linear approximation of reality, using (1) y

22 Discrete-space Loss Field Model Consider simultaneously all M pair-wise links: x = W p + n x = [X 1,... X M ] T = measured losses (db) vs. empty p = [p 1,... p N ] T = discretized loss field (db/voxel) W = [[w i,j ]] i,j = weights; n = noise

23 Shadowing Field Estimation Problems Measure x, assume known W. Estimate p. Ill-posed! Pixels links, other issues Low SNR: RSS varies without human motion in area. Linear model isn t true physics; best W is unknown.

24 Real-time Approaches to Image Estimation Real-time requirement: look for linear algorithm ˆp = Πx Projection Π needs only be calculated once Complexity: Order of # Links # pixels

25 Regularized Image Estimation Algorithms 1 Regularized inverse: minimize penalized squared error 2 f (p) = W p x 2 + α Qp 2 when Q is the derivative: [ ] 1 Π Tik = W T W + α(dx T D X + DY T D Y ) W T 2 Assume correlated p and use regularized least squares. ( ) 1 Π RLS = W T W + αcp 1 W T 2 J. Wilson and N. Patwari, Radio Tomographic Imaging with Wireless Networks, IEEE Transactions on Mobile Computing, May 2010.

26 Real-Time Implementation: Testbed DE C B DE A Crossbow Telosb, 2.4 GHz, IEEE SPIN: Token passing MAC; when one transmits, others measure RSS Open source: Packet data: latest measured RSS values C B A Laptop-connected mote overhears all traffic Complete meas t of p 3-4 times/sec (28 nodes)

27 Video Video clip: Atrium of Warnock Engineering Building

28 Alternate Algorithms: TV Use total variation (TV) norm to pull out sparsity of image 3 3 J. Wilson and N. Patwari, Regularization Methods for Radio Tomographic Imaging, in Proc Virginia Tech Wireless Symposium.

29 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

30 Through-wall Deployment Tests Tested system with 34 nodes, outside of external walls of area of house 4 30 Nodes Walls y coordinate (feet) Door Interior Area Exterior Area Door Stairs x coordinate (feet) 4 J. Wilson and N. Patwari, See Through Walls: Motion Tracking Using Variance-Based Radio Tomography Networks, IEEE Transactions on Mobile Computing, (accepted).

31 Problem: Low SNR y coordinate (feet) x coordinate (feet) RTI does not indicate actual image human location (X) Through-wall links see small attenuation effect compared to other multipath fading effects.

32 Problem: What Happened? RSS (dbm) RSS (dbm) RSS (dbm) Link (27,0) to (15.45,26.4) Link (6,0) to (20,26.4) Vacant network area Stationary human obstructing link Moving human obstructing link Time (samples) Moving people affect RSS, but change is up and down E.g.: Blocking person increases RSS ( ) E.g.: Moving person increases RSS variance (both links)

33 Idea: Use Variance to Image Motion Model: Assume variance is linear combination of motion occurring in each pixel: s = W m + n s = [s 1,... s M ] T = windowed sample variance m = [m 1,... m N ] T = motion [0, 1] W = [[w i,j ]] i,j = variance added to link i caused by motion in voxel j

34 Variance-based Radio Tomographic Imaging y coordinate (feet) x coordinate (feet) Apply regularized inversion to estimate m. VRTI image indicates actual image human location (X)

35 VRTI Video Advice: Use YouTube (>135k hits for two videos)

36 VRTI-based Tracking 1 Spot motion test: avg. error = 0.45 m 2 Track image max w/ Kalman filter: avg. error = 0.63 m (1) y coordinate (feet) Nodes Known Positions 30 Estimated Positions x coordinate (feet) (2) x coordinate (feet) y coordinate (feet) Known position Estimated position Time (samples)

37 Need: Spatial Model for Variance Where does motion have highest impact on RSS variance? 1 Near TX, RX [Yao et. al. 2008] 2 At midpoint between TX, RX [Zhang et. al. 2007] 3 Our work: In (narrow) ellipse w/ TX & RX as foci 4 Pixels which intersect link line [Kanso and Rabbat 2009] Need for measurements, analytical models

38 Variance Measurement Measurement at Bookstore, nodes on shelves Normalize link, person position s.t. x r = (-1, 0), x t = (1,0) Find average variance by human position w.r.t. RX, TX

39 Analytic Model: Intro Do simple standard multipath assumptions explain data? a Human = tall cylinder diameter D [Ghaddar et. al. 2004, Huang et. al. 2006] Scatterers/Reflectors in a plane. TX, RX, in plane z above. Propagation via single bounce a N. Patwari and J. Wilson, Spatial Models for Human Motion-Induced Signal Strength Variance on Static Links", submitted to IEEE Trans. Info. Forensics & Security,

40 Analytic Model: Details Locations: TX x t, RX x r, bounce at x Propagation mechanism (a) scattering or (b) reflection (a): P s (x) = x t x 2 x r x 2 c r (b): P r (x) = ( x t x + x r x ) np c r, c s, n p R + are propagation parameters [Nørklit & Andersen 1998, Liberti and Rappaport 1996] Variance prop. to expected total affected power (ETAP) c s

41 Analytic Model: Results Variance spatial functions: (a) Y Coordinate X Coordinate (b) Y Coordinate X Coordinate Ours & [Yao 2008]: similar to reflection ETAP, low z Those of [Zhang 2007]: high z, either modality

42 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

43 Problems: Tracking from RSS Tracking motion from image estimate is ad hoc Image estimate may be very poor when DFL is possible We d rather directly track coordinates of people in motion VRTI only tracks people in motion, not stationary people Change in mean and variance only two aspects of a r.v.

44 Goal: Statistical Inversion for Tracking What would Bayes do? Model simultaneous change in mean, variance, shape NEED: distribution parameterized by peoples locations

45 A Tale of Two Links 40 Link 1 Link 2 45 RSS (dbm) Time index

46 Outcome: Fade-Level based Model Link distributions are different, based on fade level: 5 If a link is in deep fade: RSS and variance increase when obstructed If a link is in anti-fade: RSS decreases when obstructed 5 J. Wilson and N. Patwari, "A Fade Level Skew-Laplace Signal Strength Model for Device-Free Localization With Wireless Networks", Submitted to IEEE Trans. Mobile Computing.

47 Fade-Level Model Justification P i,j is power in phasor sum of multipath. (a) When sum is in a null, change tends to pull it out (b) When in constructive sum, change will pull it down Target on LOS skew-laplace Measured Target on LOS skew-laplace Measured (a) Target off LOS Change in RSS (b) Target off LOS Change in RSS

48 Fade-Level based Model Fade level = RSS model - meas t: F = P(d i,j ) P i,j Determined during calibration (assume known sensor locations) Skew-Laplace pdf parameters: linear function of F

49 Particle Filter Need: track when meas ts are non-gaussian, non-linear Particle filtering: Bayesian coordinate est. given meas ts Convergence as links more / less likely using (a) 15% of meas ts, (b) 30% of meas ts Nodes Particles Known location 10 8 Nodes Particles Known location Y Coordinate (m) 6 4 Y Coordinate (m) (a) X Coordinate (m) (b) X Coordinate (m)

50 Tracking Results X Coordinate (m) Y Coordinate (m) Known Estimated Time Index Person walks in square path Estimate avg. error of 1 m Needs: proposal methods, human dynamics models

51 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

52 RF Sensor Networks Shadowing RTI Variance RTI Holistic RSS DFL Current Work Conclusion VRTI Repeatability (a) (b) (a) Original (Mar. 2009); (b) Repeat (May 2010)

53 Current Work: Noise Reduction Noise from regular motion can be characterized, removed: (a) (b) Compare VRTI estimators: (a) Tikhanov, (b) MAP

54 Commercialization RSS-based intrusion detection, tracking software Goal: Enable RSS DFL for practical commercial apps

55 Outline 1 RF Sensor Networks 2 Shadowing RTI 3 Variance RTI 4 Holistic RSS DFL 5 Current Work 6 Conclusion

56 Recap: RF Sensor Networks Link meas ts can benefit network in multiple ways. RSS-based device free location (DFL), Radio Tomography We can form statistical models of RSS change algorithms. Real-time imaging is possible as solution to linear equation Objects in environment change the RSS on links they cross.

57 Future: Large-scale Systems New node: RSS-sensing only, higher TX power Deploy across 10,000 sq. feet in building DFL in low-link density Multi-frequency / multi-band operation

58 Future: Deployable DFL Reconfigurable antennas Merge device & device-free: Better than either alone Simultaneous WSN localization and DFL Adaptive, real-time link model building Robust to attacks (DoS, PoM), node failure

59 Acknowledgements Dr. Joey Wilson University of Utah : Jessica Croft, Piyush Agrawal, Dustin Maas, Yang Zhao Prof. Sneha K. Kasera, School of Computing Career award ECCS , CPS award

60 RF Sensor Networks Shadowing RTI Variance RTI Holistic RSS DFL Current Work Conclusion Questions and Comments More info on