1. INSS 2009 June, 18 th 2009 Pittsburgh, USA Marcelo Martins, Hongyang Chen and Kaoru Sezaki University of Tokyo, Japan OTMCL: Orientation Tracking-based Localization for Mobile Sensor Networks

2. 2 Location awareness Localization is an important component of WSNs Interpreting data from sensors requires context Location and sampling time? Protocols Security systems (e.g., wormhole attacks) Network coverage Geocasting Location-based routing Sensor Net applications Environment monitoring Event tracking Mapping

3. 3 How can we determine location? GNSS receiver (e.g., GPS, GLONASS)  Consider cost, form factor, inaccessibility, lack of line of sight Cooperative localization algorithms Nodes cooperate with each other Anchor-based case: Reference points (anchors) help other nodes estimate their positions

4. The case of mobility in localization 4 Static networks One-time or low- frequency activity Refinement process Can be terminated Mobile networks Must be invoked periodically Update process Must exist as long as node moves

5. 5 Our goal We are interested in positioning low-powered, resource- constrained sensor nodes A (reasonably) accurate positioning system for mobile networks Low-density, arbitrarily placed anchors and regular nodes Range-free: no special ranging hardware Low communication and computational overhead Adapted to the MANET model

6. 6 Probabilistic methods Classic localization algorithms (DV-Hop, Centroid, APIT, etc.) compute the location directly and do not target mobility Probabilistic approach: explicitly considers the impreciseness of location estimates Maximum Likelihood Estimator (MLE)‏ Maximum A Posteriori (MAP)‏ Least Squares Kalman Filter Particle Filtering (Sequential Monte Carlo or SMC)‏

7. 7 Sequential Monte Carlo Localization Monte Carlo Localization (MCL)‏ [Hu04] Locations are probability distributions Sequentially updated using Monte Carlo sampling as nodes move and anchors are discovered (Movement)PredictionFiltering

8. MCL: Initialization Initialization: Node has no knowledge of its location. L 0 = { set of N random locations in the deployment area } Node’s actual position 8 Node’s estimate

9. MCL: Prediction Node’s actual position Prediction: New particles based on previous estimated location and maximum velocity, v max 9 Node’s last estimate

10. Filtering Indirect Anchor Within distance (r, 2r] of anchor Direct Anchor Node is within distance r of anchor a a 10

11. MCL : Filtering Node’s actual position r Anchor Invalid samples Invalid samples Binary filtering: Samples which are not inside the communication range of anchors are discarded 11

12. Re-sampling 1.Repeat prediction and filtering until we obtain a minimum number of samples N. 2.Final estimate is the average of all filtered samples 3.If no samples found, reposition at the center of deployment area (initialization) 12

13. Other SMC-based works MCB [Baggio08] Better prediction: smaller sampling area using neighbor coordinates MSL [Rudhafshani07] Better filtering: use information from non- anchor nodes after they are localized Samples are weighted according to reliability of neighbors (non-binary filter) 13

14. Issue: Sample degradation Problem 1: Predicted samples with wrong direction or velocity Problem 2: Previous location estimate is not well-localized Why don’t we tell where samples should be generated? 14

15. Proposal: Orientation Tracking-based Monte Carlo Localization (OTMCL) Orientation Tracking Sensor information to predict direction of movement Discover direction using a set of sensors (e.g. gyroscope, accelerometer, magnetometer) Prediction: Generated samples move inside disc area determined by α (measured angle) and β (variance) Re-sampling: If no samples are found, perform dead reckoning 15

16. Sensor bias Inertial sensor is subject to bias due to Magnetic interference Temperature variation Erroneous calibration Affects velocity and orientation estimation during movement Lower localization accuracy No assumptions about hardware Analyses use 3 categories of nodes for OTMCL based on β High-precision sensors ( β = 10 o ) Medium-precision sensors ( β = 30 o, β = 45 o ) Low-precision sensors ( β = 90 o ) 16

17. Analysis – Convergence time 17 Simulation Parameters Area: 500 x 500 m 2 Number of nodes: 320 Number of anchors: 32 Sample set: 50 Anchor density: 1 Node density: 10 Radio range: 50 m Max velocity: 10 m/s Mobility model: RWP* Full mobile scenario OTMCL achieves a decent performance even when the inertial sensor is under heavy bias relative to communication range ~ 7m stabilization phase

18. 18 Analysis – Communication overhead Reducing power consumption is a primary issue in WSNs Limited batteries Inhospitable scenarios Assumes no data aggregation, compression OTMCL needs less information to achieve similar accuracy to MSL

19. Analysis – Anchor density 19 OTMCL is robust even when the anchor network is sparse Simulation Parameters Area: 500 x 500 m 2 Number of nodes: 320 Number of anchors: 32 Sample set: 50 Node density: 10 Radio range: 50 m Max velocity: 10 m/s Mobility model: RWP* Full mobile scenario

20. Analysis – Speed variance 20 As speed increases, the larger is the sampling area  lower accuracy Simulation Parameters Area: 500 x 500 m 2 Number of nodes: 320 Number of anchors: 32 Sample set: 50 Anchor density: 1 Node density: 10 Radio range: 50 m Mobility model: RWP* Full mobile scenario

21. Analysis – Communication Irregularity 21 OTMCL is robust to radio irregularity. Dead reckoning is responsible for maintaining accuracy Degree of Irregularity [He03]

22. 22 Conclusion Monte Carlo localization Achieves accurate localization cheaply with low anchor density Orientation data promotes higher accuracy even on adverse conditions (low density, communication errors) Our contribution: A positioning system with limited communication requirements, improved accuracy and robustness to communication failures Future work Adaptive localization (e.g., variable sampling rate, variable sample number) Feasibility in a real WSN

23. Thank you for your attention martins@mcl.iis.u-tokyo.ac.jp 23

24. APPENDIX 24

25. OTMCL: Necessary number of samples Estimate error fairly stable when N > 50 25

26. Analysis – Regular node density 26 OTMCL is robust even when the anchor network is sparse Simulation Parameters Area: 500 x 500 m 2 Number of nodes: 320 Number of anchors: 32 Sample set: 50 Anchor density: 1 Radio range: 50 m Max velocity: 10 m/s Mobility model: RWP* Full mobile scenario

27. 27 Is it feasible? (On computational overhead) Impact of sampling (trials until fill sample set) AlgorithmAvg. # of sampling trials (DOI = 0.0) MCL1933.1077 MCB559.796 MSL2401.2508 ZJL597.8802 OTMCL (β = 10º)391.6977 OTMCL (β = 45º)746.1909 OTMCL (β = 90º)1109.4819

28. Radio model  Upper & lower bounds on signal strength  Beyond UB, all nodes are out of communication range  Within LB, every node is within the comm. range  Between LB & UB, there is (1) symmetric communication, (2) unidirectional comm., or (3) no comm.  Degree of Irregularity (DOI) ([Zhou04]) 28