Senseapp13 keynote

1. AUTONOMOUS SYSTEMS LABORATORY | COMPUTATIONAL INFORMATICS Dr. Raja Jurdak Principal Research Scientist Adjunct Associate Professor Research Group Leader University of Queensland CSIRO University of New South Wales Long-term Continental Scale Tracking of Flying Foxes SenseApp 2013

2. Continental Scale Tracking • Continuously track the position and state of small assets for long durations 2 |

3. Continental Scale Tracking • Continuously track the position and state of small assets for long durations • Why is it important for Australia? • Challenges: sparse population, large landmass • Applications: agriculture, biosecurity, logistics 3 |

4. Continental Scale Tracking • Continuously track the position and state of small assets for long durations • Why is it important for Australia? • Challenges: sparse population, large landmass • Applications: agriculture, biosecurity, logistics 4 | The Computing Challenge • Need to use energy hungry GPS • Operate within very tight energy budgets – Weight (30-50g) – Mobility (100s of kms a day)

5. Tracking – Current status 5 | Duration SamplingFrequency Short-term frequent sampling Long-term sparse sampling

6. Tracking – Our goal 6 | Duration SamplingFrequency Short-term frequent sampling Long-term frequent sampling Long-term sparse sampling

7. Tracking Flying Foxes The National Monitoring Program Disease Vector • Hendra Virus • Ebola in Asia/Africa • Coronavirus in Persian Gulf Seed Dispersal • Bio Security Behaviour • Not well understood • Threatened species Interaction • With other flying foxes • With other animals 7 |

8. Delay-Tolerant Networking 8 | Individuals travel between different camps and other locations Store sensor/position samples locally in flash Upload using short-range radio to gateway (3G) at known camps Base A Base B Base C DB

9. Camazotz 9 | • Multimodal sensing platform • Low power SoC R. Jurdak, P. Sommer, B. Kusy, et al. “Multimodal Activity-based GPS Sampling,” IPSN 2013.

10. Batmon Basestation 10 | ~100m radio range 3x30min active/day

11. The BatMAC protocol Every node has an assigned slot based on node ID slot = nodeID % number_of_slots Mobile nodes send announcement beacons every 5 minutes: • Node ID • Application version (e.g. 1.4) • Maximum available flash page • Voltage 11 | 1 2 3 4 5 6 7 8 9 10 1 10 slots = 5 minutes 12:00 12:01 12:02 12:03 12:04 12:05 P. Sommer, B. Kusy, A. McKeown, and R. Jurdak, "The Big Night Out: Experiences from Tracking Flying Foxes with Delay Tolerant Wireless Networking," In proceedings of (RealWSN), Como Lake, Italy, September 2013.

12. Remote Procedure Calls (RPC) Response-Request protocol • Client (basestation) • Server (mobile node) • Stateless Examples: time_get(), time_set(…), reboot() Implementation • Each RPC has a unique identifier (keys.txt) • Encapsulated into a single radio packet: identifier (2-bytes), payload (n-Bytes) 12 | time_get() 2013-07-22 09:30:00 NodeBase

13. Practical Challenge: Catching Flying Foxes 13 |

14. 14 | Current trackers work well.. but not for long

15. Can we continuously track position with bounded uncertainty? 15 | GPS Scheduler 2. GPS Duty Cycling 3. Activity Detection 4. Mobility Models GPS Sampling times and Frequencies 1. Energy Estimation

16. Online Energy Estimation • How to estimate battery state of charge? • Battery voltage unreliable in mid-range, yet relatively accurate in extreme states • Software metering captures instantaneous net energy flow yet is susceptible to long term drifts 16 | 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 State of Charge (SOC) 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4.0 4.1 BatteryVoltage[V] 29 62 99 119 240 468 1000

17. 17 | Long-term drift InstantaneousSOC variance Battery Voltage Software Bookkeeping Conflation- based approach

18. SOC estimation through conflation 18 | P. Sommer, B. Kusy, and R. Jurdak,"Energy Estimation for Long-term Tracking Applications,", To appear in proceedings of the First International Workshop on Energy- Neutral Sensing Systems (EnSys), co-located with Sensys, Rome, Italy, November, 2013.

19. Baseline Radio GPS (hot start) Typical Power Profiles of Key Tasks 19 | 0 5 10 15 20 25 30 35 Time [s] 0 10 20 30 40 50 0 5 10 15 20 25 30 Time [s] 0 5 10 15 20 0 5 10 15 20 Time [s] 0.0 0.2 0.4 0.6 0.8 1.0 Current[mA]

20. SOC Estimation Results 20 | 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Day 0 1 2 3 4 5 Current[mA] GPS Data Download

21. GPS Duty Cycling 21 | GPS fix X Assumed position Real position Uncertainty

22. GPS Duty Cycling Power GPS back on when uncertainty estimate approaches bound <Project Title> | <Project Lead>22 | GPS off X Assumed position Real position Uncertainty X

23. GPS Duty Cycling Strategy Varying the AAU according to the animal’s distance from the fence Speed models AAU: absolute acceptable uncertainty Ugps: GPS chip uncertainty s: assumed speed tL: lock time R. Jurdak, P. Corke et al., "Energy-efficient Localisation: GPS Duty Cycling with Radio Ranging," ACM TOSN, Vol. 9, Iss. 2, May 2013. R. Jurdak, P. Corke, et al. "Adaptive GPS Duty Cycling and Radio Ranging for Energy-Efficient Localization,” Sensys 2010.

24. Exploiting Radio Proximity Data Animals naturally herd closely together GPS duty cycling vs GPS DC and contact logging Combining GPS duty cycling with short range radio beaconing

25. GPS Duty Cycling with Contact Logging Using neighbors as position anchors to bound uncertainty 25 | GPS off X Assumed position Real position Uncertainty X

26. GPS Duty Cycling with Contact Logging Using neighbors as position anchors 26 | X Assumed position Real position Uncertainty X

27. Contact Logging for Data Muling 27 | Camp A (Base) Camp B Feeding Location

28. Data Muling: Open Questions Question 1: How to detect contacts between animals? • Sending announcement beacons? What additional information to include? • Synchronization? Question 2: What information should we store locally? • Logging every contact? Duration of contact? • Only during nighttime? 28 |

29. Sensor-triggered GPS Sampling 29 | • Use one or more of the cheap on-board sensors to detect activities of interest and trigger GPS samples • Some activities of interest

30. Understanding activities • Videos 30 |

31. Sensor-triggered GPS samples (Accelerometer) • Compute average vector at rest  gravity • Compute angle between current vector and gravity • Detect sustained angular shifts above 90o • 100% accuracy in detecting 11 true events • Video footage as ground truth 31 | 1.4 1.5 1.6 1.7 1.8 1.9 2 x10 5 −2 0 2 4 Sample Accelerationprojectionon meanvector(G) 1.4 1.5 1.6 1.7 1.8 1.9 2 x10 5 0 100 200 Sample Angle−currentand gravity(degrees)

32. Sensor-triggered GPS samples (Audio) 32 | • Frequency peaks at 2-4Khz • Lightweight features are based on calculating the mean signal energy and counting the number of zero crossings of a 1024 sample sliding window with an overlap of 50% • Video footage as ground truth

33. Sensor-triggered GPS samples (Audio) 33 | • Frequency peaks at 2-4Khz • Lightweight features are based on calculating the mean signal energy and counting the number of zero crossings of a 1024 sample sliding window with an overlap of 50% • Video footage as ground truth

34. Multimodal Event dissociation • When one sensor is insufficient to capture event-of-interest • Example: how to dissociate interaction events involving a collared animal from interaction events involving nearby animals only? 34 |

35. Multimodal Event dissociation • When one sensor is insufficient to capture event-of-interest • Example: how to dissociate interaction events involving a collared animal from interaction events involving nearby animals only? 35 |

36. Multimodal Activity-based Localisation Collaredevents Nearbyevents Powerconsumption DetectedEvents AveragePowerConsumption(mW) Accelerometer MAL collared only MAL nearby only MAL all events 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 9 8 7 6 5 4 2 0 1 3 Audio 36 | L ocali sat ion A p p r oach A n im al int er act ion s C ollar ed A ll D issociat ed Duty cycled GPS X A cceleromet er-t riggered X A udio-t riggered X A ccel. A ND A udio X A ccel. OR A udio X X Table 5: M A L can det ect all event s and dissociat e int er act ion event involving collared animal or near by animals. in our simulations. We compare a baseline approach of a duty cycled GPS with a period of 20s with triggered GPS sampling approaches based on the accelerometer only, audio only, or on the combination of audio and accelerometer sensors. We group all detected ground truth interactions into events that meet the 25s to 1min duration constraint. A successful detection in our simulation is when the algorithm obtains at least one GPS sample during the event. During the given time window, the duty cycled GPS mod- ule remains active for a total of 451s (including lock times) and successfully obtains GPS samples during each of the four events of interest, yielding an overall node power consumption of around 33mW. Figure 13 summarises the results of sensor-triggered GPS sampling. The accelerometer- triggered GPS manages to detect only two events (only the events from the collared bat) with a cumulative GPS active Collaredevents Nearbyevents Powerconsum DetectedEvents Accelerometer MAL collared only MAL nearby only MAL all even 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 Audio Figur e 13: Per for mance of M A L acceler om et er - and audio-t rigger ed GP can be t uned t o capt ur e eit her int er act io of t he collar ed animal, or nearby int er act i only. M A L can also det ect and dissoc types of int er act ion event s wit h compar ab consumpt ion t o audio. alongside GPS. The ZebraNet project [5] reports position records for zebras every few minutes. I make the energy problem more tractable Zebra include a solar panel, which assume that the pan silient to normal animal activities. Positioning GPS only, and the nodes propagate their infor ﬂooding in order to facilitate data acquisition by sink. Dyo e al. [3] use a heterogeneous sensor ne

37. Field sensor data for motion-based tracking 37 | • Differentiate motion/non-motion states • Slow uncertainty growth in non-motion state • Use on-board compass for motion direction – ellipse

38. The Chicken or the Egg? • Algorithms that give frequent position estimates need ground- truth validation • If we had ground truth, the problem would have been solved 38 |

39. 39 | How to design and validate continuous energy-efficient tracking algorithms over representative data?

40. Mobility Modeling • Establish mobility dependencies • Temporal • Spatial • Environmental • Social • Characterise population and individual level movement statistics • Step size • Turning angle • Diffusion • Generate long-term synthetic data that captures the statistics of short-term empirical data 40 |

41. Open Challenges • Delay-tolerant … • Data storage (what to store or not) • Sampling (maximum information for energy buck) • Communication (priorities, fairness, throughput) • Energy management (consumption, harvesting, prediction) • Tradeoffs? • Mobility model-driven sampling • How to build the model without the data  adaptive models • How flexible do these models need to be? 41 |

42. To Sum Up • Ongoing work • Modeling mobility • Progressively longer field trials – from 30 to 150 then 1000 nodes • Continuous and near-perpetual tracking based on activities, mobility models • Continental Scale Tracking • Near-perpetual monitoring of position and condition • Very challenging yet interesting research problem with real application drivers • Creates agents for microsensing 42 |

43. Contributors • CSIRO • Brano Kusy • Philipp Sommer • David Westcott • Adam McKeown • Jiajun Liu • Kun Zhao • Navinda Kottege • Chris Crossman • Phil Valencia • Leslie Overs • Ross Dungavell • Stephen Brosnan • Wen Hu 43 | • QUT • Peter Corke • UQ • Neil Bergmann • UNSW • Salil Kanhere • Sanjay Jha • Ghulam Murtaza • Lukas Li Collaborators

44. AUTONOMOUS SYSTEMS LABORATORY | ICT CENTRE Dr. Raja Jurdak Research Group Leader, Pervasive Computing Principal Research Scientist rjurdak@ieee.org Thank You