Successfully reported this slideshow.
Your SlideShare is downloading. ×

Senseapp13 keynote

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Loading in …3
×

Check these out next

1 of 44 Ad

Senseapp13 keynote

Download to read offline

Long-term outdoor localisation with battery-powered devices remains an unsolved challenge,mainly due to the high energy consumption of GPS modules. The use of inertial sensors and short-range radio can reduce reliance on GPS to prolong the operational lifetime of tracking devices, butthey only provide coarse-grained control over GPS activity. An alternative yet promising approach is touse context-sensitive mobility models to guide scheduling and sampling decisions in localisationalgorithms. In this talk, I will present our work towards continental-scale long-term tracking of flyingfoxes, as part of the National Flying Fox Monitoring Program in Australia, using a model-drivenapproach. At the core of our approach is the multimodal GPS-enabled Camazotz sensor node platformthat has been designed at CSIRO for flying fox collars, with a cumulative weight just under 30g. The project has already deployed tens of devices on live flying foxes, which have been operating in thefield for several months. We are using the data from these devices to build mobility models andalgorithms for designing the next generation of software, as we will progressively deploy more than1000 nodes within the coming months. The progressive deployment of nodes coupled with delaytolerance, constrained resources, and incremental feature development raises interesting systemschallenges and opportunities, which I will highlight. The talk will also provide a snapshot of thecurrent data collection effort, and draw lessons from our activities in this area over the past 18 months

Long-term outdoor localisation with battery-powered devices remains an unsolved challenge,mainly due to the high energy consumption of GPS modules. The use of inertial sensors and short-range radio can reduce reliance on GPS to prolong the operational lifetime of tracking devices, butthey only provide coarse-grained control over GPS activity. An alternative yet promising approach is touse context-sensitive mobility models to guide scheduling and sampling decisions in localisationalgorithms. In this talk, I will present our work towards continental-scale long-term tracking of flyingfoxes, as part of the National Flying Fox Monitoring Program in Australia, using a model-drivenapproach. At the core of our approach is the multimodal GPS-enabled Camazotz sensor node platformthat has been designed at CSIRO for flying fox collars, with a cumulative weight just under 30g. The project has already deployed tens of devices on live flying foxes, which have been operating in thefield for several months. We are using the data from these devices to build mobility models andalgorithms for designing the next generation of software, as we will progressively deploy more than1000 nodes within the coming months. The progressive deployment of nodes coupled with delaytolerance, constrained resources, and incremental feature development raises interesting systemschallenges and opportunities, which I will highlight. The talk will also provide a snapshot of thecurrent data collection effort, and draw lessons from our activities in this area over the past 18 months

Advertisement
Advertisement

More Related Content

Slideshows for you (20)

Similar to Senseapp13 keynote (20)

Advertisement

Recently uploaded (20)

Senseapp13 keynote

  1. 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. 2. Continental Scale Tracking • Continuously track the position and state of small assets for long durations 2 |
  3. 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. 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. 5. Tracking – Current status 5 | Duration SamplingFrequency Short-term frequent sampling Long-term sparse sampling
  6. 6. Tracking – Our goal 6 | Duration SamplingFrequency Short-term frequent sampling Long-term frequent sampling Long-term sparse sampling
  7. 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. 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. 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. 10. Batmon Basestation 10 | ~100m radio range 3x30min active/day
  11. 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. 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. 13. Practical Challenge: Catching Flying Foxes 13 |
  14. 14. 14 | Current trackers work well.. but not for long
  15. 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. 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. 17 | Long-term drift InstantaneousSOC variance Battery Voltage Software Bookkeeping Conflation- based approach
  18. 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. 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. 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. 21. GPS Duty Cycling 21 | GPS fix X Assumed position Real position Uncertainty
  22. 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. 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. 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. 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. 26. GPS Duty Cycling with Contact Logging Using neighbors as position anchors 26 | X Assumed position Real position Uncertainty X
  27. 27. Contact Logging for Data Muling 27 | Camp A (Base) Camp B Feeding Location
  28. 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. 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. 30. Understanding activities • Videos 30 |
  31. 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. 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. 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. 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. 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. 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 sen- sors. 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 con- sumption of around 33mW. Figure 13 summarises the re- sults 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. 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. 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. 39 | How to design and validate continuous energy-efficient tracking algorithms over representative data?
  40. 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. 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. 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. 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. 44. AUTONOMOUS SYSTEMS LABORATORY | ICT CENTRE Dr. Raja Jurdak Research Group Leader, Pervasive Computing Principal Research Scientist rjurdak@ieee.org Thank You

Editor's Notes

  • Why important to track flying foxes:

    Little is known about their behavior, track individual animals and capture its interactions
  • Since flying foxes frequently return to known location (camps), we can setup fixed infrastructure (base station) in those camps.

    Tracking data is stored locally and is uploaded to the base station upon request.

    Tens of thousands in roosting camps
  • Examples of event from flying foxes, use words interchangeably
  • Figure 5 also demonstrates the benefit of our approach over simply using ∆E for predicting SOC. Consider Node A during day 8 when the battery voltage reaches a minimum of around 3.6V. Results from Figure 2 indicate that the battery’s SOC is between 0.1 and 0.2. If we rely only on ∆E, we would predict that the battery is at a SOC of around 0.5, which is clearly incorrect and could lead to an aggressive sampling strategy with a flat battery. Our SOC estimate based on both battery voltage and ∆E provides an estimate of SOC between 0.1 and 0.2, which is much closer to reality. The battery volt- age anchors the ∆E estimates for a better overall estimate. Battery voltage on its own, however, is less useful in cap- turing short-term changes in SOC, as it is highly sensitive to instantaneous current loads and it varies slowly outside the extreme regions.
  • Proximity of flying foxes roosting camps

×