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
Unlocking the Potential: Deep dive into ocean of Ceramic Magnets.pptx
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
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
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.
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
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
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]
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
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
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 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. 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
Notas del editor
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.