Content uploaded by Raja Jurdak
Author content
All content in this area was uploaded by Raja Jurdak on Feb 15, 2016
Content may be subject to copyright.
Information Bang for the Energy Buck:
Towards Energy- and Mobility-Aware Tracking
Philipp Sommer1,3, Jiajun Liu1,4 , Kun Zhao1, Branislav Kusy1,
Raja Jurdak1, Adam McKeown2, David Westcott2
1CSIRO, Data61, Brisbane, QLD, Australia
2CSIRO, Land & Water, Cairns, QLD, Australia
3ABB Corporate Research, Baden-Daettwil, Switzerland
4Renmin University of China, Beijing, China
philipp.sommer@ch.abb.com, jiajunliu@ruc.edu.cn, firstname.lastname@csiro.au
Abstract
Context-aware services rely critically on accurate and
energy-efficient location tracking. While GPS receivers of-
fer high accuracy positioning, energy harvesting and stor-
age constraints of battery-powered devices necessitate duty-
cycling of GPS to prolong the system lifetime. Further-
more, real-world dynamics dictate that the GPS sampling
strategy adapts in real-time to achieve optimal positioning
performance. We propose an information-based approach to
solve the problem of online adaptive GPS sampling. We es-
timate the current positioning error through dead-reckoning
and schedule a new GPS sample when the error exceeds a
given threshold. The threshold adapts based on the current
energy and movement trends to balance the expected infor-
mation gain from a new GPS sample with its cost for longer-
term tracking performance. We evaluate our approach on
empirical traces from wild flying foxes and compare it to
strategies that sample GPS using fixed and adaptive duty cy-
cles. Our analysis shows that our information-based GPS
sampling strategy reduces the tracking error over the fixed
and adaptive duty cycle strategies by up to 96% and 64% re-
spectively, and approaches the performance of the optimal
offline sampling strategy.
Categories and Subject Descriptors
C.2.1 [Computer-Communication Networks]: Wire-
less Communication
Keywords
Tracking, Energy Awareness, Harvesting, Scheduler
1 Introduction
Position tracking has become an essential building block
for many applications, ranging from location-based services
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. To copy otherwise, to republish, to post on servers or to redistribute
to lists, requires prior specific permission and/or a fee.
EWSN’16, February 15–17, 2016, Graz, Austria.
Copyright © 2016 ACM 978-1-4503-1169-4 ...$10.00
for mobile phones to wildlife tracking. Tracking can ei-
ther be user-initiated, where users request location from the
device, or autonomous, where the tracking devices collect
user location information based on events of interest. Ad-
vanced tracking systems have used additional sensor inputs
for autonomous detection of traffic conditions [24] or pot-
holes [10] in a city, tracking and evaluation of health of out-
patients [26], or sharing cycling experiences [9].
Long-term location tracking, however, remains a chal-
lenge, with GPS requiring high power consumption to con-
tinuously deliver accurate positions. High energy consump-
tion can be often traded off for lower localization accuracy,
for example, by triangulating the locations of nearby Wi-Fi
access points or cellular base stations. Energy improvements
can be substantial [16] when continuously tracking the user
context and selecting the most energy-efficient localization
algorithm. Advanced techniques can use accelerometers to
improve energy efficiency of location tracking [18, 25] and
magnetometers to improve trajectory tracking [17] by a fac-
tor of three or more.
Nearly all of the existing energy-efficient tracking algo-
rithms focus on minimizing energy consumption subject to
an application-specific position error bound. Because of the
predictability of their energy budget (for instance, mobile
phones are recharged every day), most methods track posi-
tion within given accuracy constraints and aim to minimize
energy consumption in the process. The accuracy constraints
then implicitly define the energy budget. For instance, mod-
ern smartphones enable users to select their location technol-
ogy based on available energy; a user can switch off GPS and
rely only on Wi-Fi or cell-based location in urban environ-
ments if the battery power indicator is low.
In this paper, we are interested in autonomous long-term
tracking that relies on energy harvesting. As opposed to hav-
ing a fixed daily energy budget, we investigate algorithms
that adapt to the changing nature of harvested energy. Our fo-
cus is to maximize the tracking accuracy given this dynamic
energy budget, which is the reverse problem of most exist-
ing energy-efficient tracking algorithms. Adaptive tracking
highlights the question of how to schedule GPS samples in a
given future period on the basis of the available energy. His-
torical movement patterns, such as typical daily trip times,
can be used to forecast movement in the near future and thus
guide the choice of parameters for a given GPS sampling
strategy. While historical patterns can be used for strategic
sampling forecasts, it is important to re-evaluate the GPS
sampling strategy periodically in case the current behavior
of the tracked object deviates from historical data.
Building on these concepts, we propose energy- and
mobility-aware tracking that minimizes tracking errors
within the available energy budget. Our strategy uses low-
power inertial sensors to detect object motion and schedules
GPS sampling within the object motion periods. Specifically,
we use an estimate of the available energy and a forecast
of the remaining movement duration until the next recharge
event to determine suitable parameters for GPS sampling.
We use historical movement statistics coupled with instan-
taneous estimates of displacement from the most likely des-
tination to forecast movement duration. More specifically,
we predict the remaining movement distance for a tracking
interval on the basis of three estimates: (1) population-level
movement distance; (2) individual-based movement distance
during the previous interval; and (3) current distance from
the most likely destination. Each of these estimates proves
to be more accurate in a specific timeframe within the track-
ing interval.
Both the available energy and the movement forecast are
updated in real-time on the mobile node and the GPS sam-
pling strategy is continuously adapted as new estimates be-
come available. We propose an information-based GPS sam-
pling strategy that considers expected movement duration,
energy availability, and expected information gain of the next
GPS sample. The ratio of the forecast movement duration
and energy budget determines an adaptive threshold for the
expected tracking error. Tracking error is estimated through
dead reckoning and, once it exceeds the adaptive threshold,
triggers a new GPS sample. This threshold balances the ex-
pected information gain from a new GPS sample with its cost
for longer-term tracking performance.
We retrospectively analyze high-frequency GPS traces
from wild flying foxes to evaluate our energy- and mobility-
aware tracking strategy against online tracking strategies that
sample GPS periodically during motion periods with fixed
and adaptive duty cycles, as well as an optimal offline track-
ing strategy. Our results confirm that information-based
tracking that relies on energy- and mobility-awareness deliv-
ers significant reduction in tracking error compared to static
and adaptive GPS sampling and performs close to the opti-
mum especially for longer movement durations.
2 Long-Term Mobility Tracking
Long-term tracking of small mobile entities is a challeng-
ing problem with high relevance in ecology, agriculture, and
logistics. The key constraint here is energy. The very need to
track mobile entities long-term implies that their location is
unknown and not readily accessible, which limits opportuni-
ties for manually recharging the battery of tracking devices.
An alternative approach is to support energy harvesting on
tracking devices, such as through solar panels, to replenish
energy supplies in situ. With energy harvesting, the avail-
able energy budget is subject to the amount of energy that
has been harvested in the recent past. Furthermore, decisions
to acquire position samples have to be made autonomously
and in real-time by the tracking device on the basis of both
available energy and likely movement activity as communi-
cating with the device for setting parameters manually might
not be feasible.
2.1 Accuracy- vs. Energy-bound Tracking
Most recent work has focused on accuracy-bound track-
ing. Given an uncertainty or accuracy bound on location,
the goal is to minimise energy consumption while operat-
ing within this bound to maximise lifetime. The accuracy
bound de-facto defines the minimum required energy bud-
get of the position tracking, given a typical motion pattern
of the tracked object. If the available energy resources are
smaller than the tracking budget, the device will run out of
energy and the accuracy-bound tracking will fail to deliver
the required performance.
With energy harvesting, a well-designed system can op-
erate near-perpetually as long as it tailors its consumption to
its available energy budget. Given this implied longevity for
trackers that use energy-harvesting, the focus becomes on
how to best use the available energy to maximise the loca-
tion accuracy within a specific period of time, i.e., the time
period between two energy-harvesting events. We refer to
this problem as energy-bound tracking. Solving this prob-
lem naturally depends on the sampling frequency of GPS
and other supporting sensors, which is defined by the un-
derlying movement dynamics of the tracked object. Objects
that move rarely can be optimally tracked through a sparse
scheduling strategy, while determining the best schedule for
tracking highly mobile objects with high accuracy is a chal-
lenging problem. For this reason, our work here not only
considers the energy availability through harvesting sources,
but also the mobility of the tracked objects.
2.2 Limitations of Static GPS Sampling
Current commercial motion trackers use a combination
of motion-triggering and time-based duty cycling of GPS
for energy management. In particular, motion sensors are
used to detect the start and end of motion events according
to a given movement threshold and GPS samples are only
taken during the motion events. Time-based approaches set
a fixed duty cycle for GPS. A simple combination of the
two approaches uses motion sensors to detect motion and
periodically duty cycles GPS within the movement period
to reduce energy consumption. We refer to this approach
as static motion-based tracking, which effectively fixes the
GPS sampling strategy for the duration of a typical motion
event. This approach does not work well when the mobility
dynamics or the available energy vary day-to-day. A static
scheduler might miss key parts of a trajectory if it underes-
timates the length of the trajectory or the energy availabil-
ity. Autonomous long-term tracking applications, therefore,
need to adapt GPS sampling in order to reconstruct the orig-
inal trajectory as accurately and efficiently as possible.
To illustrate the energy and motion duration dynamics,
we present data from the flying fox tracking application in
Figure 1. We show an example of two trips made by the
same animal that vary in duration. We also note large dif-
Roosting Camp
Trip B
Trip A
1km
N
Foraging
Foraging
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Day
-80
-40
0
40
80
Energy [mAh]
Energy harvested Energy used
0.0
0.2
0.4
0.6
0.8
1.0
SOC
State of Charge (SOC)
Figure 1: Two GPS trajectories of a spectacled flying fox (Pteropus conspicillatus) tagged with a mobile tracking device (top).
Both example trajectories start and finish at the same roosting camp, but feature different motion patterns and durations. Dashed
circles indicate the animal’s foraging areas. Daily harvested solar energy, consumed energy of the tracking device, and battery
state-of-charge (SOC) for the same animal during a period of four weeks (bottom).
ferences in harvested energy for an individual animal on a
day-to-day basis over a 4-week period of empirical data, il-
lustrating the likelihood for differences in energy intake and
movement duration among individuals and over time for the
same individual.
2.3 Motivating Application
Our approach can be applied to tracking any mobile ob-
jects that have regular motion patterns and periodic energy
harvesting intervals, such as tracking commuting times of
people to work, tracking people during leisure activities such
as biking, roller-blading, or jogging, and tracking domes-
tic, farm, or wild animals. We focus our evaluation on a
specific application of tracking flying foxes in their habitat
across Australia. Flying foxes, also known as fruit bats, play
an important ecological role as dispersers of pollen and seed
in wet tropical areas of Africa and Asia. They come into
conflict with humans when their roost sites are located in
urban areas, because of their raiding of fruit crops and be-
cause they are vectors for a range of emerging infectious dis-
eases which have serious consequences for human health,
e.g. Hendra virus in Australia, Nipah in Asia and Ebola in
Africa. Flying foxes are highly mobile, with individuals able
to fly hundreds of kilometers in a night’s foraging. Key to
understanding and managing these animals is to understand
how they utilize landscapes and how they interact with other
disease hosts. This requires a fine-grained understanding of
their movement. However, their weight (600 g to 1 kg) lim-
its the weight and size of tracking devices that can be placed
on them [31, 23], imposing a tight energy budget on sensing
activities, such as GPS sampling.
Behavioral patterns. Flying foxes are nocturnal animals
and during the daylight rest in large groups at sites called
camps or roosts. During the day they generally only move
within the camp (within a radius of about 100-200 meters)
and for most of the time they sleep at a single location. In
contrast to the family of microbats, which are common in
Europe and the US, flying foxes do not roost in caves, but
hang upside down from the top branches of large trees ex-
posed to direct sunlight. At dusk, they leave the camp and fly
out to forage at sites that are usually tens of kilometers away.
During the night they may change location several times be-
fore returning to the camp before sunrise. The long flying
journeys between the camp and the foraging area are called
commutes. Figure 1 shows the GPS positions collected by
the collar of a single animal while it is flying from the roost-
ing camp towards two separate foraging areas on different
nights. From an ecological perspective, the movements be-
tween camps and foraging sites determine how pollen, seeds
and disease spread through the landscape, and the forage lo-
cations determine the characteristics of the interactions with
disease hosts (other animals, humans). Therefore, it is im-
portant that the tracking data captures the movement of the
tagged animal during the whole time interval of interest.
The discussion in this section has shown that the mobility
dynamics and available daily energy can vary widely among
animals or even for the same animal. Thus, a key require-
ment for maximising tracking accuracy is for the tracker to
optimally invest its available energy based on the expected
movement trends to guide the selection of optimal GPS sam-
pling parameters.
3 Energy and Mobility Estimation
In this section, we introduce our framework for loca-
tion tracking of mobile nodes that adapts to changing en-
ergy availability and motion patterns. At its core is the Task
Energy Monitoring
Mobility Prediction
GPS Accelerometer Magnetometer
Battery
Harvesting
Energy-Awareness Layer Mobility-Awareness Layer
Motion Detection
Temperature
Pressure
Sensor Process Layer
Scheduling Layer
Task Scheduler
Energy
Book-keeping Sensor Output
start/stop processes
Task 1 Task 2 Task 3 Task N
Figure 2: Building blocks of the energy- and mobility-aware scheduling framework.
Scheduler component, which uses inputs from the Energy-
Awareness and Mobility-Awareness layers to schedule sam-
pling of different hardware components in the Sensor Pro-
cess layer, such as the GPS receiver or inertial sensors. The
building blocks of the framework are depicted in Figure 2.
3.1 Task Scheduler
We assume that physical or remote access to the mobile
node may not be feasible for prolonged time periods and de-
velop mechanisms to support autonomous operation and re-
configurability. Specifically, the scheduler allows for adap-
tive execution of sensing tasks based on the input from en-
ergy and mobility layers. Sensing tasks are associated with
a specific process type, e.g., GPS or inertial sensor process,
and can be executed or terminated by the scheduler at any
time. This design principle allows us to adapt operating con-
ditions of sensing tasks in real-time. For example, the user
can configure high-frequency GPS sampling during night
time when the animals are flying to foraging areas, while
hourly GPS fixes can be acquired during day time when ani-
mals are within the roost or if available energy does not sup-
port high-frequency tracking.
3.2 Energy Awareness
Precise book-keeping of both energy consumption and
energy harvesting is key to maintaining a realistic estimate
of the available energy. The Energy Monitoring component
keeps track of the available energy stored in the batteries, the
so called State of Charge (SOC), the energy consumed by
the sensor node Eused, and the energy harvested by the solar
panels Eharvested.
Energy Consumption and Harvesting. We keep track of
the energy consumption during runtime using a software-
based book-keeping mechanism. The instantaneous over-
all consumption of a sensor node can be broken down into
the power consumption of its individual hardware compo-
nents i, which includes the microcontroller, sensors, radio
transceiver, and flash storage chip. The total instantaneous
power consumption P(t)of the sensor node can be calcu-
lated as the sum of the different components:
P(t) = ∑
i
Pi(t) = ∑
i
Vi(t)·Ii(t)(1)
where Pi(t)is the instantaneous power consumption of com-
ponent i, which depends on the voltage Vi(t)and current con-
sumption Ii(t)of its operating state. We assume that the soft-
ware is aware of the active operating state of a component
and that its current consumption is known. Consequently, we
can derive the energy consumed by the node Eused, within a
specific time interval.
The hardware platform is able to measure the solar panel
voltage Vsolar(t)and the charge current Isolar(t)into the bat-
tery while solar energy harvesting is in progress, which al-
lows to estimate the harvested energy Eharvested per interval.
Separation of Harvesting and Motion Periods. The be-
havioural patterns of flying foxes imply separation of periods
when the energy is used and when the energy is harvested.
Interestingly, such a scenario is common for mobile devices
like cellular phones or personal health monitors: the devices
are typically charged at night and used during the day. We
note that the scheduling framework presented in this paper
can operate regardless of this separation.
State of Charge Estimation. Accurate estimation of the
SOC of a battery is a non-trivial problem, since the battery
voltage Vbattery depends on the instantaneous current draw
and remains relatively constant over a large fraction of the
SOC range. Consequently, measuring the instantaneous bat-
tery voltage alone can only provide a rough estimate of the
state of charge. A large body of literature exists on different
methods for estimating the SOC of batteries in embedded de-
vices. Approaches range from dedicated hardware chips for
energy estimation [7, 13] to software metering approaches
that utilize direct battery voltage measurements [2]. In this
paper, we use a conflation based method to estimate the
SOC, which combines battery voltage measurements with
the book-keeping of energy harvesting and energy consump-
tion. Using this approach, software book-keeping can esti-
mate the energy outputs within 10% [28]. Combined with
the measurement of solar charge current, we can accurately
estimate SOC of the mobile device in software. We period-
ically update the available energy (SOC) at the end of the
interval [t,t+τ)as follows:
SOC(t+τ) = SOC(t) + Eharvested(t+τ)−Eused(t+τ)(2)
where Eharvested(t+τ)denotes the harvested energy during
the interval [t,t+τ)and Eused(t+τ)is the total energy con-
sumed by different system components.
3.3 Mobility Awareness
It is critical to capture the underlying mobility dynam-
ics of moving objects to predict their future movement accu-
rately. As these dynamics are largely dependent on the type
of tracked object, we focus on our motivating application of
long-term tracking of flying foxes.
-2
-1
0
1
2
Acceleration [g]
Accelerometer (z-axis)
0 15 30 45 60 75 90 105 120
Time [s]
0
180
360
Heading [°]
Magnetometer GPS
0
5
10
15
Speed [m/s]
GPS
Figure 3: Timeline of z-axis accelerometer and GPS speed
(top), and magnetometer heading and GPS speed (bottom)
as the animal starts flying at t=30 s.
3.3.1 Motion Detection
In order to facilitate activity detection based on low-
power sensor inputs, many mobile tracking platforms feature
inertial sensors. For example, the Camazotz hardware plat-
form (see Section 5.1) is equipped with a combined 3-axis
accelerometer and 3-axis magnetometer in a single sensor
chip. The accelerometer provides samples of the acceleration
along its three axes at a sampling frequency between 1 Hz
and several kHz. The magnetometer measures the strength of
the magnetic field at a data rate between 0.75 Hz and 220Hz
along three axes, which can detect heading changes.
Conditional Task Scheduling. In addition to continuous
sampling of acceleration and magnetic field, the accelerom-
eter further provides a programmable interrupt for detection
of motion events, e.g., when the animal changes from roost-
ing to flying. This functionality can unburden the microcon-
troller from having to continuously read samples from the
accelerometer’s output buffer so that it can remain in sleep
mode for most of the time. This interrupt will then set the
corresponding status flag in the scheduler, which can trigger
start or termination of tasks when motion has been detected
(see Figure 2).
Classification of Activity. We show a real-world example
for activity detection using inertial sensors on a flying fox
tagged with a Camazotz node in Figure 3. GPS data is used
to provide ground truth on the animal motion state, motion
duration, and heading. In our application scenario, contin-
uous sampling of accelerometer at 10 Hz provides enough
information to detect animal motion with high accuracy.
Specifically, our motion classifier accumulates acceleration
changes along the z-axis in each 5 second window. The ani-
mal is classified as flying in the time window, if the average
accumulated acceleration change is above a threshold. We
use a population-level threshold, the value of which we learn
based on a high-frequency dataset with both GPS and ac-
celerometer data that we collected in the initial phase of the
animal tracking project.
Validation. We perform 5-fold cross-validation for 20 rep-
etitions on the GPS and accelerator samples that span 137
minutes. In the overall dataset, 47% GPS samples are la-
beled as the animal being flying, and 53% samples as sta-
tionary. For each repetition, we randomly split the samples
into 5 independent folds, and use one fold to train the thresh-
Algorithm 1 Mobility Prediction
Require: dcamp(t)Current distance to camp
Require: dprev_max Previous max distance from camp
Require: ∆tmotion Estimate for motion duration
Require: dtotal(d)Estimate of total flight distance
Require: tstart,tend Observation interval
procedure PREDICT_REMAINING_DIS TANC E(t)
if dcamp(t)<dprev_max then
α←t−tstart
tend−tstart Weight factor
dremaining(t)←(1−α)·∆tmotion ·vavg +α·dcamp(t)
else
dremaining(t) = dtotal (dcamp(t))
end if
end procedure
old, which is later used in the testing phase on the rest of the
data. The speed estimate provided by the GPS is used as the
ground truth. If the GPS speed of the animal is greater than
1 meter per second, the animal is considered flying as the
ground truth. In this evaluation, our classifier reaches 0.986
average precision and 0.999 average recall in 20 repetitions.
3.3.2 Prediction of Movement Duration
Optimal scheduling depends both on the available energy
budget and the amount of energy that will be used for track-
ing movements. Therefore, it is important that the GPS pro-
cess can estimate for how long the node will be moving
within a certain time interval. Clearly, the accuracy of our
prediction will directly influence the energy consumption of
the tracking algorithm. If the actual total tracking time ex-
ceeds the predicted time, we risk running out of energy early
and we sacrifice accuracy of tracking towards the end of the
observation period. On the other hand, overestimating the to-
tal tracking time will result in conservative budgeting of en-
ergy resources and may lead to larger localization errors. The
implementation of the mobility prediction module is specific
to the designated application scenario and might take into ac-
count input from several sensors, last location from GPS, the
current time/date information and historical data.
We developed a hybrid approach to predict animal mo-
bility based on individual- and population-based models and
real-time estimation of the remaining motion duration based
on the remaining distance to the previous roosting camp
(see Algorithm 1). We start by using an initial estimate for
the total flight distance based on the previous night or an
individual-based model. If we detect that the animal has
exceeded the distance from camp observed in the previous
night dprev_max, we switch to a population-based model to
estimate the total flight distance based on historical trajecto-
ries.
Mobility models. We assume that we are given a set of his-
torical trajectories that an object of interest traverses and a
certain level of regularity of motion that allows us to extract
simple motion features. Note that this motion regularity is
shared by many other species including humans [30]. An es-
timate for the motion duration of an animal per night ∆tmotion
can be calculated by observing typical animal behavior in
historical data (see Section 5). Tracking of new animals
can be done using a population-level motion model which
maps the current distance from the camp to the total flight
distance dtotal(dcamp (t)), while a more accurate individual-
level model can be derived over time as more data becomes
available.
Distance to home base. Tracked objects may exhibit large
variations in motion patterns across the population, or across
time (see Figure 1). Prediction of motion duration that is
based on simple statistics only, such as the mean duration of
motion, will therefore be inaccurate. We observe that fly-
ing foxes usually return to the same roosting camp after the
nightly foraging (similarly to humans commuting to/from
home and work). The current location of the animal can thus
be used as a lower bound on the remaining flight distance.
This provides us with a lower bound on the remaining motion
time, which in turn determines the lower bound on energy
required to capture the return trip to the camp. Clearly, this
approach can only provide a good estimate after the animal
has reached the maximum distance from the camp, which
usually happens in the second half of the night. We there-
fore employ a linear weighting factor αwhich applies more
weight to the individual- or population-based mobility pre-
diction in the beginning of the observation period, while the
distance to base becomes more important towards the end of
the observation period.
4 GPS Sampling Strategies
The scheduling framework presented in the previous sec-
tion is agnostic to the actual GPS sampling strategy, and it
assumes that each strategy will tune its sampling parame-
ters based on the given constraints and task configuration. In
this section, we present three online and one offline strategy
to configure sampling of the GPS receiver. The remainder
of this section explains in detail how each of the strategies
schedules GPS samples.
4.1 Static Motion-based GPS Tracking
The GPS process can be configured to acquire periodic
single GPS samples based on motion events within a spe-
cific time window. For example, the GPS task can be con-
figured for periodic execution (with period Tsampling) within
the motion period ∆tmotion. By specifying the corresponding
flag in the task entry conditions we can configure the GPS
task to only start when motion has been detected by the in-
ertial sensor process. On each execution of the task a single
GPS sample is acquired and the GPS receiver is set to sleep
mode afterwards. The energy consumption in the time inter-
val ∆tinterval =tend −tstart is proportional to the GPS schedul-
ing period and can be calculated as follows:
Eused =∆tinterval ·Pbaseline +k·Thotstart ·Ptracking (3)
Pbaseline denotes the baseline power consumption of the sys-
tem, Ptracking is the additional power used during tracking op-
eration, kdenotes the number of samples taken, and Thotstart
is the average duration of a GPS hotstart (see Section 5.4).
The number of GPS samples that will be collected can be
calculated as k=b∆tmotion
Tsampling c.
Performance of the static GPS scheduler critically de-
pends on our initial estimates of the available energy re-
sources and motion patterns of the tracked object, which de-
termine the constant GPS sampling interval Tsampling. Specif-
ically, the optimal GPS sampling interval can be calculated
using the following inequality: Eused ≤Eharvested.
We can get estimates for both the harvested energy and
motion duration by studying typical behavior of a popula-
tion of animals. However, any individual deviations from the
typical behavior, for example, due to weather or food avail-
ability changes, will result in a suboptimal GPS sampling
strategy.
4.2 Adaptive Motion-based GPS Tracking
The adaptive motion-based approach schedules the maxi-
mum number of GPS samples given a certain energy budget
and an estimate of motion duration. Therefore, it will select
a GPS duty cycle that balances between energy consump-
tion and number of GPS samples taken. In contrast to static
motion-based tracking that sets a fixed Tsampl ing for the whole
time window, the adaptive approach continually updates the
GPS sampling interval to meet the energy constraints. For-
mally, the energy used Eused should be lower than the energy
budget Ebudget. Therefore, the resulting upper bound for the
number of GPS samples kcan be calculated as follows:
k(t)≤Ebudget −∆tinterval(t)·Pbaseline
Thotstart ·Ptracking
(4)
By distributing the kGPS samples uniformly over the pre-
dicted duration of motion ∆tmotion(t), we obtain the GPS sam-
pling interval Tsampling(t) and the corresponding duty cycle
D(t) as follows:
Tsampling(t) = ∆tmotion (t)
k(t)D(t) = k(t)·Thotstart
∆tmotion(t)(5)
4.3 Information-based Scheduling of GPS
Periodic duty-cycling of GPS does not take into account
the intrinsic characteristics of the trajectory, for example
sharp turns, to identify key points that will result in small
errors when used as input for interpolation. Sampling iner-
tial sensors while the node is in motion can provide a rough
estimate of the current heading while using only a small frac-
tion of the power consumption of the GPS (see Figure 3). By
employing dead-reckoning using the magnetometer heading,
the current position can be extrapolated using the last known
GPS position and speed vgps. In each time step, we calcu-
late the tracking errors e[ti]between dead-reckoning-based
position estimates at time tiand the corresponding interpo-
lated positions on a straight line between the last known GPS
position and the most recent position estimate (see Figure 4).
GPS Position e[ti]
vgps
ti=0ti=1ti=2ti=3ti=4
e[ti]
e[ti]
vgps vgps
vgps
Dead-Reckoning Position
Figure 4: Dead-reckoning based on heading estimation us-
ing magnetometer data. The current position is extrapolated
from the last known GPS position and speed vgps.
If our estimation of the trajectory based on dead-
reckoning exhibits small estimated tracking errors e[ti], there
is little benefit in taking another GPS sample. On the other
hand, a large estimated tracking error increases the value of
taking another GPS sample to avoid incurring even larger er-
rors in the future. A decision criterion is therefore needed
to define the extent of error that warrants taking a new GPS
sample. We calculate a threshold R(t)for the expected track-
ing error based on the duty cycle D(t)using the following
equation:
R(t) = R0·1
D(t)−1β(6)
As soon as the estimated maximum tracking error emax =
maxie[ti]exceeds the threshold R(t)another GPS sample is
acquired, which will restart the dead-reckoning phase. The
intuition is to set the GPS sampling policy based on the re-
lationship between the expected movement duration and the
energy budget. The ratio of movement duration and energy
budget expressed in remaining GPS sample duration is ef-
fectively the inverse of D(t)in Eq. 5. As a result, R(t)will
increase when the remaining energy budget decreases, which
will yield fewer GPS samples and a more conservative sam-
pling policy. On the other hand, the calculated threshold de-
creases if the remaining energy budget and the expected mo-
tion duration are similar (when the expected movement dura-
tion roughly matches the budgeted GPS samples), allowing
for more aggressive GPS sampling. R0and βare tuning pa-
rameters for calibrating the scale of the threshold. In case
of limβ→0,R(t)converges to a fixed threshold R0. The most
suitable settings for R0and βdepend on the characteristics
of the trajectories and can be learned from historical data.
4.4 Optimal Offline Scheduling of GPS
As a performance benchmark, we consider an offline
oracle-based optimal strategy that provides a lower bound
on the tracking errors for a given energy budget by finding
the optimal sampling points. The offline optimal strategy
cannot be implemented on a mobile device since it requires
knowledge of the complete trajectory. Let a sequence P=
{1,2,...,N}denote the points of the ground truth trajectory.
Suppose we take GPS samples at the first and last points of
the trajectory by default, and the energy budget only allows
us to take at most madditional GPS samples to form an es-
timated trajectory. The optimal estimated trajectory is a sub-
sequence of P, denoted by S={1,i1,...,ik,...,i|S|−2≤m,N},
which has the minimum worst-case tracking error εopt =
argminεmax(S)(see Section 6). This problem is equivalent to
the min-εproblem in curve approximation and can be solved
by a graph-search approach in polynomial time [4]. The ba-
sic idea is to construct a directed graph G(V,E)with the ver-
tices set V=Pand edges set E={(i,j)|di j ≤ε,i<j}and
use Gto find the optimal estimated trajectory. Here di j is
the tracking error between the estimated trajectory ˆ
i j to the
ground truth trajectory. A shortest path from vertex 1 to ver-
tex Nin Gthen represents the optimal estimated trajectory
with tracking error ε.
Figure 5: Camazotz collars (left) and a spectacled flying fox
(Pteropus conspicillatus) with collar (right).
5 Experimental Setup
In this section, we evaluate the performance of our track-
ing framework based on experimental data gathered during a
deployment on wild flying foxes.1
5.1 Hardware Platform
In this paper, we employ the Camazotz platform [15], a
low-power sensor platform for GPS location tracking, which
is optimized for small size and low weight (see Figure 5).
The platform is based on the Texas Instruments CC430F5137
system-on-chip, which combines a microcontroller with a
short-range radio transceiver in the 900 MHz band.
Energy Supply and Harvesting. Due to the stringent con-
straints in terms of weight and form factor, Camazotz em-
ploys a small single-cell Lithium-Ion battery with a capacity
of 300 mAh and relies on energy harvesting to recharge the
battery using a 34x22 mm2photovoltaic panel mounted on
the outside of the node enclosure. The panel is rated at a
maximum current of 40 mA in direct sunlight, however, we
often observe a smaller cumulative daily charge between 10
and 50 mAh under typical conditions on animal collars, e.g.,
on flying foxes roosting in trees.
Satellite-based Positioning (GPS). The Camazotz platform
is equipped with the u-blox MAX-6 GPS module, which
supports a backup mode (15 µA) where the GPS receiver
is disabled, but satellite orbit information (almanac and
ephemeris) are retained and the internal real-time clock is
still operational. This allows to reduce the idle energy con-
sumption while allowing GPS hotstarts.
Sensors and Communication. The Camazotz platform inte-
grates several sensing modalities such as accelerometer/mag-
netometer, temperature/pressure sensor and a microphone.
All on-board sensors can be put into a sleep state when not
used, resulting in a very small baseline power consump-
tion [15]. Camazotz nodes feature a flash storage chip for lo-
cal data buffering. The short-range radio transceiver is used
to offload data to gateway nodes installed at animal congre-
gation areas (flying fox roosting camps) [29].
5.2 Empirical Mobility Traces
Our empirical data set is based on tracking data from 10
free-living flying foxes that were tagged with a Camazotz
tracking device. We configured two separate GPS tasks on
the tracking device to collect GPS samples to serve as ground
1Ethical approval for experiments with animal subjects has been granted.
Animal identifier 1 2 3 4 5 6 7 8 9 10
Number of trips 3 7 7 6 7 8 3 8 5 5
Motion duration/night [minutes]: Minimum: 70.6 20.0 23.7 4.0 4.3 10.2 9.7 8.7 3.9 9.9
Average: 82.6 36.5 45.1 7.7 13.5 21.6 23.0 15.0 31.1 13.9
Maximum: 97.3 81.7 80.9 14.9 18.1 35.3 46.4 22.8 91.4 17.1
Average prediction error (individual) [minutes] 14.7 22.9 26.0 3.1 4.6 7.9 23.4 4.4 30.2 2.4
Average prediction error (population) [minutes] 57.4 15.5 20.5 18.8 12.9 7.4 17.0 11.4 22.3 12.5
Table 1: Actual and predicted trip time statistics for 10 animals and 59 trips in the flying foxes dataset where GPS position
samples with high-temporal resolution from motion-triggered continuous tracking are available.
truth location data for our retrospective analysis of differ-
ent scheduling mechanisms. During night time, sampling of
the GPS receiver is triggered using a threshold of +/-2 g on
the accelerometer and we acquire continuous GPS samples
at a rate of 1 Hz until the speed as reported by the GPS be-
comes slower than 5 m/s. After that, the GPS is put into sleep
mode until the next motion triggering event. As a fallback
in case the battery SOC falls below roughly 10%, we only
acquire a single GPS sample once every hour (static time-
based GPS scheduling) regardless of any motion triggering.
Position fixes are considered valid if the position accuracy re-
ported by the GPS is below 50 meters. GPS samples are then
stored into the external flash memory and transferred using
the short-range wireless radio when within proximity of the
base station located in bat roosting camps. During the obser-
vation period of several weeks, we select a dataset consisting
of 59 trips containing more than 100,000 GPS data points
with high temporal resolution, which allows reconstruction
of the corresponding trajectories with only minor gaps (e.g.,
due to hotstarts or temporary signal degradation). GPS head-
ing information is used to emulate magnetometer heading
data, as collecting raw samples from the inertial sensors is
not feasible due to storage constraints on the mobile nodes.
We further employ the empirical traces to learn the param-
eters used by the information-based tracking algorithm (see
Equation 6) by using data from all other trajectories (leave-
one-out).
5.3 Energy Consumption and Harvesting
We calculate the energy consumption Eused and the har-
vested solar energy Eharvested for each 24-hour period start-
ing at noon, as shown in Figure 6. Other than the energy
consumption of the GPS receiver, we also account for the
baseline consumption of the microcontroller, inertial sen-
sors, flash storage chip, and radio transceiver during idle lis-
tening, contact logging and data offload to the base station.
We report an average energy consumption of 33.0 mAh and
an average harvested energy of 27.3mAh daily.
It is important to note that continuous GPS tracking at
1 Hz during motion consumes more energy than what we can
harvest from solar panels on an average day. Therefore, such
an operation mode is not sustainable over a longer period of
time and can only be executed for a few consecutive days.
Once the battery SOC becomes too low for operating the
GPS in continuous tracking mode anymore, we switch to a
single sample every hour during the night as we always need
to guarantee a minimum energy budget for radio beacons.
0 10 20 30 40 50 60 70
Energy harvested/day [mAh]
0
10
20
30
40
50
60
70
Energy used/day [mAh]
Figure 6: Daily energy consumption and harvested solar en-
ergy for the Camazotz nodes deployed on wild flying foxes.
In this case, a configurable hysteresis hSOC on the minimum
SOC value effectively disables the motion-triggered contin-
uous GPS sampling task until the battery has recovered to a
reasonable charge state.
5.4 Performance of GPS Tracking
GPS receivers need to lock onto the signals from several
satellites to obtain an accurate position estimate, a so called
GPS fix. Thereby, the time to the first fix (TTFF) after en-
abling the GPS, highly depends on the internal state of the
receiver logic. If the receiver has no up-to-date satellite in-
formation and the current time, it needs to acquire this infor-
mation first by listening to the data transmitted by the GPS
satellites, which can take up to a minute or longer (coldstart).
Subsequent starts of the GPS can then rely on satellite infor-
mation locally stored in the receiver’s memory and provide
a first fix much faster (hotstart). During nightly foraging,
GPS hotstarts are dominating (96 %) with an average TTFF
of 5.9 seconds. We use this value to estimate the energy con-
sumption associated with a GPS hotstart in our simulation
framework (see Section 6).
5.5 Animal Motion Characteristics
We present a summary of motion statistics based on the
high-resolution GPS dataset collected with ten tagged ani-
mals in Table 1. It can be observed that daily motion dura-
tion can vary greatly for individual animals and for the whole
population. For each animal, we evaluate performance of
trip duration prediction based on individual and population
statistics. Specifically, we calculate the prediction error of
a trip as the difference between the trip duration and the
mean duration of all other trips in the dataset (constrained
Max. distance (yesterday) [m]
Max. distance (today) [m]
Max. distance from camp [m]
Total flight distance [m]
Figure 7: Maximum distance from the camp for two consec-
utive days (left), and maximum distance from the base camp
and the total flight time (right).
to the trips of the same animal for individual-based predic-
tion). Population-based estimation performs better for newly
tagged animals for which we do not have many historical
traces. However, individual estimates start working better if
we have a large enough sample set.
Due to the relatively small sample size of the high reso-
lution dataset, we also consider time periods where we do
not have continuous high-resolution GPS trajectories avail-
able, but where missing parts of the trajectory can be inter-
polated based on hourly GPS positions. We then calculate
the maximal distance from the camp and the total flight dis-
tance within that day (see Figure 7) for a total of 314 trips of
10 nodes. We can observe that today’s and yesterday’s max-
imum distance from the camp are weakly correlated (Pear-
son correlation = 0.35), confirming that yesterday’s distance
alone is not sufficient for predicting today’s distance. Fur-
thermore, the maximum distance from the camp is also re-
lated to the total flight distance on that day (Pearson corre-
lation 0.86). We use this to build a model for the total flight
distance dtotal(dcamp (t)) based on the current distance from
camp dcamp(t)(see Algorithm 1).
6 Performance Evaluation
We use empirical GPS traces from 10 free living flying
foxes to evaluate the performance of different tracking strate-
gies under realistic energy constraints. We select three dif-
ferent power budgets for the operation of the GPS receiver
and the inertial sensors. We evaluate our framework in three
modes: static motion-based,adaptive motion-based, and
information-based. In the static motion-based mode, a fixed
duty cycle is selected at the beginning of the night to drive
GPS sampling during motion periods that are detected by
the inertial sensor. In the adaptive motion-based mode, our
framework periodically adjusts the GPS sampling frequency
during motion based on its most recent estimates of energy
resources and remaining flight time. Information-based sam-
pling uses an adaptive error threshold for GPS sampling de-
cisions based on remaining flight time and on energy. We
also compare the performance of our framework to an opti-
mal offline schedule calculated with full access to the trajec-
tory.
Metrics. We focus on two metrics when evaluating tracking
strategies, namely the tracking error and the power consump-
tion. Given a sampling strategy, the error metric is used to
measure the accuracy of the obtained trajectories, and the
power consumption characterizes its energy efficiency. The
tracking error at the discrete time instant tiis defined as the
Euclidean distance between the ground truth location (x,y)
and our estimated location (ˆx,ˆy):
e[ti] = q(x[ti]−ˆx[ti])2+ (y[ti]−ˆy[ti])2.(7)
We define the average tracking error eavg and the worst-case
tracking error emax as follows:
eavg =1
T
T
∑
i=0
e[ti]emax =max
0<i≤Te[ti](8)
6.1 Comparison of GPS Tracking Strategies
We first provide an example GPS trajectory with a de-
tailed analysis of energy usage and motion prediction for
three different sampling strategies. Figure 8 shows an exam-
ple flight from our empirical dataset and demonstrates the ex-
ecution of the static, adaptive, and information-based track-
ing strategies. All three maps show the original trajectory for
reference. Clearly, the adaptive and information-based ap-
proaches capture the original trajectory well. In contrast, the
static approach underestimates the flight time as this animal
is flying longer than the population-based estimate would
suggest, so the node depletes its energy resources early dur-
ing the night and suffers high position errors from that point
on. The information-based approach captures the tortuous
sections of the trajectory better than the adaptive approach,
while both perform similarly for relatively straight sections
of the trajectory.
Figure 9 shows the energy consumption and predicted
motion duration for the three approaches. In the static ap-
proach, the node starts with the population-based estimation
of flight time and calculates a fixed GPS duty cycle accord-
ingly. It then quickly reaches its energy budget just before
midnight and stops sampling after that point. The remaining
flight trajectory back to the roosting camp is interpolated,
as shown by the dashed line in Figure 8. For the adaptive
and information-based strategies, the node starts out with
an individual-based historical estimate of movement dura-
tion, and as the time progresses, the weight of this esti-
mate progressively decreases, while the weight of the cur-
rent distance from base increases. By the end of the night,
the node’s travel time estimate converges to the actual flight
time. Information-based tracking continuously adapts the er-
ror threshold for sampling, and thereby the sampling sched-
ule, according to updated estimates of remaining flight time
and energy budget. As a result, the sampled trajectory clearly
shows non-uniform sampling patterns, where long straight
segments are sampled infrequently and more tortuous seg-
ments are sampled with higher frequency. Information-based
sampling captures the trajectory well while preserving more
than 10% of the energy budget.
These three example traces demonstrate the benefits of
adaptive motion-based and information-based scheduling
strategies which take into account an updated estimate for the
remaining flight time and available energy budget for tuning
Groundtruth
GPS Samples
(a) static motion-based tracking, population-based
prediction of movement duration
Groundtruth
GPS Samples
(b) adaptive motion-based tracking, individual-
based prediction of movement duration
Groundtruth
GPS Samples
(c) information-based tracking, individual-based
prediction of movement duration
Figure 8: GPS sampling points for static motion-based (a), adaptive motion-based (b), and information-based tracking (c) on an
example GPS trajectory of a flying fox. The unit length of the coordinate system is one kilometre. Map source: mapbox.com
0
10
20
30
40
50
60
70
80
Energy [J]
Static motion-based
Adaptive motion-based
Information-based
17 18 19 20 21 22 23 00 01 02 03 04 05 06 07
Time of Day [Hour]
0
50
100
150
Time [min]
Individual-based prediction
Predicted time to base Population-based prediction
Observed motion duration
Figure 9: Energy consumption (top) and motion duration
(bottom) for an example GPS trajectory of a flying fox.
tracking policy online. It is further apparent that there exists
a tradeoff for both the adaptive and information-based strate-
gies between the weight of the initial population-based pre-
diction of flight time and the online motion prediction mech-
anism proposed in Section 6 for the flying foxes scenario.
6.2 Tracking Performance
We further study the tracking performance on the fly-
ing foxes dataset for all strategies given a fixed available
energy budget and compare them to an optimal offline al-
gorithm. The distributions of the maximum tracking er-
rors for the different tracking algorithms, energy budgets
and individual-/population-based motion prediction modes
are shown in Figure 10. We can observe that adaptive ap-
proaches clearly outperform static scheduling in terms of
the maximum tracking error. Interestingly, the static ap-
proach performs even worse with increasing energy budgets.
This is due to the fact, that the initially calculated duty cy-
cle is higher and the node will take samples more aggres-
sively, which will cause the energy budget to be exceeded
faster when the flight time is longer than predicted, caus-
ing high errors towards the end of the night. The benefits
of adaptive scheduling become apparent when the motion
period exceeds the average motion duration for the whole
population, which is around 30 minutes. We observe that
adaptive scheduling based on the previous night’s motion
statistics (adaptive-prev-population) only performs slightly
worse than an individual-based estimate on multiple trajecto-
ries (adaptive-individual). Information-based sampling with
individual-based prediction delivers consistently better ac-
curacy than both adaptive and static approaches for longer
movement durations of 60 minutes or more, while showing
comparable accuracy performance for shorter trips. The key
differentiator for information-based sampling compared to
adaptive sampling is that it tailors the sampling frequency
to the estimated information gain from taking the next sam-
ple, while considering the forecast movement duration and
energy budget. In contrast, adaptive sampling uses periodic
sampling of GPS during movement, while adapting the sam-
pling period regularly as it updates its movement forecast.
The results for our dataset of 59 empirical 14-hour traces
from 10 wild flying foxes are summarized in Figure 11.
While adaptive population-based sampling significantly im-
proves accuracy over static sampling, it is slightly outper-
formed by adaptive individual-based sampling that consid-
ers individual histories in forecasting movement duration.
Information-based sampling performs best of all online ap-
proaches with an accuracy improvement of 96%, 75% and
64% over the static, population-based, and individual sam-
pling strategies respectively for a power budget of 1.5 mW. It
delivers these improvements by capturing detailed trajectory
features through more frequent sampling, and by reducing
sampling rates when the trajectory can be well interpolated
between samples.
We also note room for improvement given the tracking
error of the optimal offline schedule through sophisticated
models for flight destination prediction. As we collect more
flight traces from wild flying foxes, this data will improve
our prediction model, for example by matching the current
trip to a set of previously visited locations or frequently trav-
eled waypoints. This would allow a more accurate estimation
of the flight destination and the remaining motion duration.
7 Related Work
Localisation of mobile devices have been studied ex-
tensively in the past motivated by applications in naviga-
tion, tracking and rescue. Location awareness is also a
key requirement for emerging mobile applications such as
location-based search, turn-by turn navigation, social media,
or participatory sensing [10, 9, 24].
Wildlife tracking. GPS technology was first utilized on
larger animals, but with the maturing of the technology the
size of the units, and therefore the size of the species it can
be used on is decreasing. However, certain wildlife tracking
applications cannot rely on GPS due to the application sce-
nario, e.g., tracking badgers underground [8], and thus em-
ploy alternative localization techniques such as RFID tags.
Lindgren et al. [20] equipped seals with short-range radio
0 20 40 60 80 100
Movement duration [minutes]
Tracking error (max) [m]
static-population
adaptive-prev-population adaptive-individual
information-individual offline
(a) Budget: 1.5 mW
0 20 40 60 80 100
Movement duration [minutes]
Tracking error (max) [m]
static-population
adaptive-prev-population adaptive-individual
information-individual offline
(b) Budget: 2.25 mW
0 20 40 60 80 100
Movement duration [minutes]
Tracking error (max) [m]
static-population
adaptive-prev-population adaptive-individual
information-individual offline
(c) Budget: 3 mW
Figure 10: Distribution of the resulting maximum tracking error on the flying foxes dataset for different tracking strategies
given an energy budget corresponding to an average power consumption of 1.5, 2.25 and 3mW during the observation period.
1.5 2.25 3
Power budget [mW]
Tracking error (avg) [m]
static-population
adaptive-prev-population adaptive-indvidual
information-individual offline
Figure 11: Average tracking errors for different strategies
given a specific energy budget on the flying foxes dataset.
transceivers to collect contact logs between tagged animals
instead of using GPS to determine the location of the ani-
mal. With our focus on tracking in remote areas with limited
cellular network coverage, we rely on GPS only to deter-
mine the current position of tagged animals. Although the
power consumption of GPS receivers has been continually
improving, it remains a major part of the overall energy bud-
get of mobile devices. Liu et al. [21] proposed to reduce the
power consumption by offloading the GPS signal processing
to the cloud. However, this requires to store and transfer a
large amount of raw data, which creates another bottleneck
for resource-constrained applications.
GPS Duty-Cycling. Continuous tracking applications ex-
hibit a trade-off between accuracy of localization and the
amount of energy spent to do so. Therefore, mechanisms
to adapt the duty cycle of the GPS receiver and other lo-
calization modalities have been extensively studied in pre-
vious work [32, 9]. Paek et al. [25] use the location-time
history of a user and a blacklist based on cell tower signal
strength to decide when to activate the GPS and to avoid un-
availability of GPS indoors. SensLoc [16] uses algorithms
to detect the user context, such as frequently visited places,
and movement detection to adapt the GPS duty cycle. Ju-
rdak et al. [14] propose to use short-range radio contact-
logging to bound the position uncertainty when duty-cycling
the GPS. SmartDC [5] provides adaptive GPS duty-cycling
using prediction mechanisms for regularities in user mobil-
ity. eNav [12] employs dead-reckoning using the accelera-
tion in smartphones to enable the GPS only when the user
reaches the next navigation waypoint.
Energy Management. Adaptive localization algorithms,
such as EnLoc [6], EnTracked [17], and a-Loc [19], dy-
namically select the localization modality that provides the
most energy-efficient position estimate within provided un-
certainty bounds. Another approach to duty-cycling the GPS
is to rely on low-power sensors such as accelerometer or
compass to trigger GPS sampling only when the device is
moving [16, 25] or when it changed its heading [17]. Jig-
saw [22] manages battery life of mobile phones by balanc-
ing power consumption of sensing processes with quality re-
quirements of the application and user context.
Unlike most adaptive localization algorithms, we focus on
the inverse problem of determining the most accurate GPS
sampling strategy given energy constraints. Consequently,
instead of sampling GPS location when the uncertainty of the
current location estimate increases above a certain limit, we
constrain the GPS sampling subject to the available energy.
Our approach will not strictly enforce that all location esti-
mates are within the error bounds, but will strive to achieve
best performance overall given the energy constraints.
Energy-aware process scheduling frameworks have been
proposed to adjust the long-term system performance based
on prediction of harvested energy [3]. While such frame-
works consider long-term operation of solar energy harvest-
ing systems over years at a guaranteed minimum perfor-
mance level, the goal of this paper is to manage short-term
energy usage under daily varying harvesting and mobility.
8 Discussion
This paper has presented information-based tracking that
maximizes the tracking information return subject to an en-
ergy budget. Our strategy uses historical motion statistics
and distance from the destination to forecast activity du-
ration to strategically select the best sampling GPS policy.
The algorithm then updates its forecast and adapts its ac-
ceptable error threshold in response to observed changes in
distance from base and energy availability. We evaluated
our approach on empirical traces from wild flying foxes and
showed that our approach can significantly increase position-
ing accuracy for a given energy budget.
While our work was motivated by tracking flying foxes,
our proposed energy- and mobility-aware scheduling frame-
work is applicable to a broad range of other tracking appli-
cations. For instance, most living beings, including humans,
exhibit very strong preferential return behaviour to one or
two locations, typically home and work or study [11]. There
are also strong time-of-day correlations with the most likely
destination (going to work at 8 am or returning home at 5
pm [1]). As such, the concepts for scheduling of GPS by
predicting motion duration on the basis of previous days,
population-based statistics, or current distance from the most
likely destination are clearly transferable to tracking people
and indeed many other living species [27]. A related area
is asset tracking in logistics. Unpowered ground assets of-
ten need to be tracked over long durations. Using historic
motion patterns, whether for the asset group or individual
assets, and energy availability to schedule position sampling
can provide more accurate asset tracking.
An interesting direction for future work is to improve
motion duration prediction algorithms for more accurate
scheduling of GPS samples. Interestingly, information-
based sampling can deliver better accuracy while preserv-
ing budgeted energy. This is likely a result of training the
tuning parameters over a large number of trajectories, which
can lead to more conservative sampling for some trajecto-
ries. We leave a deeper investigation of this aspect for future
work. For flying fox tracking, it will be important to de-
tect whether the animal is flying to a new location, or is on
a typical foraging trip returning to the same roosting camp.
In addition, temperature and solar exposure inputs can be
used as proxies for weather conditions that may better pre-
dict the extent of movement and energy demand for a sin-
gle forecast period. Finally, it will be important to evaluate
the information-based tracking strategy for other application
scenarios, such as for people or cattle. We expect our work
to be an important step towards highly autonomous tracking
with energy harvesting that delivers detailed trajectory data
near-perpetually for a broad range of applications.
9 Acknowledgments
This work was supported by the Batmon Project in
CSIRO’s Sensor and Sensor Networks Transformation Ca-
pability Platform.
10 References
[1] A. Bamis and A. Savvides. Lightweight Extraction of Frequent
Spatio-Temporal Activities from GPS Traces. In Proc. RTSS, 2010.
[2] B. Buchli, D. Aschwanden, and J. Beutel. Battery State-of-Charge
Approximation for Energy Harvesting Embedded Systems. In Proc.
EWSN, 2013.
[3] B. Buchli, F. Sutton, J. Beutel, and L. Thiele. Dynamic Power Man-
agement for Long-term Energy Neutral Operation of Solar Energy
Harvesting Systems. In Proc. SenSys, 2014.
[4] W. S. Chan and F. Chin. Approximation of polygonal curves with min-
imum number of line segments or minimum error. International Jour-
nal of Computational Geometry & Applications, 6(01):59–77, 1996.
[5] Y. Chon, E. Talipov, H. Shin, and H. Cha. Smartdc: Mobility
prediction-based adaptive duty cycling for everyday location moni-
toring. IEEE Trans. Mobile Comput., 13(3):512–525, 2014.
[6] I. Constandache, S. Gaonkar, M. Sayler, R. R. Choudhury, and L. Cox.
EnLoc: Energy-Efficient Localization for Mobile Phones. In Proc.
INFOCOM, 2009.
[7] P. Dutta, M. Feldmeier, J. Paradiso, and D. Culler. Energy Metering
for Free: Augmenting Switching Regulators for Real-Time Monitor-
ing. In Proc. IPSN, 2008.
[8] V. Dyo, S. A. Ellwood, D. W. Macdonald, A. Markham, C. Mascolo,
B. Pásztor, S. Scellato, N. Trigoni, R. Wohlers, and K. Yousef. Evolu-
tion and Sustainability of a Wildlife Monitoring Sensor Network. In
Proc. SenSys, 2010.
[9] S. B. Eisenman, E. Miluzzo, N. D. Lane, R. A. Peterson, G.-S. Ahn,
and A. T. Campbell. Bikenet: A mobile sensing system for cyclist
experience mapping. ACM Trans. Sen. Netw., 6(1):6:1–6:39, 2010.
[10] J. Eriksson, L. Girod, B. Hull, R. Newton, S. Madden, and H. Balakr-
ishnan. The Pothole Patrol: Using a Mobile Sensor Network for Road
Surface Monitoring. In Proc. MobiSys, 2008.
[11] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi. Understanding
individual human mobility patterns. Nature, 453(7196), 2008.
[12] S. Hu, L. Su, S. Li, S. Wang, C. Pan, S. Gu, M. T. Al Amin, H. Liu,
S. Nath, R. R. Choudhury, and T. F. Abdelzaher. Experiences with
eNav: A Low-power Vehicular Navigation System. In Proc. Ubi-
Comp, pages 433–444, 2015.
[13] X. Jiang, P. Dutta, D. Culler, and I. Stoica. Micro Power Meter for
Energy Monitoring of Wireless Sensor Networks at Scale. In Proc.
IPSN, 2007.
[14] R. Jurdak, P. Corke, D. Dharman, and G. Salagnac. Adaptive GPS
Duty Cycling and Radio Ranging for Energy-Efficient Localization.
In Proc. SenSys, 2010.
[15] R. Jurdak, P. Sommer, B. Kusy, N. Kottege, C. Crossman, A. Mck-
eown, and D. Westcott. Camazotz: Multimodal Activity-based GPS
Sampling. In Proc. IPSN, 2013.
[16] D. H. Kim, Y. Kim, D. Estrin, and M. B. Srivastava. SensLoc: Sensing
Everyday Places and Paths using Less Energy. In Proc. SenSys, 2010.
[17] M. B. Kjaergaard, S. Bhattacharya, H. Blunck, and P. Nurmi. Energy-
efficient trajectory tracking for mobile devices. In Proc. MobiSys,
2011.
[18] M. B. Kjaergaard, J. Langdal, T. Godsk, and T. Toftkjaer. EnTracked:
Energy-Efficient Robust Position Tracking for Mobile Devices. In
Proc. MobiSys, 2009.
[19] K. Lin, A. Kansal, D. Lymberopoulos, and F. Zhao. Energy-Accuracy
Trade-Off for Continuous Mobile Device Location. In Proc. MobiSys,
2010.
[20] A. Lindgren, C. Mascolo, M. Lonergan, and B. McConnell. Seal-
2-Seal: A Delay-Tolerant Protocol for Contact Logging in Wildlife
Monitoring Sensor Networks. In Proc. MASS, 2008.
[21] J. Liu, B. Priyantha, T. Hart, H. S. Ramos, A. A. F. Loureiro, and
Q. Wang. Energy Efficient GPS Sensing with Cloud Offloading. In
Proc. SenSys, 2012.
[22] H. Lu, J. Yang, Z. Liu, N. D. Lane, T. Choudhury, and A. T. Camp-
bell. The Jigsaw Continuous Sensing Engine for Mobile Phone Ap-
plications. In Proc. SenSys, 2010.
[23] A. McKeown and D. Westcott. Assessing the accuracy of small satel-
lite transmitters on free-living flying-foxes. Australian Ecology, 37,
2012.
[24] P. Mohan, V. N. Padmanabhan, and R. Ramjee. Nericell: Rich Moni-
toring of Road and Traffic Conditions using Mobile Smartphones. In
Proc. SenSys, 2008.
[25] J. Paek, J. Kim, and R. Govindan. Energy-Efficient Rate-Adaptive
GPS-based Positioning for Smartphones. In Proc. MobiSys, 2010.
[26] J. Ryder, B. Longstaff, S. Reddy, and D. Estrin. Ambulation: A Tool
for Monitoring Mobility Patterns over Time Using Mobile Phones. In
Proc. CSE, 2009.
[27] S. Scellato, M. Musolesi, C. Mascolo, V. Latora, and A. Campbell.
NextPlace: A Spatio-temporal Prediction Framework for Pervasive
Systems. In Proc. Pervasive. 2011.
[28] P. Sommer, B. Kusy, and R. Jurdak. Energy Estimation for Long-Term
Tracking Applications. In Proc. ENSSys, 2013.
[29] P. Sommer, B. Kusy, A. McKeown, and R. Jurdak. The Big Night
Out: Experiences from Tracking Flying Foxes with Delay-Tolerant
Wireless Networking. In Proc. REALWSN, 2013.
[30] C. Song, Z. Qu, N. Blumm, and A.-L. Barabási. Limits of predictabil-
ity in human mobility. Science, 327(5968):1018–1021, 2010.
[31] D. Westcott et al. The spectacled flying-fox, pteropus conspicillatus.
Technical report, June 2001.
[32] P. Zhang, C. M. Sadler, S. A. Lyon, and M. Martonosi. Hardware
Design Experiences in ZebraNet. In Proc. SenSys, 2004.
