Camazotz: multimodal activity-based GPS sampling

Abstract

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, but they only provide coarse-grained control over GPS activity. In this paper, we introduce our feature-rich lightweight Camazotz platform as an enabler of Multimodal Activity-based Localisation~(MAL), which detects activities of interest by combining multiple sensor streams for fine-grained control of GPS sampling times. Using the case study of long-term flying fox tracking, we characterise the tracking, connectivity, energy, and activity recognition performance of our module under both static and 3-D mobile scenarios. We use Camazotz to collect empirical flying fox data and illustrate the utility of individual and composite sensor modalities in classifying activity. We evaluate MAL for flying foxes through simulations based on retrospective empirical data. The results show that multimodal activity-based localisation reduces the power consumption over periodic GPS and single sensor-triggered GPS by up to 77% and 14% respectively, and provides a richer event type dissociation for fine-grained control of GPS sampling.

Full text

Camazotz: Multimodal Activity-Based GPS Sampling
Raja Jurdak1, Philipp Sommer1, Branislav Kusy1, Navinda Kottege1,
Christopher Crossman1, Adam McKeown2, David Westcott2
1Autonomous Systems Lab, CSIRO ICT Centre, Brisbane, QLD, Australia
2CSIRO Ecosystem Sciences, Cairns, QLD, Australia
firstname.lastname@csiro.au
ABSTRACT
Long-term outdoor localisation with battery-powered de-
vices 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, but
they only provide coarse-grained control over GPS activity.
In this paper, we introduce our feature-rich lightweight Ca-
mazotz platform as an enabler of Multimodal Activity-based
Localisation (MAL), which detects activities of interest by
combining multiple sensor streams for fine-grained control of
GPS sampling times. Using the case study of long-term fly-
ing fox tracking, we characterise the tracking, connectivity,
energy, and activity recognition performance of our module
under both static and 3-D mobile scenarios. We use Cama-
zotz to collect empirical flying fox data and illustrate the
utility of individual and composite sensor modalities in clas-
sifying activity. We evaluate MAL for flying foxes through
simulations based on retrospective empirical data. The re-
sults show that multimodal activity-based localisation re-
duces the power consumption over periodic GPS and single
sensor-triggered GPS by up to 77% and 14% respectively,
and provides a richer event type dissociation for fine-grained
control of GPS sampling.
Categories and Subject Descriptors
C.2.1 [Computer-Communication Networks]: Wireless
Communication
Keywords
Wireless Sensor Networks, Tracking
1. INTRODUCTION
Embedded systems technology has been developing at re-
markably fast rates, which has led to heightened expecta-
tions for a wide range of applications. End-users now ex-
pect platforms to continuously follow intuitive trends, such
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.
IPSN’13, April 8–11, 2013, Philadelphia, Pennsylvania, USA.
Copyright 2013 ACM 978-1-4503-1959-1/13/04 ...$15.00.
as shrinking in size and weight while having longer battery
lives. In order to deliver on all those expectations concur-
rently, system developers typically reduce the number of
sensing modalities on monitoring platforms and the sensor
sampling frequencies. For many mobile sensing applications,
including bird tracking at continental scales, the require-
ment for multiple sensing modalities and durable lightweight
platforms are continuously in tension.
Recent work [1] has aimed for lightweight long-distance
and long-term tracking of the endangered Whooping Crane
in North America, using custom-designed platforms that
include GPS, cellular, and weigh just above 100g. While
this work has provided a proof-of-concept of large scale and
lightweight wildlife tracking, this technology cannot be used
for tracking smaller flying animals, as the sheer weight of the
devices would prevent the animals from flying freely. This
paper is motivated by the need to track one such species of
particular interest, namely flying foxes. Flying foxes, also
known as fruit bats, are megabats that spread virulent and
deadly diseases such as Ebola, Hendra, and the recently dis-
covered SARS-like Coronavirus [26], at a global scale.
Tracking flying foxes requires platforms and algorithms
that can deliver position and activity information from highly
mobile individual animals over long-durations. While posi-
tion monitoring can use GPS as the main sensor modal-
ity, behaviour and activity classification require additional
sensor modalities, such as inertial, acoustic, and air pres-
sure sensors. For instance, audio signals can be used for
detecting previously unknown congregation areas, or roost-
ing camps, for flying fox populations. The introduction of
new sensing modalities places a burden on the limited node
energy, processing, and memory resources, as well as an in-
direct cost of more complex system management. The un-
controlled 3-D mobility associated with bird tracking can
also have unpredictable effects on the performance of node
components, particularly transceivers and solar panels. All
of the above challenges highlight the need for holistic design
of feature-rich mobile sensing platforms with early prototyp-
ing to incorporate these subtle dependencies among system
components.
This paper introduces Camazotz1, a lightweight and feature-
rich mobile sensing platform, which aims at long-term wildlife
tracking. Camazotz uses a CC430 system on chip (SoC) with
a low power GPS, inertial, acoustic, air pressure and temper-
ature sensors, two solar panels, 300 mAh Li-Ion battery, with
a total weight just under 30g targeted at tracking smaller
wildlife such as flying foxes. We describe our holistic design
1Camazotz was a bat god in Mayan mythology.

process for the platform that relies on early prototyping and
empirical evaluation of three key aspects: (1) the impact
of 3-D mobility on radio performance using an unmanned
aerial vehicle for controlled mobility experiments; (2) the
performance of low power GPS as a function of shutoff time
and its implications on node lifetime; and (3) the ability to
perform in situ activity recognition using audio and inertial
sensors through basic signal processing on Camazotz. Based
on the evaluation results, we show how Camazotz can enable
Multimodal Activity-based Localisation (MAL) that detects
activities of interest by combining multiple sensor streams
for fine-grained control of GPS sampling times.
Our evaluation shows that 3-D mobility has limited effect
on Camazotz’s radio connectivity to a ground base station.
We mainly find signal degradation for high angles of align-
ment between Camazotz and the base. Our GPS evaluations
confirm a weak dependence of the time to first fix on GPS off
time, and our on-bat experiments show that the GPS design
of Camaztoz achieves consistent position accuracies below
10 m. Solar experiments from nodes on bats yield an esti-
mate of 3 mA average solar current during the day, which we
use to set duty cycles that deliver energy-neutral operation
to Camaztoz. Our activity recognition results from flying fox
experiments with Camazotz demonstrate the detection of in-
teraction and waste removal events with audio and inertial
sensors respectively, and confirm that air pressure sensors
can provide a much more precise estimate of altitude than
GPS. Finally, we use these results to demonstrate MAL by
considering how simple fusion of audio and inertial sensor
events through logical OR and AND operations can disso-
ciate event types, deliver fine-grained activity-based control
of GPS samples, and by doing so, save power consumption.
The remainder of the paper is organised as follows. Sec-
tion 2 motivates the design of Camaztoz. Section 3 presents
our empirical validation experiments to evaluate the plat-
form’s performance. Section 4 shows how Camazotz can
enable multimodal activity-based localisation to accurately
detect events and extend node lifetime. Section 5 discusses
related work, and Section 6 concludes the paper.
2. CAMAZOTZ PLATFORM
2.1 Motivating Application
Surprisingly little is known about flying-fox ecology be-
haviour due to difficulties associated with studying animals
that are nocturnally active and which roost in large aggre-
gations (often 40-50000 animals at a single site [23]). Recent
research shows that the source of our difficulties in study-
ing these animals lies in the extra-ordinary mobility exhib-
ited by individuals and by flying-fox populations. Studies
show that individual animals are highly mobile, travelling
on average 20km to their first feeding site in a night and
over 100km during nightly foraging [25]. Over weeks and
months individuals can move hundreds or thousands of kilo-
metres [4]. This mobility is also observable at the scale of
the population with flying-fox populations moving in and
out of regions, often over periods of just days [23].
Flying-foxes are of great interest to wildlife managers in
Australia. On the one hand these animals are listed as
threatened species, and at the same time, they are recog-
nised as agricultural pests, causing as much as $20 mil-
lion of damage to fruit crops per year [24]. Most impor-
tantly, flying-foxes are effective vectors of a number of viru-
Texas
Instruments
CC430F5137
Ublox
Max 6
Bosch
BMP085
ST Micro
LSM303
Audio Mic
Atmel
AT25DF
I2C
SPI
ADC
Power
Supplies
Solar
Panels
Li-Ion
Charger
Li-Ion
Battery
Serial
Flash
Low Power
GPS
CC430F5137
System-on-
Chip:
MCU/Radio
Pressure
sensor
3-D Inertial
sensors
Microphone
(a) Functional components of
Camazotz (b) Flying fox with
Camazotz attached
Figure 1: The Camazotz platform couples a SoC
with GPS, inertial, acoustic, air pressure and tem-
perature sensors, two solar panels, 300 mAh Li-Ion
battery, with a combined weight <30 g.
lent emerging infectious diseases that threaten both humans
and livestock. Developing effective management responses
to these flying-fox impacts requires that assumptions are
made about their mobility.
Our application requirements are to obtain day roost loca-
tions for comparison with surveyed camp locations. Where
these locations do not match with known camps, we need
to know whether the animal is roosting alone or in a small
group, or, whether it is at an unknown camp. This requires
visiting each such location – an impossible task given that
such events could happen multiple times each day across the
range of the species (i.e. 2800 km) and will often be in in-
accessible locations. Since flying-foxes in camps are highly
vocal animals [19] an alternative would be to use collar node-
based microphones to make recordings of flying-fox noise to
provide an index of the number of animals at a roost.
Most flying-fox risks, such as crop damage and transmis-
sion of disease, are incurred away from camps at the lo-
cations where flying-foxes are feeding. Predictive models
of how these risks will be distributed within a landscape
require an understanding of how flying-foxes respond, in
terms of their movement and choice of foraging locations, to
the structure of landscapes and the distribution of resources
within them, and how this varies across landscapes, seasons
and individuals. Developing such an understanding requires
high temporal and spatial resolution data on movement dur-
ing nighttime foraging sessions, with sample frequencies as
high as 1 Hz potentially under some circumstances.
The devices that can achieve the goals outlined above
must be capable of: (1) collecting regular daytime fixes (with
an accuracy of 10 m) at camps to identify new camps; (2)
collecting high-frequency nighttime fixes to monitor move-
ment patterns and landscape use, and doing this with an
accuracy of 10 m or less using inertial sensors during fine-
scale movements; (3) making daytime audio recordings to
allow estimation of camp size; (4) operate over long periods,
i.e. 12 months, and preferably longer; and (5) provide data
download capability.
Such simultaneous goals are clearly at odds with current
technology but can be approached through smart power,

sensor and data management algorithms and flexible duty-
cycling. Available technologies, such as Platform Termi-
nal Transmitter (PTT) and GPS tags [18], cannot hope to
achieve these goals because their power demands conspire
to make the tags useful for only single aspects of the study
described above. That is, they can either collect a hand-
ful of daytime fixes at regular intervals over long-periods, or
they can collect high-frequency movement data over short
periods. No tags are currently capable of collecting audio
data. This paper addresses this gap for accurate, flexible
and energy-efficient position tracking of flying-foxes. This
work is part of a large national project that aims to deploy
hundreds of tracking nodes (up to 1000) on individual flying
foxes.
2.2 Design Challenges
bioacoustic signals alongside inertial and altitude infor-
mation for real-time activity classification.
To achieve the above goals, we need to address specific de-
sign challenges relating to dimension constraints of tracking
nodes and to the mobility dynamics of migratory birds. Ac-
cording to animal ethics regulations in Australia, the weight
of any objects placed on flying foxes must not exceed 5%
of their body weight, corresponding to target cumulative
weights of 30 to 50g for all the electronics and enclosures for
tracking. The range of target weights stems from the weight
differences of individual animals between adolescents and
larger males. This weight restriction obviously constrains
the size and capacity of batteries on the devices.
Size is another issue, where tracking devices cannot exceed
a few centimetres in height or length and 2cm in depth to
ensure that the devices do not hinder or affect the animal’s
ability to fly freely. In addition to placing further constraints
on the battery and electronics, this size restriction is likely
to affect the size of the recommended GPS antenna which
may in turn impact location accuracy.
The mobility dynamics of flying foxes represent yet an-
other major challenge. Flying foxes are able to fly up to
100 km in a single night and they are known to visit truly
remote areas at continental and transcontinental scale. Be-
cause they are likely to spend significant portions of time
either in remote areas or across country borders, cellular
coverage may not be available. Additionally, cellular mod-
ules would add significant weight, size, and energy cost to
the overall platform. We choose to transfer position data
by installing base stations at known roosting camps and us-
ing short-range radio communication opportunistically when
the animals are in these camps. However, only a small pro-
portion of the camps where these animals congregate to
roost is known, and there is no deterministic mapping of
an individual animal to one or more known roosting camps.
These dynamics suggest that the design of tracking devices
has to account for a high degree of delay tolerance in both
hardware, providing enough memory to store position and
activity data for periods of disconnectivity, and software, to
Device Size (mm) RAM (Bytes) Flash (KB)
TI CC430F5137 7×7 4096 32
Freescale MC12311 8×8 2048 32
Nordic nRF9E5 5×5 256 4
Atmel ATA8743 5×5 256 4
Table 1: Comparison of 900 Mhz SoC devices.
compress stored data and opportunistically deliver it once
connectivity returns.
2.3 Hardware
Given such tight constraints on the size and weight of the
platform, we select a system on chip (SoC) for the micro-
controller and radio transceiver. In particular, we compare
the size, RAM, and Flash capacity of existing SoC options,
as shown in Table 1. Based on this comparison, we use the
Texas Instruments CC430F5137 which includes an MSP430
core and a CC1101 radio. Apart from its favourable physi-
cal size/capacity advantage, the MSP430 core supports low
power operation and offers high compatibility with popular
sensor network operating systems. The CC1101 equivalent
radio transceiver provides a GFSK communication in the
915 MHz band.
Figure 1(a) illustrates the functional components of Ca-
mazotz. Key to its success in the field is to maximise lo-
cation accuracy for the available size, weight, and energy
resources. We adopt the u-blox MAX-6 GPS module, which
optimises for size and power consumption and provides the
high performance of the u-blox 6 series positioning engine.
The localisation hardware also includes the GPS antenna,
which involves a design choice between antenna size/type on
one hand, and the GPS signal directionality and strength on
the other. Table 2 shows two considered GPS antenna op-
tions: a Taoglas GPS patch antenna with integrated ampli-
fier, which while providing strong signal reception will also
have high directionality and take up a reasonable amount
of space and weight; and a Fractus small planar monopole
GPS antenna [9], which is lighter, less directional, yet pro-
vides weaker signal reception. The smaller antenna’s omni-
directional radiation pattern maintains consistent GPS re-
ception in any orientation, which is particularly favourable
for the 3-D mobility of flying foxes. We therefore select the
Fractus antenna and augment it with a 20dB low noise am-
plifier (LNA) to boost its signal to comparable levels as the
patch antenna, whilst consuming less power.
One disadvantage with our antenna choice is that it re-
quires an adjacent ground plane that is nearly 12 times its
size to work efficiently. We opportunistically match the size
of the ground plane to the overall Camazotz board footprint,
so that the same Camazotz board that includes all the func-
tional components has a dual role as the ground plane for
the GPS module. This design choice keeps the overall node’s
weight and size within their respective targets.
Figure 2 shows the top and bottom views of the Camaztoz
platform, while Figure 1(b) shows the node within its enclo-
sure on a bat during one of our field trials. Note the dual
solar panels in Figure 1(b) on opposite sides of the enclosure,
in order to maximise the chances of energy harvesting when
foxes are roosting in a camp (typically in the upside down
position, where the node slightly flops down), or flying at
the beginning or end of the day, which exposes the top-side
panel to the sun.
The energy charging architecture of Camaztoz is yet an-
other design consideration. Since nodes on flying foxes will
Type Size(mm) Gain (dB) Power@1.8 V
Taoglas AP.10F 10×10×4 -10 + 20 9
Fractus Geofind + LNA 10×10×0.9 1.5 + 20 5.61
Table 2: Comparison of small GPS antennas.

Figure 2: Top (left) and bottom (right) view of Ca-
mazotz prototype device without battery and solar
panel. Dimensions are 54×30×14 mm.
have unpredictable and intermittent access to solar energy,
we need a flexible energy architecture that opportunistically
exploits available solar energy. In particular, there will be
situations when node batteries are either fully charged or
fully flat. In both these cases, it is beneficial to power the
node directly through the solar panel to make the most of
the node’s sun exposure. We therefore incorporate a solar
bypass circuit to enable Camazotz to consume energy di-
rectly from the solar panels during extreme battery states.
The flat battery state occurs when the battery is completely
flat and is only trickle charging at a very low current. In
this case, the bypass allows the device to power up cleanly,
rather than relying on the flat battery to power up, which
would risk oscillation around a minimum voltage threshold
leading to data loss. During the fully charged state, the by-
pass circuit on Camazotz can use any excess solar energy
(that would otherwise be wasted) for increased sampling or
computation.
The design of Camazotz also adopts a low power approach
in its selection of sensors and in the integration of these sen-
sors into the board. We select low power sensors for Cama-
zotz to suit its restricted energy budget, with an eye towards
Multimodal Activity-based Localisation for fine-tuned GPS
sampling control. In particular, we select the Bosch BMP085
pressure sensor, the STMicroelectronics LSM303 3-axis ac-
celerometer/magnetometer and a Knowles microphone. The
BMP085 pressure sensor draws only 12 µA, and when com-
bined with a static node’s pressure reading, can provide us
with a more accurate height measurement (see Figure 11(a)).
The LSM303 accelerometer consumes 830 µA of current and
allows for detection of different behaviours (see Figure 10).
The final sensor is the Knowles microphone, which is con-
nected to the 12-bit ADC on the microcontroller, and con-
sumes less than 1 mA in operation. The microphone can
be used in conjunction with other sensors for more robust
activity detection.
Integrating these sensors into Camazotz requires a design
decision in itself. Having access to a limited energy supply
and a goal of long term operation dictates that we duty-cycle
the node components. This is a particular focus to ensure
that we could minimise the energy consumption in the sleep
state. Rather than putting all the peripheral components
into their standby modes, which are in the order of 40µA
total, we create a single digital line that can cut power to
all peripherals surrounding the SoC prior to entering sleep
state for the lowest possible energy consumption of 12µA
on average.
2.4 Software
The Camazotz platform runs the Contiki operating sys-
tem, which provides a threaded programming environment
using the C programming language. We add two key fea-
tures on top of the Contiki core: remote procedure calls, and
a logging abstraction.
As the Camazotz device will be deployed on wild ani-
mals, retrieval the node for reconfiguration is not an option.
To address this issue we implement remote procedure calls
(RPC), allowing us to send a radio command to the device to
perform certain actions (e.g. reading memory blocks or sta-
tus information) or to adjust configuration parameters such
as the GPS duty-cycle. Every RPC command is sent as a
unicast or broadcast packet containing a unique command
identifier and a list of arguments. Our implementation of
RPC commands serves as a basic building block to support
additional functionality for future use cases. For example,
a base station located at a roosting camp can query status
information of a mobile node to request sensor data stored
in the flash memory to be sent over the radio.
Logging on the Camazotz device is critical to its success,
due to the delay-tolerant nature of the flying fox applica-
tion. Communication outage times may range from hours,
days, weeks or even months before there is an opportunity
to offload data. Initially, data will be logged at a high sam-
ple rate to a Secure Digital (SD) flash card, for board and
code verification, and then switch to external flash for fi-
nal testing. To address this requirement for interchangeable
storage, we introduce a data logging abstraction, which pro-
vides a consistent application programming interface (API),
regardless of the underlying storage mechanism. The ad-
vantage of the logging abstraction approach is that we can
log high sample rate sensor data to the SD card while in
the development and testing phases, then for the final ver-
sion we reduce the sampling rate, required by our energy
budget, and with minimal code changes, switch to use the
external flash for logging. Current mechanisms supported
by the logging abstraction include the radio, external flash
and SD card.
3. EVALUATION
This section empirically evaluates the Camazotz platform’s
communication, energy, and sensing features.
3.1 Mobility
Flying foxes are active animals that can cover distances of
up to 100 km at cruising speeds of 7-8 ms-1. They participate
in complex social behaviour while at roosting camps that
frequently result in their location change. Our deployment
setup includes a base station with 3G connectivity that is
deployed within the roosting camp close to the ground. The
Camazotz platform needs to be able to communicate with
the base station from the surrounding trees, within a dis-
tance of 200 m, as well as enable bats-to-bat communication
outside of roosting camps. In this section, we study the im-
pact of the height and mobility of the Camazotz transceiver
on the received radio signal quality. We study three antenna
types to maximise packet reception rates at the base station.
3.1.1 Experimental Platform
We use AscTec Pelican UAV platform [22], which is a
flexible quad-copter platform designed for easy integration
with a variety of payloads, up to a maximum of 650g. The

100
80
60
40
20
0
-40
-50
-60
-70
-80
-90
-100
Packet Loss (%)
RSSI (dBm)
0 10 20 30 40 50 60 70 80
Range (m)
ChipS Loss
ChipL Loss
Whip Loss
ChipS RSSI
ChipL RSSI
Whip RSSI
Figure 3: Comparison of mobile to base RSSI and
packet loss for two chip antennas (large and small)
and a whip antenna.
platform is equipped with inertial and GPS chips and an
autopilot that enables non-experts to pilot the platform out
of the box after a few minutes of training. Our payload con-
sists of the Camazotz prototype with radio and GPS chips,
high-capacity Li-Ion battery to power the prototype, and a
mount for different test antennas. The maximum flight time
with our payload is about 30 mins. Altogether, our experi-
ments include more than 10 hours of flight data covering a
distance of more than 20 km without any major incident, a
testament to the robustness and reliability of today’s UAV
technology.
The Camazotz node on the UAV broadcasts packets with
a payload of 32 bytes and a frequency of 8 packets per sec-
ond. The base station node is installed at approximately
1.5 m above the ground and is equipped with a ground plate
with a diameter of approximately 20cm. We record the ra-
dio signal strength indicator (RSSI) for each received packet.
The packet reception rate (PRR) is estimated using sequence
numbers included in the broadcasted packets and each packet
is timestamped by the PC time at the reception.
The Pelican platform provides a software development kit
that enabled us to log GPS and inertial data from the Pel-
ican autopilot. This data is useful in evaluating GPS ac-
curacy, as well as correlating Camazotz radio performance
to the relative speed, height, or distance between Cama-
zotz and the base station. We use an XBee connected to
a laptop to communicate with the UAV and recorded the
autopilot data at 4 Hz. We use the UAV’s GPS for latitude,
longitude, and heading and UAV’s inertial sensors for alti-
tude and speed estimates to compensate for GPS errors. We
have also written a Python interface to Google Earth that
shows the location of the UAV in real time and can replay
the recorded experiments.
3.1.2 Antenna Selection
We consider two basic antenna types for our platform: an
EZConnect 868 MHz chip antenna manufactured by Frac-
tus [9] and a quarter-wavelength whip antenna. We test
two versions of the chip antenna as the ground-plane in the
development kit was significantly larger than the footprint
of our platform. The smaller configuration was designed to
match the footprint of Camazotz.
We run experiments for each antenna flying the UAV
along random trajectories, covering a number of different
heights, speeds, and distances. The overall length of the
recorded data for each antenna is approximately the same.
The distance between the flying node and the base station is
estimated from the UAV GPS data and the known location
of the base station.
-50
-55
-60
-65
-70
-75
-80
-85
-90
RSSI (dBm)
Range (m)
RSSI (0 ms-1)
RSSI (2 ms-1)
RSSI (4 ms-1)
0 10 20 30 40 50 60
Figure 4: Correlation of mobile to base RSSI signal,
conditional on the relative speed of the UAV and
the base station.
-50
-55
-60
-65
-70
-75
-80
-85
-90
RSSI (dBm)
0 10 20 30 40 50 60
Range (m)
RSSI (0o)
RSSI (25o)
RSSI (50o)
Figure 5: Correlation of mobile to base RSSI signal
over their distance, conditional on the altitude angle
between the UAV and the base station.
We plot our results in Fig. 3. The chip antenna with a
small ground-plane (ChipS) is clearly performing the worst
and experiences significant packet losses at distances of 20-
30 m. Somewhat surprisingly, the simple whip antenna out-
performs the unmodified chip antenna (ChipL) at most dis-
tances and experiences almost no packet loss at all tested
points. In addition, its performance is much more depend-
able as shown by the smaller variance of the RSSI signal.
Our conclusion is to use the quarter-wavelength whip an-
tenna as it provides superior performance at a lower cost.
3.1.3 Impact of Speed
We next study the impact of the relative speed of the Ca-
mazotz radios on the packet reception rates. We have flown
our UAV platform in a series of experiments designed to re-
semble flying fox flight patterns at up to half of their cruising
speeds. Figure 4 does not show any significant correlation
between the speed and the received signal quality, so the
radio scheduling algorithm thus does not need to constrain
packet transmissions based on the speed.
3.1.4 Impact of Angle
Finally, we study the impact of the node orientation on
the radio reception. Due to the constraints on the payload of
our UAV test platform, we did not study the impact of the
node heading as simulating a bat would require attaching a
one litre bottle of water to Camazotz. We thus focus our
study on the altitude angle between the node and the base
station. As the transmission pattern of antennas is not a
perfect sphere, radio performance is expected to decrease at
higher angles (when nodes are directly above base station).
Figure 5 confirms our expectations, albeit showing only a
minor degradation of the signal quality at higher angles. We
attribute this to the ground plate used at the base station
that helps to reflect some of the energy from the antenna

null areas. However, even with the ground plate, the ra-
dio communication is sensitive to the altitude angle which
should be considered during the deployment phase. In par-
ticular, we need to refrain from installing the base station
directly under the trees populated by flying foxes.
3.2 GPS
We conduct two different experiments to evaluate the per-
formance of Camaztoz’s GPS module. The first experiment
uses a Camazotz board in a static outdoor setup, while in the
second experiment the Camazotz board has been attached
to a captive live flying fox in a large outdoor cage.
For the first experiment, we attach the Camazotz board
to a tree on our campus to have similar conditions as in a
camp. The GPS receiver has a partly unobstructed view of
the sky. We configure the Contiki application running on
the Camazotz to switch off the u-blox MAX-6 GPS receiver
60 s after a position fix has been acquired. During the off
phase, only the backup voltage of the GPS module is active
which powers its real-time clock and the RAM. Therefore,
the GPS receivers can still keep the ephemeris information in
RAM and is able to do a warm start. We select the off time
interval uniformly at random between 10 s and 60 min and
measure how long it takes to acquire the first fix after power
to the module has been enabled again. The measurement
results for the time to first fix are shown in Figure 6(a). Our
results indicate that the time to first fix (TTFF) is correlated
with the time interval the GPS receiver was switched off,
confirming results from older GPS modules [6].
010 20 30 40 50 60
GPS off time (min)
0
5
10
15
20
25
30
Time to first fix (s)
(a)
0 2 4 6 8 10 12 14 16 18 20
Accuracy (m)
10%
20%
30%
Percentage of fixes
0%
25%
50%
75%
100%
Cumulative percentage of fixes
Calculated
Measured
Calculated,
(cumulative)
Measured,
(cumulative)
0%
(b)
Figure 6: GPS performance of Camazotz; (a) time
to first position fix for various off-time intervals,
(b) comparison of reported accuracy estimate and
a measure of true accuracy.
During the experiment on the living flying fox, the Ca-
mazotz logged 1 Hz GPS data to its SD card, and continu-
ously sent status update messages via radio to a base station
nearby. We assess the true accuracy of the GPS against the
accuracy reports from the GPS module. We set the true ac-
curacy as the distance of the GPS locations from the known
location of the animal for a period of time when it was roost-
ing in a single location. We choose a one hour period of the
day when we observed that the animal was in the one lo-
cation, hanging from a roosting location and occasionally
grooming and fanning itself. During this period the GPS
was in tracking mode where it was collecting fixes contin-
ually at 1 Hz, collecting 3600 fixes. We took the average
location of all of these fixes over the time period and use
it as the true location of the animal. Geo-referenced high
Component Power Duty Cycle Power
100% (mW) %DC (mW)
GPS 74 3 2.2
Radio 99 2 2
Cpu 13.2 5 0.7
Flash 40 1 0.4
Acc/Mag 2.6 10 0.3
Pressure/Temp 0.1 100 0.1
Mic 3.3 1 0.03
Totals 232.2 5.7
Table 3: Power consumption at 100% and target
duty cycles of Camaztoz components.
resolution imagery [11] with a spatial accuracy of 1 m was
then used to confirm the coordinates for this location, which
was considered the true location. To measure the accuracy
we calculate the distance from each fix to this true roosting
location, and for each fix compare this figure to the accuracy
estimate calculated by the GPS unit.
Figure 6(b) summarises our results. The measured ac-
curacy (M=5.9, SD=3.0) is significantly lower than the re-
ported accuracy ((M=7.2 , SD=1.3); t(7206)=24.2 , p<0.01)
from the GPS module, with a minimum accuracy value re-
ported by the GPS of 3.9 m, while the measured accuracy
data indicates that 3% of the fixes were within 1m accu-
racy. The calculated accuracy values are much more clus-
tered than the measured values, with 87% of the calculated
values being between 5 and 9m. These results indicate that
the GPS unit generally provides conservative estimates of its
accuracy in our experiment, as the true positions are more
accurate than the reported accuracy measurements suggest.
We note that the GPS results in our experiments are only
indicative of the module’s performance in our specific testing
scenario. Both the TTFF and the reported accuracy may
vary as flying foxes move to different environments. While
this paper introduces GPS sampling based on multimodal
sensor inputs, an interesting future direction for this work is
to adapt GPS sampling schedules to observed variations in
GPS module performance in addition to observed context.
3.3 Long-term Operation
3.3.1 Solar on collars
The Camazotz platform makes use of solar panels to help
ensure long term operation. Figure 7 shows the result of an
experiment logging solar charge current at 1Hz that includes
two Camaztoz platforms, one on a live bat and another
nearby on the ground in full sun exposure. The large dips
shown in the static node’s solar charge current are caused
by shadows from the structure of the bat enclosure that the
device was deployed in. The measured solar charge current
of the bat node is significantly lower than the the reference
static node, as one would expect given the non optimal ori-
entation of the bat node and also the bat’s insistence on
resting in a shady location for the majority of this experi-
ment. The small peaks over 5mA shown in the bat node’s
solar charge current occurred when small glimpses of sun-
light were caught by the solar panel, and are more represen-
tative of what would be achieved depoyed on bats in the wild
that tend to spend long periods in the sun while roosting.
3.3.2 Lifetime Implications
Building on these solar current experiments, we now ex-

12:00:00 13:00:00 14:00:00 15:00:00 16:00:00
Local Time
0
10
20
30
35
25
15
5
40
Current (mA)
Bat
Static Node
Average
Figure 7: Solar current captured from static node
versus bat node.
plore the lifetime implications for our nodes. Camazotz will
have a 300mAh Li-Ion battery with an average voltage of
3.8 V (range of about 3.3 -4.3 V depending on charge state).
The on-bat solar experiments indicate that we can obtain
an average of at least 3mA for 12 hrs from the solar panels,
which equates to about 5.7 mW average power input for a
full day.
We aim to keep the battery close to full charge to avoid
trickle charging, so we design for energy neutral operation
considering the energy consumption and harvesting. Table 3
shows the power consumption of the Camazotz components
in full operation mode and in operation at our selected duty
cycles. The average power consumption with these settings
is just below 5.7 mW (recalling that 5.7 mW is a conservative
estimate for harvestable energy), which meets the energy
neutral target and promotes long-term operation subject to
the lifetime of the physical components and the recharge
cycles of the battery.
Note that within the allowable 3% duty cycle of the GPS
module, setting a sampling schedule for the module remains
an issue for further investigation. Our forthcoming paper [13]
provides a detailed empirical analysis on the GPS tradeoffs
involving the off time, target position accuracy and its corre-
sponding energy consumption. We aim to characterise the
performance of the newer u-blox MAX-6 GPS module on
Camazotz and use a similar analysis to determine the most
appropriate GPS schedules.
3.4 Activity Recognition
In addition to tracking where flying foxes go, we are inter-
ested in what they are doing. The key activities of interest
are shown in Table 4. Flying is a key activity that requires
position estimates at high sampling frequencies (ideally in
the order of seconds) and should be trackable through the
GPS, air pressure (for height), and potentially inertial sen-
sors for wing beat frequency. Frequent daily interactions at
roosting camps among multiple animals fighting for territory
or in mating advances can be captured through a combina-
tion of distinctive sequential sounds and increased move-
ment. Urinating and defecating are important to detect and
localise for determining where and how flying foxes spread
seeds from fruits they have eaten. For that reason, GPS fixes
are desirable when these events occur. Detection is possi-
ble using the accelerometer to capture the instance that an
animal switches from its normal upside down stance to a
right side up stance. Grooming animals are typically in the
upside down position yet moving their head/neck to groom
themselves, which may be detectable through the inertial
sensors, and using their claws to scratch their bodies, which
creates a scratching sound. Resting is the default state for
most animals in a roosting camp, which typically does not
require the position lock and can be used as a baseline low
activity state for differentiating from other states. The lack
of a hardware motion trigger on the accelerometer indicates
this state, and pressure sensors can be used to estimate the
height at which animals are roosting for establishing hierar-
chies within a camp.
It is clear that detecting the above activities requires mul-
tiple sensory modalities, and in some cases the detection of
an activity should trigger a position lock through the GPS.
The remainder of this section empirically evaluates to what
extent the sensory modalities on our platform can capture
these activities of interest.
3.4.1 Audio
While acoustics associated with echolocating micro-bats
has been well covered in the literature [10], not much at-
tention has been given to using flying fox vocalisations as
an automated means of monitoring them. Some very early
work by Nelson [19] and relatively recent work by Parijs et
al. [21] presents some characterisation of the different calls
made by flying fox species. Adult and juvenile flying foxes
emit calls within the human audible frequencies mostly dur-
ing interaction events [17].
The in-built microphone on the Camazotz node is used
to capture audio at a sampling rate of 22.4 kHz. This is
sufficient to cover the full spectrum of sounds emitted by
flying foxes based on our initial studies using high quality
audio sampled at 96 kHz. These recordings at known roost-
ing camp sites show the most energetic part of the signal
to be within the 2 - 4 kHz range and the upper harmonics
to start fading away around 8 kHz. The audio data used
in this paper is first down sampled to 16 kHz and high-pass
filtered with fc= 1 kHz. We then process the audio stream
via an energy based detector as described in [2] to extract
acoustic events. These include vocalisations from the col-
lared individual as well as other bats within range. The
events include other background ‘noise’ such as bird calls
and anthropogenic sounds depending on the geographical
location of the roosting camp. In our dataset, we also have
loud scratching sounds when the bat is scratching the collar
node along with human voices, construction site sounds as
well as motor vehicle sounds.
Amongst the bat vocalisations, we identify multiple dif-
ferent call patterns. We focus on the sustained repetitive
call associated with aggressive interaction events [19]. We
use a set of three simple features to detect these particular
calls which are associated with interaction events: (1) mean
sound level, (2) call duration and (3) mean normalised fre-
quency. An example of a repetitive call associated with an
interaction event along with these features are shown in Fig-
ure 8. To facilitate implementation on the Camazotz node,
the 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 %. This gives us a
convenient and non-resource intensive method for extract-
ing acoustic features with reasonable accuracy by heuristi-
cally setting the threshold levels. Figure 9 illustrates the
process of selecting a suitable threshold for acoustic activ-

Activity
Sensors Timing
Audio Inertial Air Solar Event Event GPS Sampling
Pressure Duration Frequency Period
Flying X X hours daily high
Interacting X X seconds frequent on event
Urinating/Defecating X seconds frequent on event
Grooming X X seconds very frequent none
Resting X X X hours daily infrequent
Table 4: Key activities of flying foxes, their timing profile, and the sensors we use to detect them.
−15
−10
−5
0
5
10
15
20
25
30
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Time (s)
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
-30
-40
-60
-80
-100
-120
-140
Relative Sound Level (dB)
6.0
5.0
4.0
3.0
2.0
1.0
0.0
7.0
8.0
Frequency (kHz)
3.5 4.00.0 0.5 1.0 1.5 2.0 2.5 4.5 5.0
Time (s)
3.0
Sound Level (dB)
3.5 4.00.0 0.5 1.0 1.5 2.0 2.5 4.5 5.0
Time (s)
3.0
Normalized frequency
Mean
sound level
Duration of
sound event
Mean
normalized
frequency
Figure 8: Top: spectrogram of typical audio interac-
tion event. Middle: corresponding sound level and
zero crossings. Bottom: normalised frequency. Ar-
rows show derived acoustic features.
ity classification. Figure 9 plots accuracy, precision and the
performance metric [15] as the threshold is increased from
0. The performance metric is the product of accuracy, preci-
sion, sensitivity and specificity and serves as an indicator for
selecting a threshold which gives the highest accuracy while
maintaining a high level of precision. Figure 9 also shows the
receiver operator characteristic (ROC) curve which plots the
true positive rate vs. the false positive rate as the threshold
is varied. The indicated operating point corresponds to the
selected threshold of 0.002. Two-fold cross validation was
done over 1000 iterations to evaluate the performance of the
classification by splitting the dataset in half. This resulted
in a mean accuracy of 77.5 % and a mean precision of 70.5 %
relative to manually marked ground truth obtained via video
footage and external audio recordings.
3.4.2 Inertial
The inertial sensors on our platform enable the detection
of activities such as interaction among multiple animals, uri-
nating/defecating, and grooming behaviour, either individ-
ually or in combination with other sensors. For instance, ac-
celerometers can be combined with acoustic sensor data to
Threshold [x10-2]
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy, Precision & Performance
0.0 1.0 2.0 3.0 4.0
Accuracy
Performance metric
Precision
Used threshold
False positive rate
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
True positive rate
Operating point
Figure 9: Plot of accuracy, precision and perfor-
mance metric vs. classification threshold (left), and
receiver operator characteristic (ROC) curve for
acoustic activity classification (right) showing the
operating point corresponding to the used thresh-
old of 0.002.
detect interactions among multiple animals. Alternatively,
accelerometers can independently detect the full reversal of
orientation that occurs when flying foxes engage in waste
removal from their bodies.
We examine accelerometer signals collected from a fly-
ing fox collar at 128 Hz during the captive bat experiments.
Video footage and visual inspection serve as the ground
truth for this experiment. In order to visually distinguish
the angular inversion that occurs during urination activi-
ties, we compute the mean three-dimensional vector during
a 7 min portion of the experiment. The reason for choosing
the mean vector is that the flying fox remains in a down
facing position for most of the experiment, which indicates
that the mean vector should provide a decent estimate of
the constant gravitational force and serve as a reference for
orientation reversal. Figure 10 (top) shows the XYZ com-
ponents of the accelerometer signal projected on the mean
vector. There are clear sign inversions in all the accelerome-
ter dimensions in two instances in the trace. However, using
sign inversions to detect orientational flips is susceptible to
corner cases where one of the accelerometer dimensions is
orthogonal to the gravity vector.
We detect inversion events instead by computing the an-
gle θbetween the current 3-D acceleration vector ~c and the
inferred gravity vector ~g, using the following equation:
tan(θ) = norm(~g ×~c, ~g ·~c) (1)
where θis in degrees, and norm is the vector norm function.
The rationale for using angular shifts is that any 180°inver-
sion in orientation will result in a significant shift in θthat
is greater than 90°for a sustained period, which can only
correspond to waste removal events in flying foxes.
Figure 10 (bottom) shows the resulting angles. The rect-
angular boxes indicate two detected instances of inversion
events, while the left and right images show the correspond-

1.4 1.5 1.6 1.7 1.8 1.9 2
x 105
−2
0
2
4
Sample
Acceleration projection on
mean vector (G)
1.4 1.5 1.6 1.7 1.8 1.9 2
x 105
0
100
200
Sample
Angle − current and
gravity (degrees)
Figure 10: Top: Projected 3-D axis acceleration val-
ues over the duration of the experiment; Bottom:
Angle between the inferred gravity vector and cur-
rent acceleration vector. Two events with a sus-
tained angle shift are detected.
900
850
800
750
Time (min)
0 30 60 90 120 150 180
Height above mean sea level (m)
GPS
Pressure
Ground Height
(a) GPS verses air pressure
height above mean sea level
Time (min)
0
1
2
3
4
5
Height above base node (m)
0 50 100 150 200
(b) Flying fox altitude rela-
tive to base node.
Figure 11: Air pressure for estimating altitude.
ing video frames. The central image shows the typical fly-
ing fox orientation during the remainder of the experiment.
Examining the signals corresponding to the two detected
events, we can see a sustained angular shift of above 90°in
(θ) for the two detected events.
In order to verify our ability to automatically classify in-
version events on the Camazotz inertial signals, we manu-
ally marked the ground truth inversion events for 3hrs of
recorded on-bat accelerometer data during the afternoon,
using the video footage for ground truth. The ground truth
showed 11 true inversion events during this time period, with
an average duration of 5.91s for each event. We then ran
a classifier on the entire dataset to detect angle-shift events
by identifying contiguous samples of at least 4 s where the
angle was shifted by at least 90o. Our classifier detects all
11 true events, yielding 100% accuracy and precision.
3.4.3 Air Pressure
The altitude at which flying foxes fly or roost is of high
significance to ecologists, in order to characterise individual
and social behaviour. GPS is notoriously poor at providing
altitude information, where the vertical error is estimated
as twice the horizontal error on average. While a typical
GPS fix will have a horizontal error within 10 m, the verti-
cal error of 20 m does not provide sufficiently granular data
for understanding fine-grained flying fox interactions, such
as positional hierarchies in a roosting camp. We rely instead
on air pressure sensors for altitude estimation. Air pressure
itself can provide inaccurate estimates altitude because of
variations in atmospheric conditions. However, air pressure
measurements can use a ground-based reference measure-
ment in order to provide fairly accurate estimates of the
mobile nodes ground elevation. In the flying fox applica-
tion, known roosting camps will have base station nodes, so
it is easy to include an air pressure sensor at these nodes to
serve as ground reference.
When flying foxes are near roosting camps, they can use
the latest air pressure measurement from the nearest base
station. When they are far away from roosting camps, we
can revert to a nominal air pressure at sea level as a refer-
ence. An interesting direction for future work would be to
try to fuse air pressure and GPS altitude data to determine
if there is a performance gain for estimating altitude.
We conduct experiments with a collar-based mobile node
that measures air pressure on a flying fox at a bat hospital.
The flying fox is free to move within a large cage with a
variable height of up to 5 m, and a base station on the ground
measures air pressure for reference. Both the mobile node
and the base station measure air pressure in Pa, which does
not map linearly to altitude. We use the following equation
to convert the sensor readings from each of the two nodes
into the estimated height above mean sea level:
H= 44330 ∗(1 −(P /1013.25) 1
5.255 ) (2)
where P is the measured air pressure in hPa. In order to
estimate the ground height of the mobile node, we simply
take the difference between the H values from the mobile
node and the base node.
Figure 11(a) compares the altitude estimation of GPS and
air pressure on a Camazotz node on a bat collar. The air
pressure estimate is based on the ground reference at the
base station, which is located at 797 m above sea level. The
figure clearly shows the stability and consistency of the al-
titude estimate based on air pressure compared to the ex-
tremely noisy GPS altitude estimate.
Figure 11(b) illustrates the estimated ground height (the
difference between the bat and ground node altitudes from
Figure 11(a)) of the flying fox during a field trial of nearly
6 hrs. The data has been averaged over 1min time windows.
During the first 20 min, the fox is being fitted with the col-
lar in a 1 m high cage before being released into the larger
cage for the remainder of the experiment. The dip at around
200 min into the experiment happens at feeding time when
the bats descend. The height estimates were verified to be
representative of the animal’s movements through visual in-
spections and video recordings. Fluctuations in consecutive
samples appear to be within 0 to 50 cm, which establishes
an uncertainty bound for height estimates.
4. MULTIMODAL ACTIVITY BASED LO-
CALISATION
The tight weight, size, and energy constraints of long-term
localisation mean that the GPS module has to be aggres-
sively duty cycled. Because we require position fixes when
flying foxes engage in activities of interest, we use the diverse

0 400 800 1200
−2
0
2
Time (seconds)
Acceleration
ACCXACCYACCZDetected Interaction Events
(a)
0 400 800 1200
0
50
100
150
Time (sec)
Angle (degrees)
Changes in mean
angular shift
Angular shift
(b)
Time (s)
400 800 12000
Mean sound level (dB)
−40
−30
−20
−10
0
200 600 1000
Acoustic activity
(c)
Figure 12: Activity detection; (a) accelerometer sig-
nal - thick black line indicating detected interaction
events, (b) angular shift between gravity and cur-
rent acceleration vector. The changes in mean an-
gular shift clearly identify the start of the two true
events, (c) mean sound level - dashed lines show
detected acoustic interaction activity that involve
nearby animals but not the collared animal.
sensor modalities on Camazotz to detect these activities.
While some of these activities, such as urination/defecation,
can typically be detected with a single sensor, others, such
as interactions of multiple animals, require the fusion of mul-
tiple sensor outputs in order to determine that the activity
is taking place and whether the collared animal is engaged
in it.
We focus on detecting and locating interaction events that
involve the collared animal interacting with nearby animals
or the nearby animals interacting among themselves, as this
aids in mapping the social dynamics within a roosting camp.
We are particularly focused on longer interaction events that
may last from 25 s up to 1 min rather than spurious interac-
tions in the order of a few seconds, especially since localising
this activity may require a multi-second start-up time from
the GPS module. During these interactions, the animals
tend to repeatedly bend their body from their upside-down
stance, and on many occasions this movement is associated
with multiple sequential vocalisations. Both the accelerom-
eter and microphone can detect interactions involving the
collared animal, but only the microphone also can detect
interactions among nearby animals. We investigate further,
using 20 min accelerometer and audio data traces from our
captive bat experiments, and using video footage as ground
truth.
For the accelerometer trace, we observe that interaction
events exhibit much shorter term inversions than urination /
defecation events in the accelerometer traces, sometimes for
just above 10 ms, before reverting back to normal stances.
What distinguishes this activity is the repetitiveness of the
inversions within a window of several seconds. We can there-
fore distinguish interaction events through the average an-
gular shift between gravity and the current orientation over
a window in time corresponding to typical interaction event
durations. In particular, interaction events involving the
collared animals begin with an initial jerk where the animal
is agitated and nearly changes orientation before engaging
in repetitive short-term angular shifts.
We identify only two such events in our 3-hour data trace
(with durations of 25 and 54 seconds) using video footage
as ground truth. We use these two events to empirically
define thresholds for interaction event detection through ac-
celerometers to demonstrate the MAL concept, and we leave
the validation of threshold for when more data becomes
available. The accelerometer trace and true events are shown
in Figure 12(a), while the angles are shown in Figure 12(b).
It is clear that true events correspond to repeated short-term
inversions of 120°or more in the angle trace, while some false
events also exhibit sporadic angular shifts above 120°. We
differentiate these shifts by averaging angular shifts. For ev-
ery sample with at least 120°angular shift, we compute the
average angular shift in the subsequent 54 second window,
and take the derivative of the resulting average angular shift.
The starting times of the two true events in our data trace
correspond to the highest differentials in average angular
shifts of 43.8°and 60°(indicated by arrows in Figure 12(b)),
which capture initial jerks by a flying fox when engaging an-
other fox in aggressive interactions. The next highest peak
corresponding to a non-interaction event is at 18°. We there-
fore adopt 30°as our empirical threshold for the differential
in mean angular shift to distinguish interaction events of the
collared animal.
The mean sound level and detected interaction events are
shown in Figure 12(c). Each vertical line in the acoustic
activity plot represents an instance of sustained repetitive
vocalisation lasting for approximately 5s. As seen from this
plot, acoustic activity of other bats within range is also cap-
tured by the in-built microphone of the Camazotz node and
detected by the acoustic activity detection mechanism de-
scribed in Section 3.4.1.
Table 5 summarises how MAL contributes to better char-
acterisation of activities. While the accelerometer can cap-
ture the main interaction events of the collared animal, it
does not provide any information on interactions among
neighbouring animals, thereby missing two of the four events
of interest. The audio sensor is capable of detecting all four
events, but it is not able to distinguish which events involve
the collared animal. It is only through the combination of
these two sensor modalities that we can dissociate these two
types of events.
MAL can achieve the event type dissociations with mini-
mal increases to energy consumption over single-sensor trig-
gered approaches. We analyse the proportion of detected
events and the node power consumption that arises from
each strategy. We use the component power consumption
data from Table 3 and GPS lock time data from Figure 6(a)

Localisation Approach Animal interactions
Collared All Dissociated
Duty cycled GPS X
Accelerometer-triggered X
Audio-triggered X
Accel. AND Audio X
Accel. OR Audio X X
Table 5: MAL can detect all events and dissociate
interaction event involving collared animal or nearby
animals.
in our simulations. We compare a baseline approach of a
duty cycled GPS with a period of 20 s 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 1 min 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 451 s (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 33 mW. 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
time of 21 s and power saving of 86 % over the GPS duty cy-
cled approach. In comparison, the audio-triggered GPS can
detect all four interaction events of interest while keeping the
GPS active for a total of 64 s, corresponding to a node power
consumption of 7.42 mW. However, the audio-triggered ap-
proach can only determine that interaction events are occur-
ring nearby, but not whether the collared animal is involved.
MAL can be tuned to capture only interaction events in-
volving the collared animal, with comparable detection to
accelerometer and slightly higher power consumption for
powering the audio sensor. Alternatively, MAL can be tuned
to capture only nearby interaction events, yielding a 14%
reduction in power consumption over audio and correct de-
tection of the two interaction events involving only nearby
animals. Triggering the GPS on the basis of both the ac-
celerometer and audio activity detectors yields comparable
energy consumption to audio and correctly dissociates the
two types of detected events.
The main benefit of MAL is that it provides users with the
flexibility to tune performance to their current activities of
interest. If users are interested in collared bat interactions
only, they can simply use accelerometer triggers for obtain-
ing GPS samples and save energy in the process. If they
are interested in the cumulative set of interaction events
regardless of individual animal association with activities,
then audio is sufficient. If, on the other hand, users are in-
terested in pinpointing individual animals associated with
each activity, multimodal triggering of the GPS can provide
the data granularity for dissociating these event types.
5. RELATED WORK
The Networked Cow project [8] used PDAs with GPS and
adhoc-mode WiFi to route position information to a base
station. The work in [6] extends this cattle tracking ap-
plication to use short-range radio for relative localisation
Collared events Nearby events Power consumption
Detected Events
Average Power Consumption (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
Figure 13: Performance of MAL against
accelerometer- and audio-triggered GPS. MAL
can be tuned to capture either interaction events
of the collared animal, or nearby interaction events
only. MAL can also detect and dissociate both
types of interaction events with comparable power
consumption to audio.
alongside GPS. The ZebraNet project [5] reports individual
position records for zebras every few minutes. In order to
make the energy problem more tractable ZebraNet collars
include a solar panel, which assume that the panels are re-
silient to normal animal activities. Positioning is done by
GPS only, and the nodes propagate their information by
flooding in order to facilitate data acquisition by the mobile
sink. Dyo e al. [3] use a heterogeneous sensor network con-
sisting of RFID-based tags and base stations to track Euro-
pean Badgers over a prolonged period of time and highlight
the importance of interaction with domain scientists and
early prototyping, which are also central to our methodol-
ogy in design Camazotz. Our work shares the long-term
monitoring goals and network topology with [3], but Cama-
zotz includes GPS modules on the wildlife tags and aims to
push the size, weight, and lifetime of the nodes to new limits
through aggressive duty cycling based on MAL.
Anthony et al. [1] developed the CraneTracker system for
long-range long-duration tracking of the endangered whoop-
ing Crane. Their platform, weighing about 100 g, includes
GPS and inertial sensors as well as cellular and an Atmel
RF230 radio for short-range communication. Their design
aims at two GPS fixes/day and a communication latency of
less than 24 hours. While our work also targets long-range
and long-duration tracking of small birds, our target applica-
tion tracking flying foxes has much stricter design goals. For
instance, the device can not weigh more than 30 to 50 g or
5% of the bodyweight of the animals. Additionally, we aim
for position logs at the frequency of at least once every half
hour which results in a much higher utilisation of the GPS
module. The combined smaller footprint and higher GPS
sampling frequency for our application motivates our design
of the Camazotz platform. The use of accelerometers has
also been proposed as a low power indicator of movement to
supplement GPS duty cycling [20] [14]. Guo et al. [12] also
consider the use of directional and angular speed for cat-
tle behaviour classification. The work in [7] addresses the
tradeoff between localisation accuracy and energy efficiency.
A key difference with our work is that we use multiple sensor
modalities to trigger GPS duty cycling for more fine-grained
activity detection.
Recently, Liu et al. [16] proposed a sample-and-process
approach to dramatically reduce the active time for GPS
position sampling by up to three orders of magnitude. While

this approach is promising for reducing power consumption,
it requires post-facto offline processing to recover positions
and involves storing and transferring large amounts of data
per fix. An interesting direction for future work is to explore
the energy-implications of this sample-and-process approach
for long-term flying fox tracking.
6. CONCLUSION
This paper has introduced the feature-rich lightweight Ca-
mazotz platform for long-term tracking of flying foxes. We
have provided a comprehensive empirical evaluation of Ca-
maztoz in both laboratory and on-animal experiments. Our
results reveal a moderate radio communication dependency
on communication angle in 3-D mobile environments, and
confirm that whip antennas perform best. We have charac-
terised the time-to-first-fix of our GPS design as a function
of off-time on the ground. This was followed by on-bat ex-
periments that showed most of the GPS positions that the
GPS module accuracy estimate was generally conservative.
We also evaluate the expected solar charge for our design,
and plan the scheduling of our node components accordingly.
We have shown how multiple sensor modalities on Cama-
zotz can individually or collectively detect flying fox activi-
ties. Based on these findings, we have proposed and evalu-
ated the utility of Multimodal Activity-based Localisation,
where multiple sensors can jointly trigger the GPS for lo-
calising interaction events. Our results demonstrate that
combining sensor event detections can dissociate on-collar
and surrounding interactions for fine-grained control of GPS
sampling.
7. ACKNOWLEDGMENTS
This work was supported by the Batmon Project in CSIRO’s
Sensor and Sensor Networks Transformation Capability Plat-
form. The authors thank the paper shepherd Jakob Eriksson
and the anonymous reviewers for their valuable comments
that improved the paper quality.
8. REFERENCES
[1] D. Anthony et al. Sensing through the continent:
towards monitoring migratory birds using cellular
sensor networks. In Proc. IPSN, pages 329–340, 2012.
[2] B. Croker and N. Kottege. Using feature vectors to
detect frog calls in wireless sensor networks. The
Journal of the Acoustical Society of America,
131(5):EL400–EL405, 2012.
[3] V. Dyo et al. Evolution and sustainability of a wildlife
monitoring sensor network. In Proc. Sensys, pages
127–140, 2010.
[4] P. Eby. Seasonal movements of grey-headed
flying-foxes. Wildlife Research, 18:547–559, 1991.
[5] P. Zhang et al. Habitat Monitoring with ZebraNet:
Design and Experiences, chapter Wireless Sensor
Networks: A Systems Perspective. Artech House, 2005.
[6] R. Jurdak et al. Adaptive gps duty cycling and radio
ranging for energy-efficient localization. In Proc.
Sensys, pages 57–70, 2010.
[7] S. Tilak et al. Dynamic localization control for mobile
sensor networks. In IPCCC, pages 587–592, 2005.
[8] Z. Butler et al. Virtual fences for controlling cows. In
Proc. ICRA, pages 4429–4436, 2004.
[9] Fractus. Dipole antenna. http:
//www.fractus.com/index.php/fractus/srw_868/.
[10] M.A. Gadziola, J. M. S. Grimsley, P. A. Faure, and
J. J. Wenstrup. Social vocalizations of big brown bats
vary with behavioral context. PLoS ONE, 7(9):e44550,
2012.
[11] Queensland Government. North queensland 50cm
digital ortho-rectified aerial photography. Technical
report, Queensland Government, 2011.
[12] Y. Guo and G. Poulton et al. Using accelerometer,
high sample rate gps and magnetometer data to
develop a cattle movement and behaviour model.
Ecological Modelling, 220(17):2068 – 2075, 2009.
[13] Raja Jurdak, Peter Corke, Alban Cotillon, Dhinesh
Dharman, Chris Crossman, and Guillaume Salagnac.
Energy-efficient localisation: Gps duty cycling with
radio ranging. Transactions on Sensor Networks, 9(3),
2013. in press.
[14] M. B. Kjaergaard et al. Entracked: energy-efficient
robust position tracking for mobile devices. In Proc.
MobiSys, pages 221–234, 2009.
[15] Navinda Kottege, Frederieke Kroon, Raja Jurdak, and
Dean Jones. Classification of underwater broadband
bio-acoustics using spectro-temporal features. In
Proceedings of the Seventh International Conference
on Underwater Networks and Systems, 2012.
[16] Jie Liu, Bodhi Priyantha, Ted Hart, Heitor Ramos,
Antonio A.F. Loureiro, and Qiang Wang. Energy
efficient gps sensing with cloud offloading. In Proc.
SenSys, November 2012.
[17] J. Marko. “Pteropus scapulatus” (On-line), 2005.
http://animaldiversity.ummz.umich.edu/site
/accounts/information/Pteropus scapulatus.html.
[18] A. McKeown and D.A. Westcott. Assessing the
accuracy of small satellite transmitters on free-living
flying-foxes. Australian Ecology, 37:295–301, 2012.
[19] J. E. Nelson. Vocal communication in australian flying
foxes (pteropodidae; megachiroptera). Zeitschrift f¨
ur
Tierpsychologie, 21(7):857–870, 1964.
[20] J. Paek et al. Energy-efficient rate-adaptive gps-based
positioning for smartphones. In MobiSys, pages
299–314, 2010.
[21] S. M. Van Parijs and P. J. Corkeron. Ontogeny of
Vocalisations in Infant Black Flying Foxes, Pteropus
alecto. Behaviour, 139(9):1111–1124, 2002.
[22] http://www.asctec.de/uav-applications/
research/products/asctec-pelican/.
[23] L. A. Shilton et al. Landscape-scale redistribution of a
highly mobile threatened species. Austral Ecology,
33(4):549–561, 2008.
[24] C.R. Tidemann and M.J. Vardon. Pests, pestilence,
pollen and protein: the need for community-based
management of flying-foxes in australia. Australian
Biologist, 10:77–83, 1997.
[25] DA Westcott et al. The spectacled flying-fox, pteropus
conspicillatus. Technical report, June 2001.
[26] A.M. Zaki, S. van Boheemen, T.M. Bestebroer,
A.D.M.E. Osterhaus, and R.A.M. Fouchier. Isolation
of a novel coronavirus from a man with pneumonia in
saudi arabia. New England Journal of Medicine, 2012.