of 8
Sensing Group Proximity Dynamics of Firefighting Teams
using Smartphones
Sebastian Feese
earable Computing Lab.
ETH Zurich
Bert Arnrich
Wearable Computing Lab.
ETH Zurich
Gerhard Tr
Wearable Computing Lab.
ETH Zurich
Michael Burtscher
Department of Psychology
University of Zurich
Bertolt Meyer
Department of Psychology
University of Zurich
Klaus Jonas
Department of Psychology
University of Zurich
Firefighters work in dangerous and unfamiliar situations un-
a high degree of time pressure and thus team work is of ut-
most importance. Relying on trained automatisms, firefight-
ers coordinate their actions implicitly by observing the ac-
tions of their team members. To support training instructors
with objective mission data, we aim to automatically detect
when a firefighter is in-sight with other firefighters and to vi-
sualize the proximity dynamics of firefighting missions. In
our approach, we equip firefighters with smartphones and use
the built-in ANT protocol, a low-power communication radio,
to measure proximity to other firefighters. In a second step,
we cluster the proximity data to detect moving sub-groups.
To evaluate our method, we recorded proximity data of 16
professional firefighting teams performing a real-life training
scenario. We manually labeled six training sessions, involv-
ing 51 firefighters, to obtain 79 minutes of ground truth data.
On average, our algorithm assigns each group member to the
correct ground truth cluster with 80% accuracy. Considering
height information derived from atmospheric pressure signals
increases group assignment accuracy to 95%.
Author Keywords
mobile sensing; radio-based proximity; group clustering;
ACM Classification Keywords
H.1.2 User/Machine Systems; H.3.4 Systems and Software:
uted systems; J.4 Social and Behavioral Sciences
During firefighting missions each firefighter fulfills a spe-
function and relies on his peers. Firefighting teams
usually split into sub-groups to work in parallel on different
tasks. Depending on mission complexity and the comman-
der’s strategy, these sub-groups are more or less stable and
Permission to make digital or hard copies of all or part of this work for personal or
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. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
, September 9–12, 2013, Zurich, Switzerland.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-2127-3/13/09...$15.00.
can merge and split again at any time. As the whole firefight-
team works towards a common goal, it is crucial that the
sub-groups coordinate their actions. However, coordination
between members of different sub-groups is complicated by
the fact that they might not be in visual contact.
Wearable computing can provide details on these group
dynamics by automatically measuring how group structure
changes during a mission. A graphical representation of who
was when in close proximity to whom illustrates mission de-
velopment over time allowing instructors to pinpoint possi-
ble coordination problems, which can be addressed in further
In this paper, we present a methodology to measure and visu-
alize group proximity dynamics of firefighting teams. Using
the built-in ANT
radio of smartphones, we scan nearby de-
fast and efficiently in order to detect sub-groups based
on the measured proximity. In particular, we make the fol-
lowing contributions:
1. We investigate the use of the low-power ANT radio to mea-
sure proximity between individuals and detail our search
strategy to detect nearby devices. Further, we characterize
discovery time and search distance.
2. We present a methodology to cluster moving sub-groups
within action teams using ANT-based proximity informa-
tion and extend the approach to also incorporate height in-
formation derived from atmospheric pressure signals.
3. For an easy understanding of group dynamics, we visual-
ize the group clusters over time in form of narrative charts
which represent who was when in a sub-group with whom.
4. We evaluate our group clustering algorithms in real-world
firefighting training sessions and compare the results to
manually annotated ground truth. We further show, how
a firefighting training mission evolves over time and high-
light important steps of the mission.
Related Work
Several research projects funded by the European Union
at supporting and increasing work safety of firefighters.
The ProeTEX project [3] developed a smart textile to moni-
tor the physiological status of firefighters. To support tacti-
cal navigation under poor visibility, a beacon based relative
positioning system was proposed during the wearIT@work
project [5]. To better integrate current practices of firefighting
brigades the approach was adapted in the ProFiTex project [4]
and resulted in a Smart Lifeline which enabled relative posi-
tioning. The NIST Smart Firefighting Project [2] combines
smart building technology, smart firefighter equipment and
robotics. Like in previous projects the aim is to provide
real-time information on firefighter location, firefighter vital
signs, and environmental conditions. The Fire Information
and Rescue Equipment project [1] at UC Berkeley combined
wireless sensor networks (WSN) and head-mounted displays
to support firefighters. A pre-installed WSN enabled room-
level localisation of emergency responders within a build-
ing [15]. The benefits and drawbacks of preinstalled location
systems, wireless sensor systems and inertial tracking sys-
tems for emergency responders were compared in [8].
In contrast to the above systems, we focus on group proxim-
ity rather than on localisation to capture mission development
and team activity. Our primary goal is to support post inci-
dent feedback with objective mission data. Although previous
system prototypes were tested in simulated scenarios none of
them were used in real-world trainings. In this paper, we de-
ploy and evaluate our method in real-world training sessions.
In the data mining community spatial-temporal data is mined
for moving objects by clustering methods which combine
time and location information [11, 10, 9]. Kalnis et al. [10],
split trajectory data in time slices and used a density based
clustering method to group close objects. Similar clusters
found in consecutive time slices were then considered as a
moving cluster of objects. In previous work [16], we have
extended the approach to handle noisy data and applied the
clustering method to GPS-trajectories of people walking in
groups through a city.
In the field of reality mining, the works by Eagle and Pent-
land have first explored the use of the mobile phone to mea-
sure proximity to others using repeated Bluetooth scans [7].
They showed that communities and daily routines of persons
can be identified from Bluetooth proximity networks. More
recent work, discovered human interactions from proximity
networks using topic models [6]. In both approaches, the
measured interaction data is aggregated in time slices of at
min duration and thus the discovered patterns are on
an even larger time scale.
Contrary to previous work, we use the low-power ANT pro-
tocol to scan for nearby devices. This allows us to detect
devices in close proximity much faster, usually in less than
ms compared to
s of a typical Bluetooth scan. This
increased time resolution by a factor of up to 50 enables us to
measure how groups of firefighters split and merge during a
mission in real-time.
ANT, similar to Bluetooth Low Energy, is an ultra low-
wer, low bandwidth wireless protocol which operates in the
Pop Device ID
from SearchList
Configure channel ch
Open channel ch
ch = ch + 1
Log (Device ID,RSSI)
Close channel ch
Append Device ID
to SearchList
ch = 1
N = min(MAX_CHANNELS, length(SearchList))
ch <= N
Device Found
ch > N
Start timer ch
Figure 1. Implemented list search to discover nearby devices.
range. Contrary to Bluetooth Low Energy it allows
a node to be master and slave at the same time and thus sup-
ports many network configurations. Currently, ANT is most
used in fitness devices such as chest-belts and pedometers.
ANT chips support up to eight logical channels on one physi-
GHz radio link using time division multiplexing. Each
ANT channel is identified by a tuple of network ID, type ID
and device ID. Configuring a channel includes setting the ID,
the frequency and the period.
Search Strategy
ANT offers different strategies to discover other devices: one
search a device with a known ID, search for devices
which match certain properties, e.g. are of particular type,
search near devices using proximity search and one can uti-
lize background searches. However, because the ANT chip
included in our Sony Xperia Active phones did not support
proximity nor background search mode, we implemented a
list search strategy. On each device one master channel con-
stantly transmits a device ID with a specified period and seven
slave channels are used to search in parallel for devices spec-
ified in a search list. In Figure 1 the search strategy is pre-
sented in form of an UML Activity Diagram. Given a list of
devices to search for, the first device ID is popped from the
list, a slave channel with the desired device ID is opened and
a search timer is started. In case that a device is found or
the search times out, the channel is closed, the device ID is
appended at the end of the list, the next device ID is popped
from the list and the channel is reopened. In case that the de-
vice is found, the received signal strength indicator (RSSI) is
The implemented list search is a simple device discovery
strategy. It is robust, as the search result of one device is inde-
pendent of the search results of other devices. But the search
strategy does not scale to a large search list as it takes time to
find all devices before a device can be searched again. To han-
dle large search lists, one could utilize a collaborative search
strategy, in which devices also send their known neighbours
to reduce the number of devices that have to be searched on
each device.