Occupancy Detection using Gas Sensors
Andrzej Szczurek, Monika Maciejewska and Tomasz Pietrucha
Faculty of Environmental Engineering, Wroclaw University of Technology,
Wyb. Wyspiańskiego 27, 50-370 Wroclaw, Poland
Keywords: Indoor Air, Occupancy Detection, Gas Sensor, VOC, Carbon Dioxide.
Abstract: Room occupancy is an important variable in high performance building management. Presence of people is
usually detected by dedicated sensing systems. The most popular ones exploit physical phenomena. Such
sensing solutions include passive infrared motion detectors, magnetic reed switches, ultrasonic, microwave
and audible sensors, video cameras and radio-frequency identification. However, in most cases either human
movement is needed to succeed in detection or privacy issues are involved. In this work, we studied occupancy
detection using chemical sensors. In this case, the basis for detecting human presence indoors is their influence
of chemical composition of air. Movement of people is not needed to succeed and privacy of occupants is
secured. The approach was reported effective when using carbon dioxide, which is one of major human
metabolites. We focused on volatile organic compounds (VOCs). Their consideration is justified because
numerous human effluents belong to this group. The analysis showed that VOCs’ sensors, such as
semiconductor gas sensors, offer comparable occupancy detection accuracy (97.16 %) as nondispersive
infrared sensor (NDIR) (97.36 %), which is considered as the benchmark. In view of our results,
semiconductor gas sensors are interesting candidates for nodes of sensor nets dedicated to detection of human
presence indoors. They are smaller, cheaper and consume less energy.
1 INTRODUCTION
Occupancy, is commonly recognized as the act of
occupying. The information about occupancy is
useful for numerous applications. First of all, it is an
important variable in determining the heating and
cooling loads as well as ventilation rates necessary to
maintain appropriate thermal comfort and indoor air
quality.
The availability of occupancy information allows
to significantly reduce energy consumption by heat,
ventilation and air conditioning (HVAC) systems
(Erickson and Cerpa, 2010; Erickson et al., 2011;
Brooks et al., 2014; Brooks et al., 2015; Goyal et al.,
2015). It is also extensively used for determination
of occupancy profiles, which are widely applied in
building simulations (Erickson et al., 2014).
Occupancy should be taken into account by building
commissions, for proactive building management,
diagnosis of indoor air quality complaints, and
investigation of building energy consumption.
Real time detection of occupants presence is
fundamental for control of lighting. It plays key role
in security systems for detecting abnormal human
activity. It was shown that the detection of changes in
occupancy patterns can be even helpful in revealing
clinical diseases such as depression (Dickerson et al.,
2011).
Occupancy is typically sensed by especially
dedicated sensing systems, which employ passive
infrared (PIR) motion detectors, magnetic reed
switches, ultrasonic, microwave and audible sensors,
video cameras, radio-frequency identification
(RFID), gas sensors, etc. (Nguyen and Aiello, 2013).
Currently, most of commercial systems which
perform occupant detection are based on PIR motion
detectors. These devices measure infrared light
radiating from objects in their field of view. Apparent
motion is detected when an infrared source with one
temperature, such as a human body, passes in front of
an infrared source with another temperature, such as
a wall. PIR motion detectors do not generate or
radiate any energy for detection purposes. PIR
detectors are particularly effective for controlling
lighting in infrequently occupied, small, closed
spaces such as storage rooms where a defined
detection pattern is required. They are ineffective for
more open layouts such as offices (Neida et al., 2001;
Szczurek A., Maciejewska M. and Pietrucha T.
Occupancy Detection using Gas Sensors.
DOI: 10.5220/0006207100990107
In Proceedings of the 6th International Conference on Sensor Networks (SENSORNETS 2017), pages 99-107
ISBN: 421065/17
Copyright
c
 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
99

Dodier et al., 2006). Although willingly applied for
occupancy detection, PIR sensors are rarely deployed
alone, because human motion is required to trigger
these detectors. In most practical applications, they
are used in conjunction with other sensors, e.g. with
magnetic reed switch door sensors to detect
occupancy for controlling HVAC (Agarwal et al,
2010).
Microwave and ultrasonic detectors utilize
Doppler shift principle. They transmit respectively
high frequency microwaves or sound waves in the
area. The occupancy is detected from the change of
pattern in the reflected wave. If the reflected pattern
is changing continuously then room occupancy is
assumed. If the reflected pattern remains the same for
a pre-set period of time, the detector sends
information about the lack of people inside.
Compared to other types of occupancy detectors,
microwave detectors have high sensitivity to objects
which move, as well as much greater coverage
(detection range). Additionally, they can detect
through glass. Therefore, careful consideration of
location is required in certain applications. The major
drawback associated with such detection principle is
that ultrasonic detectors, similar as microwave ones,
fail to detect objects which remain relatively still or
inactive for some time. For this reason, these devices
are mostly used in conjunction with detectors based
on audible sound sensors. However, audible sound
sensors are not able to distinguish between human
and non-human noises, which makes them inclined
toward false alarms.
Occupancy detection can also rely on the use of
video cameras (Ramoser et al., 2003). Camera
networks are perhaps the most common type of
sensor network. They are deployed in a variety of
real-world applications including surveillance,
intelligent environments and scientific remote
monitoring (Funiak et al., 2006). A wireless camera
sensor network installed in large multi-function
building for collecting data regarding occupancy
estimated occupancy with an accuracy of 80%
(Erickson et al., 2009). In many applications human
detection systems using video cameras are not a
preferred choice. It results from the intrusive nature
of these devices, the cost associated with their
deployment, large amounts of data for storing and
complexity of image processing.
Radio-frequency identification (RFID) uses
electromagnetic field to automatically identify and
track tags attached to objects. It can be applied for
occupancy detection. As demonstrated (Scott et al.,
2011), the authors put the RFID tags on the house
keys and RFID receiver was plugged in the server to
detect whenever a user carrying a key enters a house.
The information was used for automatic control of
house heating. Systems based on RFID identification
and video cameras can be used to detect occupancy,
but mostly they have been applied for security
purposes rather than in building control systems. In
general, practical applications prefer to avoid
techniques which require changes in user behaviour
such as RFID or which are a concern to user privacy
such as video cameras.
Current systems for occupancy detection usually
require the installation of dedicated sensors. They
need to be purchased, fitted, calibrated, powered and
maintained. This poses a number of critical
constraints, especially in domestic environments. On
the other hand, in typical households many sensor
devices are already available and they can be used to
perform occupancy detection in an opportunistic
manner. These devices can contribute to improve the
overall reliability of the detection system or reduce its
cost (Kleiminger et al., 2013a; Kleiminger et al.,
2013b). Examples involve taking advantage of GPS
coordinates from residents’ mobile phones, traces of
connections to WiFi access points or other data like
readings from digital electricity meters (Kleiminger
et al., 2011; Kleiminger et al., 2013a; Kleiminger et
al., 2013b). Other example is the occupancy sensing
where existing IT infrastructure can be used to
replace and/or supplement dedicated sensors in
determination of building occupancy (Melfi et al.,
2011). This approach is largely based on monitoring
MAC and IP addresses in routers and wireless access
points, followed by correlating these addresses to the
occupancy of a building, zone, and/or room.
Occupancy data obtained in this way can be used to
control lighting, HVAC, and other building functions
to improve building functionality and reduce its
energy use.
Despite the large number of potential capabilities,
detection of building occupancy is still the complex
and unsolved problem. Systems based on gas sensors
are the worthwhile option in applications such as
control of ventilation, heating and cooling
installations. According to the definition, given by the
International Union of Pure and Applied Chemistry
(lUPAC), these devices transform chemical
information, ranging from the concentration of a
specific sample component to total composition
analysis, into an analytically useful signal. In other
words, gas sensors can detect the presence of
appropriate gases in a near surrounding of these
devices.
Occupants themselves affect the chemical
composition of indoor air. They are a source of
SENSORNETS 2017 - 6th International Conference on Sensor Networks
100

emission of various chemical substances, e.g. carbon
dioxide, water vapour, volatile organic compounds as
well as biological agents. This emission results
mainly from normal metabolic processes. The
occupants influence on indoor air arises also from
their activities, lifestyle and behaviour. For these
reasons, chemical composition of indoor air very
often reflects the occupancy state, especially in non-
industrial environment.
Different gases can be used as indicators of this
factor. For example, under some circumstances the
number of occupants can be estimated from the real-
time measurements of the CO
2
concentration (Wang
et al., 1999; Jiang et al., 2016; Labeodan et al., 2015).
Occupancy information is also included in the total
concentration of volatile organic compounds emitted
by people inside room. Therefore, it can be extracted
from signals generated by CO
2
and VOCs sensors.
Different operating principles cause that
measurement characteristics, price and requirements
for use of these devices are also different.
The aim of this work is to compare properties of
CO
2
and VOCs sensors as detectors of occupancy.
2 EXPERIMENTAL
The occupancy detection was studied using a
university classroom as an example of the
periodically occupied space. The schematic drawing
of the classroom is shown in Fig. 1. Its dimensions
are: 7.35 x 9.60 x 3.20 m. The space was designed to
host forty students and the lecturer. The room is fitted
with openable windows and naturally ventilated.
The study focused on working days i.e. when the
classroom was busy with students. Eighteen days of
this kind were considered.
In the classroom there were performed
measurements involving gas sensors. The daily
measurement session started in the morning and it
was completed in the evening. The session consisted
of a survey and instrumental measurements. The main
aim of survey was to collect information about room
occupancy. For this purpose, an appropriate enquiry
form was prepared and used. The data about room
load was later used as reference, for training
occupancy detection models.
Instrumental measurements consisted in
recording responses of gas sensors of various kind.
The following devices ware used in the study: non
dispersive infrared sensor (NDIR) for CO
2
concentration measurement, photo-ionization sensor
(PID), flame ionization sensor (FID) and
semiconductor gas sensors for the determination of
total concentration of VOCs. In the last group there
were included commercial sensors offered by Figaro
Engineering, Japan (www.figarosensor.com). The
following sensors were chosen: TGS800, TGSn822-
A0, TGS823, TGS825, TGS826, TGS830, TGS832,
TGS842, TGS2180, TGS2600, TGS2602, TGS2620,
TGS2104, TGS2444, TGS2201-gasoline (two),
TGS2201-diesel (two). The sensor data was the basis
for occupancy detection.
Figure 1: The schematic drawing of the classroom and the
location of measurement point (circle).
The listed sensors may be applied to measure
chemical characteristics of indoor air. The response
of NDIR sensor is proportional to CO
2
concentration
in gas mixture. The response of PID sensor is
indicative for the amount of volatile compounds
contained in gas sample, which can be ionized by the
PID lamp under photon emission energies of 10.6 eV.
These compounds include: aromatics, mercaptans,
organic amines, ketones, ethers, esters, acrylates,
aldehydes, alcohols, alkanes and some inorganics,
like ammonia and hydrogen sulfide. PID cannot
detect water vapor. FID sensor response is indicative
for the amount of organic compounds contained in
gas sample, which are ionized during combustion in
a hydrogen flame. Flame ionization detectors cannot
detect inorganic substances and some highly
oxygenated or functionalized species. Both PID and
FID sensor determine total concentration of all
species they detect. Semiconductor gas sensors
respond to wide range of species contained in air
samples. In particular, the sensing element of Figaro
gas sensors is a tin dioxide (SnO
2
), semiconductor
which has low conductivity in clean air. In the
presence of a detectable gas, the sensor's conductivity
Occupancy Detection using Gas Sensors
101

increases depending on the gas concentration in the
air. We chose sensors from two series 8xx and 2xxx.
Devices in the first series have ceramic base. They are
featured by good long term stability, but typically
consume around 830 mW for their operation. Sensors
in the second series were manufactured using thin
film technology. In most cases they have lower
energy consumption and smaller dimensions
compared with series 8xx.
Passive sampling was used for measurements
involving NDIR sensor. Measurements with PID
sensor, FID sensor and TGS sensors were performed
with the application of dynamic sampling. In the last
case, air sample was drawn from the measurement
point and it was delivered to the measurement
devices. Teflon tubes were used for this purpose.
Indoor air was monitored in one location, as
shown in Fig. 1. Data was recorded continuously, in
real time, with constant time resolution of 1 min.
3 METHODS
The following assumptions were made about the
occupancy detection.
1. Occupancy detection intends to deal with only
two states - when occupants are present or
absent in the space of the room.
2. Classifier is applied for distinguishing
between the two states.
3. The basis for the distinction are measurements
performed indoors using gas sensor.
4. The measurement data has the form of time
series {X
i
, i = 1, …, n} where n is the most
recent moment of data acquisition.
5. Occupancy detection is performed on-line, for
the current time moment n. The time
resolution of detection is the same as data
collection.
3.1 Classification
Occupancy detection was represented as a
classification problem. It consisted in distinguishing
two categories of room state: presence and absence of
people.
Two kinds of features were considered as the basis
of detection: sensor response X
n-j
where j{0,…, L},
and change of sensor response X
k
=X
n
-X
n-j
, where
j{0,…, L}. We used time lag j, j=0, 1, …, L to move
back from the time point of occupancy detection in
order to identify the period when the time series
contains information, which is useful for the purpose
of occupancy detection. In this work, we assumed
L=30 min.
In this work, there were considered three kinds of
feature sets.
Type 1 feature sets consisted of values of sensor
responses {X
n-j
}. The following sets of features were
involved in classification: A1={X
n
}, A2=A1{X
n-1
},
A3=A2{X
n-2
}, A4=A3{X
n-3
}, A5= A4{X
n-4
},
A10=A5{X
n-10
}, A15=A10{X
n-15
},
A20=A15{X
n-20
}, A25=A20{X
n-25
},
A30=A25{X
n-30
}. With these sets of features we
tested the usefulness of sensor responses recorded
between 0 and 30 min back from the moment of
occupancy detection.
Type 2 feature sets consisted of changes of sensor
responses {X
n-j
}. The following sets of features
were involved in classification: B1={X
n
},
B2=B1{X
n-1
}, B3=B2{X
n-2
}, B4=B3{X
n-
3
}, B5=B4{X
n-4
}, B10=B5{X
n-10
},
B15=B10{X
n-15
}, B20=B15{X
n-20
},
B25=B20{X
n-25
}, B30=B25{X
n-30
}. With
these sets of features we tested the usefulness of
changes of sensor responses encountered between 0
and 30 min back from the moment of occupancy
detection.
Type 3 feature sets consisted of both, values and
changes of sensor responses {X
n-j
, X
n-k
}. The
following sets of features were involved in
classification {A1, B1}, {A1, B2}, …, {A30, B30}.
Sets of features were constructed individually for
each sensor.
For classification, we applied k-Nearest
Neighbors (k-NN) algorithm (Webb, 1999; Park and
Kim, 2015). The major reason for choosing it was that
the properties of the classifier fit the characteristics of
data used for occupancy detection. K-NN is a non-
parametric method. i.e. none assumptions are made
about the distribution of the input data. Test vector
(whose label is unknown) is classified by assigning
the label which is most frequent among the k training
vectors nearest to the vector in question. Training
vectors are simply stored in the memory and no
explicit training phase is involved. Considerable
advantage is the simplicity of k-NN algorithm, which
makes it is easily implementable in hardware
solutions.
Certain characteristics of data used for occupancy
detection caused that k-NN was chosen for
classification. 1) It was observed that the
measurement data which represents distinct
categories of room state exhibit very limited
grouping. In such circumstances, parametric
classification approaches would not be favoured. 2)
SENSORNETS 2017 - 6th International Conference on Sensor Networks
102

In feature space, its parts occupied by data
representing categories presence and absence of
people heavily overlapped. In such cases, data point-
to-data point distance is preferred as the basis of
classification, compared with distance between the
data point and the centre of the group.
Classification of test vectors was performed in
leave one out mode. By trial and error method we
chose the parameter of the classifier k=3.
3.2 Performance Assessment
The occupancy was detected with predefined
temporal resolution. For the purpose of algorithm
evaluation, each result of detection was compared
with the true state of the room. The possible
combinations of detection outcomes versus possible
true states of the room are shown in confusion matrix
(Table 1).
Table 1: Confusion matrix for the detection of classroom
occupancy.
True state
People
present
People
absent
Detected
state
People present TP FP
People absent FN TN
The true positive case (TP) was when the presence
of people was detected and really, there were people
in the room at that time. True negative case (TN) was
when the absence of people was detected and really
the space was empty. The false positive case (FP) was
when the presence of people was detected, while in
reality there was no one in the room. False negative
case (FN) was when the absence of people was
detected while there were people inside.
The accuracy of occupancy detection was defined
as the ratio of TP and TN cases jointly to the overall
number of detections:

 =
 + 
+++
(1)
In this work, we additionally analysed false
negative rate:
 =

 + 
(2)
and false positive rate:
 =

 + 
(3)
False negative rate indicated how frequently the
occupied room was classified as empty. False positive
rate showed how frequently the empty room was
classifies as occupied.
4 RESULTS
We studied occupancy detection in lecture room. The
complete study spanned over 18 working days.
During 61 % of this time, the room was occupied and
it stayed empty over the remaining 39 %. Presence of
people was associated with the classes held.
According to the general rules, the duration of classes
at the university is 45 min or 90 min. They are
separated by 15 min or 10 min breaks, depending on
the time of the day. In reality, the temporal variation
of room occupancy was much more complex. On
many occasions classes started or finished earlier or
later than assumed. Sometimes, the adjacent classes
were aggregated by cancelling the break. As reported,
the number of people in the lecture room was constant
during most of the classes. The size of groups varied
between 9 and 43 students. However, during some
classes the number of occupants varied, in particular
at the very beginning and at the very end of classes.
The occupancy detection was expected to cope with
both kinds of room load variation, temporal and
related to the number of people.
Figure 2: Time series of sensor responses recorded in the
classroom during an exemplary day together with
occupancy indication (black). Responses of the following
sensors are presented: NDIR sensor (red), PID sensor
(green), FID sensor (blue), TGS2201g2 (magenta).
In Fig. 2 we present the time series of scaled
responses of gas sensors recorded during an
exemplary day together with room occupancy. The
group of semiconductor sensors was represented by
one selected sensor TGS2201g2.
Occupancy Detection using Gas Sensors
103

Figure 3: Accuracy of occupancy detection based on
individual sensor responses recorded prior to the moment
of detection. The following sets of features were
considered: A1={X
n
}, A2=A1{X
n-1
}, A3=A2{X
n-2
},
A4=A3{X
n-3
}, A5= A4{X
n-4
}, A10=A5{X
n-10
},
A15=A10{X
n-15
}, A20=A15{X
n-20
}, A25=A20{X
n-
25
}, A30=A25{X
n-30
}. The lowercase subtrahend {0, 1, 2,
3, 4, 10, 15, 20, 25, 30} is the time lag associated with a
particular feature set. Bottom panel refers to semiconductor
gas sensors, which are grouped according to energy
consumption.
Based on results displayed in Fig. 3 to Fig. 5, the
most accurate occupancy detection was attained when
applying the sequence of sensor responses recorded
prior to the moment of detection (Fig. 3). Time series
of changes of sensor response were informative as
well (Fig. 4), but the accuracy of detection was
weaker when using features of this kind. In both
cases, increasing the length of the time series resulted
in the improved occupancy detection. We
demonstrated that in case of some sensors the time lag
of 30 min was sufficient for achieving high accuracy
(NDIR, TGSs), but when using other sensors longer
lags should be involved (PID, FID). Interestingly
(Fig. 5), combining values of sensor responses and
their changes in one feature set did not improve
occupancy detection as compared to values of sensor
responses only.
Figure 4: Accuracy of occupancy detection based on
changes of individual sensor responses recorded prior to the
moment of detection. The following sets of features were
considered: B1={X
n
}, B2=B1{X
n-1
}, B3=B2{X
n-2
,
B4=B3{X
n-3
}, B5= B4{X
n-4
}, B10=B5{X
n-10
},
B15=B10{X
n-15
}, B20=B15{X
n-20
},
B25=B20{X
n-25
}, B30=B25X
n-30
}. The lowercase
subtrahend {0, 1, 2, 3, 4, 10, 15, 20, 25, 30} is the time lag
associated with a particular feature set. Bottom panel refers
to semiconductor gas sensors, which are grouped according
to energy consumption.
Although NDIR sensor performed best in
occupancy detection (97.36 %), attention shall be
payed to the fact that highly competitive accuracy
was achieved with semiconductor gas sensors. A
number of sensors of this kind were only by a fraction
of percent weaker in occupancy detection accuracy
than NDIR sensor. These were TGS2201g2 (97.16
%), TGS2201g1 (96.86 %), TGS2444 (96.86 %),
TGS2201d2 (96.59%). Moreover, all semiconductor
gas sensors involved in the study offered high
performance. In general ACC in this group exceeded
93.99 %. Surprisingly, PID sensor demonstrated
lowest suitability for occupancy detection (74.51%).
SENSORNETS 2017 - 6th International Conference on Sensor Networks
104

FID sensor performed much better (91.22%)
compared with PID, but still it was not as good as
semiconductor gas sensors.
Figure 5: Accuracy of occupancy detection based on feature
sets A30 (blue), B30 (red) and A30  B30 (yellow).
Figure 6: False positive rate (FPR) and false negative rate
(FNR) occupancy detections using various sensors.
In Fig. 6 we analyse cases of room occupancy
misdetection. False positive rate and false negative
rate were applied for this purpose. As shown, for all
sensors, false positive rate was greater compared with
false negative rate. The smallest number of cases
when the empty room was wrongly classified as
occupied occurred with NDIR sensor (FPR=3.25%)
and with TGS2201g2 (FPR=3.25%). The smallest
number of cases when the occupied room was
wrongly classified as empty occurred with NDIR
sensor (FNR=2.23%) and with TGS2444
(FPR=2.39%). From the point of view of securing
proper conditions for people who stay indoors false
positive detections would be preferred to false
negative ones. Multiple false negative detections
could prevent maintaining human comfort by
stopping the work of supporting installations while
their operation is needed. Unfortunately, false
positive detections are not indifferent as well. They
make the supporting installations operate in vain and
cause unjustified energy consumption.
5 DISCUSSION
In this work we studied occupancy detection using
gas sensors: NDIR sensor, PID sensor, FID sensor
and semiconductor gas sensors. They characterize
indoor air from chemical point of view. It is known
that human presence indoors influences air
composition. This justifies the usefulness of gas
sensors for occupancy detection.
Our results confirm best detectability of people
presence indoors when applying NDIR sensor
(ACC=97.36 %). High performance of the device
which measures carbon dioxide concentration is in
line with earlier findings of other researchers
(Labeodan et al., 2015; Candanedo and Feldheim,
2016; Jiang et al., 2016).
The major achievement of this work is the
demonstration that there are other sensors, which
offer comparable accuracy of occupancy detection.
The competitive alternative to NDIR are
semiconductor gas sensors. With the best of them we
achieved 97.16% accurate detection. Some of these
sensors are already commercialized as indoor air
sensors. However, their use is not widespread yet. To
our knowledge, this is the first work which
demonstrated high performance of semiconductor gas
sensors in the application to occupancy detection.
Tendencies in indoor air monitoring
instrumentation incline toward PID sensor. Roughly,
it addresses the same aspect of indoor air as
semiconductor gas sensors. But, as we showed, PID
sensor is practically useless from the point of view of
occupancy detection (ACC= 74.51%). This fact
actually raises a concern about the ability of PID
sensor to follow human borne impact on indoor air.
In view of our results, semiconductor gas sensors
are very interesting candidates for sensing elements
in sensor nets for occupancy detection. As compared
to NDIR sensor, these devices are several times
cheaper. They are also smaller, and the
miniaturization is constantly in progress.
Additionally, their power consumption is
competitive. NDIR sensors typically consume 600
mW while semiconductor gas sensors, which we
selected as best consume, 502 mW (TGS2201g2) or
56 mW (TGS2444).
It shall be mentioned that occupancy detection
presented in this work engaged raw measurement
data. Some authors report advantages of initial data
Occupancy Detection using Gas Sensors
105

filtration or data smoothing techniques when
estimating occupancy level (Wang, 1999; Jiang et al.,
2016). In case of detection exclusively, the added
value resulting from this kind of pre-processing is not
obvious. However, the issue shall not be overlooked.
The drawback of the proposed approach to
occupancy detection is related to the use of classifier.
It causes that the detection model has to be tuned to
the space in which it is supposed to operate. However,
so far, solutions which do not involve classifier offer
considerably worse performance in terms of detection
accuracy.
6 CONCLUSIONS
This work focussed on occupancy detection in an
indoor space. The basis for detection were responses
of gas sensor. We considered NDIR sensor, PID
sensor, FID sensor and wide range of semiconductor
gas sensors.
Occupancy was detected in an exemplary lecture
room. In occupancy periods this space was populated
by 9 to 43 people. The detection was done with time
resolution of 1 min.
Our results showed that best sources of
information about presence of people in the room
were NDIR sensor (ACC = 97.36 %) and
semiconductor gas sensors, in particular TGS2201g2
(ACC = 97.16 %), TGS2201g1 (ACC = 96.86 %),
TGS2444 (ACC = 96.86 %) and TGS2201d2 (ACC =
96.59%). Interestingly, the source of least informative
data was PID sensor. The best achieved accuracy of
detection was very high, considering that responses
of individual sensors were used.
We demonstrated that time series of sensor
responses, recorded prior to the moment of
occupancy detection, are very useful for realizing this
task. The relevant information was available within
the time lag of at least 30 min. Changes of sensor
responses were considerably less informative that
their values.
ACKNOWLEDGEMENTS
This contribution was supported by the project: "The
variability of physical and chemical parameters in
time as the source of comprehensive information
about indoor air quality". The project is financially
supported by the National Science Center, Poland,
under the contract No. UMO-2012/07/B/ST8/03031.
REFERENCES
Agarwal, Y., Balaji, B., Gupta, R., Lyles, J., Wei, M.,
Weng, T., 2010. Occupancy-driven energy
management for smart building automation. In
Proceedings of the 2nd ACM Workshop on Embedded
Sensing Systems for Energy-Efficiency in Building.
ACM, 1–6.
Brooks, J., Kumar, S., Goyal, S., Subramany, R., Barooah,
P., 2015. Energy-efficient control of under-actuated
HVAC zones in commercial buildings, Energy and
Buildings, 93, 160–168.
Brooks, J., Goyal, S., Subramany, R., Lin, Y., Middelkoop,
T., Arpan, L., Carloni, L., Barooah, P., 2014. An
experimental investigation of occupancy-based energy-
efficient control of commercial building indoor climate.
In: Proceeding of the IEEE 53rd Annual Conference
on, IEEE, Decision and Control (CDC), Los Angeles,
CA, 5680–5685.
Candanedo L.M., Feldheim V., 2016. Accurate occupancy
detection of an office room from light, temperature,
humidity and CO
2
measurements using statistical
learning models. Energy and Buildings, 112, 28-39.
Dickerson, R. , Gorlin, E., Stankovic. J., 2011. Empath: A
continuous remote emotional health monitoring system
for depressive illness. In Proc. Wireless Health'11.
ACM.
Dodier, R., Henze, G., Tiller, D., Guo, X., 2006. Building
occupancy detection through sensor belief networks,
Energy and Buildings, 38(9), 1033–1043.
Erickson, V.L., Lin, Y., Kamthe, A., Brahme, R., Surana,
A., Cerpa, A. E., Sohn, M. D., Narayanan S., 2009.
Energy Efficient Building Environment Control
Strategies Using Real-time Occupancy Measurements,
In Proceeding of BuildSys '09 Proceedings of the First
ACM Workshop on Embedded Sensing Systems for
Energy Efficiency in Buildings, 19-24.
Erickson, V., Cerpa, A., 2010. Occupancy based demand
response HVAC control strategy. In Proceedings of the
2nd ACM Workshop on Embedded Sensing Systems for
Energy-Efficiency in Building (BuildSys 2010), 7–10.
Erickson, V.L., Carreira-Perpinán, M.Á., Cerpa, A.E. 2011.
OBSERVE: Occupancy-based system For efficient
reduction of HVAC energy. In Proceedings of the 10th
International Conference on, IEEE, Information
Processing in Sensor Networks (IPSN), Chicago, IL,
258–269.
Erickson, V.L., Carreira-Perpinán, M.Á., Cerpa, A.E.,
2014. Occupancy modeling and prediction for building
energy management, ACM Trans. Sensor Netw.
(TOSN), 10(3), 42.
Funiak, S., Guestrin, C., Paskin, M., Sukthankar, R., 2006.
Distributed Localization of Networked Cameras, In the
Fifth International Conference on Information
Processing in Sensor Networks, Proceedings of the
Fifth International Conference on Information
Processing in Sensor Networks, IPSN 2006, Nashville,
Tennessee, USA.
SENSORNETS 2017 - 6th International Conference on Sensor Networks
106

Jiang Ch., Masood M.K., Soh Y. Ch., Li H., 2016. Indoor
occupancy estimation from carbon dioxide
concentration, Energy and Buildings, 131, 132-141.
Kleiminger, W., Beckel, Ch., Santini, S., 2011.
Opportunistic Sensing for Efficient Energy Usage in
Private Households, In Proceedings of the Smart
Energy Strategies Conference, 1–6.
Kleiminger, W., Beckel, C., Dey, A., Santini, S., 2013a.
Poster Abstract: Using unlabeled Wi-Fi scan data to
discover occupancy patterns of private households, In
Proceedings of the 11
th
ACM Conference on Embedded
Networked Sensor Systems, 47.
Kleminger, W., Beckel, Ch., Staake, T., Santini, S. 2013b.
Occupancy Detection from Electricity Consumption
Data. In Proceedings of the 5
th
ACM Workshop on
Embedded Systems For Energy-Efficient Buildings,
BuildSys'13, 1–8.
Labeodan T., Zeiler W., Boxern G., Zhao Y., 2015.
Occupancy measuremnt in commercial office buildings
for dmend-driven control applicaions – A survey and
detection system evaluation, Energy and Buildings, 93,
303-314.
Melfi, R., Rosenblum, B., Nordman, B., Christensen, K.,
2011. Measuring Building Occupancy Using Existing
Network Infrastructure, In Proceeding IGCC '11
Proceedings of the 2011 International Green
Computing Conference and Workshops, 1-8.
Neida, B., Maniccia, D., Tweed, A., 2001. An analysis of
the energy and cost savings potential of occupancy
sensors for commercial lighting systems, Journal of the
Illuminating Engineering Society of North America,
111-125.
Nguyen, T.A., Aiello, M., 2013. Energy intelligent
buildings based on user activity: A survey. Energy and
Buildings, 56, 244–257.
Park Ch. H., Kim S. B., 2015. Sequential random k-nearest
neighbor feature selection for high-dimensional data.
Expert Systems with Applications, 42, 2336–2342.
Ramoser H., Schlogl, T., Beleznail, C., Winter, M., Bischof
H. 2003. Shape-based detection of humans for video
surveillance applications. In Proc. of IEEE Int. Conf. on
Image Processing, 1013–1016.
Scott, J. , Brush, A.B., Krumm, J., Meyers, B., Hazas, M.,
Hodges, S., Villar, N., 2011. Preheat: controlling home
heating using occupancy prediction, In Proceedings of
the 13th international conference on Ubiquitous
computing. ACM, 281–290.
Wang S., Burnett J., Chong H., 1999. Experimental
validation of CO
2
-based occupancy detection for
demand controlled ventilation. Indoor and built
environment, 8, 377-391.
Webb A., Statistical Pattern Recognition, Arnold, 1999.
http://www.figarosensor.com/
http://www.gassensor.com.cn.
Occupancy Detection using Gas Sensors
107