A Knowledge-Based Proactive Intelligent System for Buildings

Occupancy Monitoring

Marie Unmack Baerentzen

, Jalil Boudjadar

1 a

, Saif Ul Islam

2 b

and Carl Peter Leslie Schultz

1 c

Aarhus University, Denmark

Institute of Space Technology, Pakistan

Keywords:

Domain-Speciﬁc Language, Occupancy Analysis, Energy Efﬁciency, Optimization.

Abstract:

Occupancy monitoring for buildings is a key component to enable cost-effective allocation of spaces and

efﬁcient resources utilization. The occupancy monitoring systems rely on networks of sensors and cameras

to achieve high accuracy, however the main challenges are the privacy concerns and the computation cost.

This paper proposes the design of an intelligent energy-efﬁcient and privacy-aware system to track, monitor

and analyze buildings occupancy. The core idea is that rather than collecting large amounts of sensor data to

perform occupancy analysis post hoc, our proposal adopts a top-down approach where, using the knowledge

about the activity expected to be taking place it proactively identiﬁes the minimal data relevant to the actual

state following the semantics of the expected activity. Thus, it switches on/off the sensors in accordance with

such a subsequent dynamics and reduces the data amount to collect and the computation cost. The proposed

system has been built using a domain-speciﬁc language, implemented in C++ and tested for a building case

study. Our experimental results show that, while achieving a considerable reduction in computation cost (up

to 35%) and energy consumption (up to 31%), our system maintains high accuracy for occupancy tracking

compared to the state of the art solutions.

1 INTRODUCTION

Building occupancy monitoring enables to track the

state of the different spaces in real-time (Azimi and

O’Brien, 2022; Elkhoukhi et al., 2018; Pradeep Ku-

mar, 2016; Perra et al., 2021). The primary goal

of such systems is to provide a decision support to

buildings management to achieve a better spaces al-

location and resources-efﬁcient operation (Salimi and

Hammad, 2019; M et al., 2021). Particularly with the

ever-increasing prices of electricity, the need to op-

timize energy efﬁciency is of capital interest (Lasla

et al., 2019).

Occupancy tracking and analysis technology

has the potential to transform buildings efﬁciency

by automatically monitoring and identifying when

and where resources utilization can be minimized

(Elkhoukhi et al., 2018). However, current ap-

proaches for occupancy monitoring are driven by Big

Data Analysis so that to collect large quantities of sen-

sor data (video, audio, step panels, etc.) and employ

predeﬁned classiﬁers to detect patterns in the data

https://orcid.org/0000-0003-1442-4907

https://orcid.org/0000-0002-9546-4195

https://orcid.org/0000-0001-7334-6617

(Zhang et al., 2019) (Jiang and Yin, 2015). Not only

this analysis is slow, far to be real-time (Elkhoukhi

et al., 2018), it has relatively low accuracy (Salimi

and Hammad, 2020; Pan et al., 2014). Moreover, such

comprehensive, brute-force data collection processes

consume a high amount of energy and may raise pri-

vacy concerns, particularly in light of the new EU

General Data Protection Regulations (GDPR)

The primary aim of this paper is to design and im-

plement a prototype for building occupancy monitor-

ing and analysis using blind sensors, with the main

goal being able to intelligently and proactively con-

trol system sensors. So that to actuate the power

consumers in the building spaces, such as light and

air conditioning, following the actual and predicted

states.

Our key contribution is, rather than continuously

(and blindly) monitoring and processing all available

sensor data post-hoc to detect critical events (purely

bottom-up, highly computationally inefﬁcient), our

approach is knowledge-driven: identify which high-

level symbolic facts are needed to change the cur-

rent state into a critical event, and trace these facts

back through a well deﬁned semantics of the adopted

https://eugdpr.org

680

Baerentzen, M., Boudjadar, J., Islam, S. and Schultz, C.

A Knowledge-Based Proactive Intelligent System for Buildings Occupancy Monitoring.

DOI: 10.5220/0012144000003538

In Proceedings of the 18th International Conference on Software Technologies (ICSOFT 2023), pages 680-687

ISBN: 978-989-758-665-1; ISSN: 2184-2833

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

domain-speciﬁc language to identify which minimal

combination of sensors can supply data to conﬁrm

or contradict these facts (top-down, then bottom-up).

The start point of each iteration of this exploration is

the event expected at that time point.

This intelligent sliding of the sensors operation

stems from the problem that reactive systems, which

gather and process massive amounts of data, drain

a lot of power from the sensors. This challenge is

much severe especially when using battery powered

sensors.

The system we consider monitors a public build-

ing for academic and education activities. It has

various activities such as Lecture; Lesson; Seminar;

Meeting; Empty, which all have recognizable fea-

tures, each of which is identiﬁable through a subset

of the sensors with different conﬁdence levels. The

proactivation process consists in sliding the sensors

frequency, including switch On/Off, to optimize data

sampling in an intelligent way following the actual

state and the semantics of the expected/identiﬁed ac-

tivities. To automate the data analysis and the proacti-

vation process, we deﬁned a domain speciﬁc language

formed by a set of events that have well-deﬁned se-

mantics. A prototype of the proposed system has been

implement and tested against the state of the art (static

frequency) occupancy monitoring systems in terms of

computation cost and energy consumption.

Finally, Section 7 concludes the paper.

2 BACKGROUND

This section presents the background related to occu-

pancy analysis and optimization.

2.1 Occupancy Analysis

Occupancy analysis is the process of identifying the

occupancy state of a given building space, at different

time points, by recognizing human presence and the

activities conducted in such a building space (Ahmad

et al., 2021). It relies on collecting data from a sen-

sor network, or alternative technology, which usually

captures the impact made by human presence through

different features such as breathing, motion and noise.

This knowledge is mainly used to ﬁgure out how

many people are in a building space and subsequently

what is most likely ongoing activity at any time point.

To leverage the beneﬁt of using occupancy monitor-

ing, the occupancy analysis needs to be real-time so

that proper actions can be taken, either by a build-

ing manager or automated system, with respect to

minimizing resources waste and improving the indoor

comfort (Seghezzi et al., 2021).

Given that such systems are recently deployed on

local micro-controllers (edge or embedded devices)

using wireless battery-powered sensors, computation

cost and energy consumption are two key perfor-

mance metrics to optimize. One way of conducting

such an optimization action is by reducing the data

sampling and the underlying analysis, but this ap-

proach can have drastic impact on the outcomes ac-

curacy and reliability.

To back up the occupancy analysis, usually a

model is required (Yang et al., 2016). In fact, the

model contains information about known data, ac-

tivities and potentially the underlying semantics of

such activities. Real-time data is then run through

the model to identify the most likely ongoing activ-

ity. Occupancy analysis models are mostly domain-

speciﬁc due to the given building functionality, sen-

sors used, and building layout (Jiang et al., 2022).

However, there are occupancy monitoring models that

use extensively trained machine learning algorithms

(Zhang et al., 2022) before being put into actual use.

Either cases, the overall goal of occupancy analysis

is to track how building spaces are utilized, and also

to optimize the use of energy consumption and rooms

allocation (Sun et al., 2020).

2.2 Optimization

Optimizing a system consists in minimizing and/or

maximizing different performance metrics such as re-

sponse time, resources utilization and energy con-

sumption (Minoli et al., 2017). While improving

some of the performance indicators, the optimiza-

tion must not lead to deteriorating other metrics con-

siderably neither impacting the system functional-

ity (Boudjadar and Khooban, 2020), (Boudjadar and

Tomko, 2022).

Building spaces are usually multi-functionality

where a given space can be allocated to run different

activities. Thus tailoring the building space towards

the actual activities is of capital interest in the opti-

mization of buildings functionality and resources. For

occupancy analysis, this usually comes in the form

of reducing the power spent on measurements and

calculations while still maintaining low computation

cost and high accuracy outcomes. The key driver for

such an optimization is the real-time knowledge and

estimation of the occupancy state so that a sliding

of the sensors frequency (fewer data measurements,

longer durations in between and even possibly turned

off) can applied following the need of data to con-

ﬁrm/deny/investigate such states (Rault et al., 2014).

A Knowledge-Based Proactive Intelligent System for Buildings Occupancy Monitoring

681

A trade-off between the accuracy and the sliding of

the sensors functionality needs to be established and

maintained through an objective function. Accord-

ingly, building equipments (lights, ceiling fans, air

conditioning, etc) can be tuned following the actual

occupancy state.

3 RELATED WORK

Different occupancy monitoring and analysis models

have been proposed in the literature (Luo et al., 2017;

Rai et al., 2015; McKenna et al., 2015; Ahmad et al.,

2021; Perra et al., 2021; Ortiz Perez et al., 2018;

Abraham and Li, 2014). Mainly, the state of the art

approaches differ in terms of the technology used to

track the occupancy, analysis techniques and the fea-

tures used in the optimization (accuracy, energy, sen-

sors functionality, response time).

The authors of (Luo et al., 2017; Perra et al., 2021)

proposed an agent-based occupancy analysis, that

amounts at tracking the individual occupants rather

than measuring the occupants impact. This occu-

pancy monitoring alternative often models each occu-

pant individually to allow capturing where occupants

are expected to be at any time point. Although this

approach can have high accurate estimation and conﬁ-

dent occupancy predictability, it suffers from the scal-

ability and the computation cost given that the system

size is dependent on the total number of occupants.

Sanish et al. (Rai et al., 2015) combined the

agent-based occupancy analysis with a graph-based

model. This enabled to abstract away the detailed

occupant behavior and interactions such as position

tracking. Although this abstraction reduced the occu-

pancy model complexity, scalability remains a chal-

lenge in a similar way to (Luo et al., 2017).

A state-based occupancy model has been pro-

posed in (McKenna et al., 2015). It consists of cre-

ating stochastic data in terms of how probable a state

will occur. This process enables to infer the location

of people and their activities. Although the approach

enables ﬁne grained modeling and estimation of the

occupancy state, the bottleneck can be related to the

massive data gathering which drains sensor batteries

and computation resources.

Knowledge-based occupancy analysis (Ahmad

et al., 2021; Perra et al., 2021) amounts at design-

ing a domain-speciﬁc model encoding the activities to

be carried out in a given building space. By running

actual data through the model, a potential activity is

identiﬁed, thus either conﬁrming or denying the ex-

pected schedule. The only concern is being domain

speciﬁc, it requests expert knowledge to replicate the

occupancy analysis for different applications.

We align with the work in (Ahmad et al., 2021)

where we track occupancy by monitoring the impact

occupants make in a given building space. Moreover,

our events deﬁnition patterns are inspired from (Perra

et al., 2021). However, our proposed system enables

to identify the minimal data to look for, following the

real-time state, in order to recognize the ongoing ac-

tivity. This has led to reduce the operation cost, in

terms of data gathering, computation cost and energy

consumption.

Compared to the state of the art, our experimen-

tal results show that the proposed proactivation sys-

tem outperforms existing occupancy monitoring tech-

niques, where sensors are permanently active (Or-

tiz Perez et al., 2018; Abraham and Li, 2014), by re-

ducing drastically the data and energy requirements

to track occupancy. Moreover, it can achieve the

same knowledge level as (Perra et al., 2021; Ahmad

et al., 2021), where sensors operate with static sam-

pling time intervals, but with 30% less data gathering.

This results in less energy consumption for both data

gathering and data processing.

4 OCCUPANCY MONITORING

This section describes the model of our occupancy

monitoring system. The building spaces we consider

consist of closed/structured rooms that can serve as

ofﬁces, classrooms, meeting rooms, etc. Each in-

dividual space is monitored through a set of CO2

sensors, motion sensors, noise sensors and light sen-

sors. The sensors are grouped into different hubs, dis-

tributed in the room space to monitor, each of which

contains a sensor from each category. The light sen-

sors measure data in lux, CO2 in parts per million

(ppm), audio in decibels (dB), and motion in four lev-

els: none (0), low (1), medium (2), and active (3).

Moreover, the sensors can be conﬁgured on-the-ﬂy to

operate different frequencies.

To monitor and analyze the occupancy we deﬁne

a domain-speciﬁc language. This includes the gram-

mar and how to handle different levels of conﬁdence

to match the events the system can recognize. The

language describes all potential events and features in

a semantic way, thus enabling a thorough formal anal-

ysis of the occupancy state in real-time.

The proposed domain-speciﬁc language to model

and track the occupancy state is deﬁned as follows:

ICSOFT 2023 - 18th International Conference on Software Technologies

682

L , E1 | E2 | . .. | En

E , (F, c) | (F, c)∧E

F , S | S ∧ F

S , V ∈ [a, b]

c , H | M | L

Where E is an event, F is a feature, S is a sensor, V is

a sensor value that must be within a range [a,b], and c

is the conﬁdence level of a feature to identify an event

and can have values high (H), medium (M) or low (L).

E and F are non-terminal elements of the grammar,

which means they can create different constructs re-

cursively. This is in fact practical as different events,

respectively features, can have different numbers of

features, respectively different sensors associated.

The sensor inputs are combined in different ways

with different expected value ranges to identify the

various features of the system. The features in turn,

together with different conﬁdence levels, are com-

bined to deﬁne the system events.

Namely, the important features of our language,

derived from observations on the case study, are the

following:

• F1: 1 or 2 people in front,

• F2: Lecturing (consistent noise in front),

• F3: No lecturing,

• F4: Smartboard on,

• F5: No movement,

• F6: Smartboard off,

• F7: Many occupants present,

• F8: Few occupants present,

• F9: No occupant present,

• F10: Occupants quiet,

• F11: Occupants walking and talking,

• F12: Many people talking,

• F13: People talking and taking notes,

• F14: No or little noise,

• F15: Teacher sitting,

• F16: consistent sound from the back,

• F17: Front lights on,

• F18: Back lights off/dimmed,

• F19: No lights on.

A feature can be recognized using different sen-

sors and for different values. As an example, F1 ,

S2 ∈ [410, 700ppm] ∧ S3 ∈ [2, 3]. One can see that

some features, such as F11 and F19, may need spe-

ciﬁc interpretation in order to be speciﬁed and cap-

tured in the grammar. As an example, F11 feature

can be recognized if people are walking, i.e. having

a high value for the motion sensor and high value of

the noise sensor. To improve the creditability of the

events recognition, we consider predicates rather than

single-point values, where an occupant is walking for

example if the motion measure is within 70 to 100%

of the maximum sensor value. However, for other fea-

tures such F11, a feature is recognized only if the un-

derlying sensor measurement is the highest value of

the sample range. Each event of the system is then

made up of a number of features, each with a conﬁ-

dence level c.

The system we deﬁne can specify a large set of

events, recognize different features and incorporate a

large number of sensors. However, for the sake of

illustration in this work, we consider only 8 sensors

grouped into 2 hubs each located in one end of the

rooms to monitor. Moreover, we limit the set of events

to analyze the ﬁve following events present in the use

case: E1 is a lecture, E2 is a break, E3 is a meeting,

E4 is a lesson (exercise session), and E5 refers to an

empty building space. As an example, we specify a

lecture event to be:

E1 , < F1, H > ∧ < F2, H > ∧ < F4, M > ∧

< F7, M > ∧ < F10, L > ∧ < F18, L >

We may need to state that the deﬁnition of the

ﬁve events is not standard, however it is inspired from

(Perra et al., 2021) and formalized based on the obser-

vations made during data gathering of the use case.

The semantics of an event is then given by its set

of sensors, inferred from the underlying features, to-

gether with conﬁdence levels and measurement inter-

vals that are required to recognize and conﬁrm the oc-

currence of that event. The conﬁdence levels of an

event dictate the importance of each feature for the

event, which is important for optimizing the sensor

frequencies, for example sensors of features having

low conﬁdence for a given event e can be discarded

while e is being conﬁrmed through high conﬁdence

features. The calculation and optimization of sensors

frequency will be deﬁned in the next section.

A state of the system s is simply made up of a

value (v(S

)) from each of the eight sensors. Although

sensors data is time stamped to ensure consistency, to

simplify the state deﬁnition we omit to use time as

part of the state, s(t). Thus s refers to the actual state,

i.e. having the latest sensor readings, can simply be

made as follows:

s = hv(S

), v(S

), . . . , v(S

We overloaded V to be a valuation function that

returns the value of each sensor.

A Knowledge-Based Proactive Intelligent System for Buildings Occupancy Monitoring

683

5 PROACTIVATION: OPTIMIZED

OCCUPANCY ANALYSIS

This section elaborates further on how the occupancy

analysis in made together with how the sensors are

planned on-the-ﬂy and how the frequency sliding is

computed following the expected event and the actual

state.

To trigger the occupancy analysis, one may need

to have an expected event ε. The expected event can

be provided by the space schedule from the building

management. From that event, our analysis process

identiﬁes the sensors to be used to capture ε. This is

done according to the semantics of event ε as deﬁned

in Section 4.

It is important to trigger the analysis with an ex-

pected event ε, since this enables the algorithm to

know what it has to look for, and thus operates the

sensors needed only. The expected event can change

dynamically either by outside means or following the

knowledge built from actual sensors data, i.e. in case

occupancy data does not match the expected event ε.

In the absence of an expected event, our occupancy

analysis can start with a brute force data sampling

from all sensors then the event recognized can be set

to be the expected one.

Sensing Plan. A sensing plan R dictates which sen-

sors must be On and how often those sensors sam-

ple occupancy data. R is formally given by R =

h(S

, r

), (S

, r

), (S

, r

), . . .i, where S

are sensors

and r

are frequencies. The value of each r

is derived

from the conﬁdence value of the feature each sensor

belongs to. If a sensor belongs to multiple features

of a given event, the highest conﬁdence value is used

then. As an example, if an event e

is given by e

F1, H > ∧ < F2, L > where F1 , S

|V (S

) ∈ [a, b] and

F2 , (S

, S

)|S

∈ [c, d] then S

will be assigned high

conﬁdence for both features, which means S

will run

the same frequency for both F1 and F2, except that the

value of S

has to satisfy both [a, b] and [c, d] ranges

in order to conﬁrm occurrence of event e

5.1 Occupancy Analysis

To perform occupancy analysis, our algorithm com-

pares the sensors readings to the features of the ex-

pected event, or any event of the domain-speciﬁc lan-

guage L in case the expected event has already been

disapproved or not provided at all. This can in fact

lead to 3 cases: 1) good matching; 2) no matching; 3)

partial matching.

Good Matching. In case that sensor readings match

a single event e, then the expected event ε will be up-

dated to this new event e, i.e. ε := e. From the seman-

tics of e, a sensing plan is computed as follows:

R = h(S

, r

), .., (S

, r

) | ∀i ∃c

, c

) ∈ ||e|| ∧



2 if c

= H

4 if c

= M

8 otherwise

||e|| is the semantics of event e returning the sen-

sors needed, data ranges and conﬁdence levels to rec-

ognize e. The calculation of frequency measurement

for each sensor is based on the conﬁdence level of the

event each sensor is assigned to. By default, we as-

sign frequencies of 2, 4 and 8 minutes respectively to

sensors having high, medium and low conﬁdence lev-

els respectively. Those are commonly adopted sen-

sors frequencies for buildings occupancy monitoring

solutions (Ortiz Perez et al., 2018; Abraham and Li,

2014; Beck et al., 2021).

In fact, if an event is being conﬁrmed for few con-

secutive sampling iterations, an optimization of the

sensing plan is needed so that we start attenuating sen-

sor frequencies, or even switching off, those having

lowest conﬁdence levels as follows.

R = R \ {(S

, r

) | ∀i (S

, L) ∈ ||e||}

Further elaboration on the optimization of sensing

plans is provided in Section 5.2.

No Matching. No-match situations occur when the

actual sensing plan leads to match no event of the

given domain-speciﬁc language. In such a case, a new

sensing plan should be to turn on all sensors with high

frequency (2 minutes) to be able to get a starting point

so that either a good matching or a partial matching of

an event occurs.

R = h(S

, 2), (S

, 2), . . . , (S

, 2)i

The no-match case is also a disapproval of the ac-

tual expected event ε. In practice, one has to be care-

ful when re-conﬁguring the sensors frequency as the

no-match case is the most expensive in terms of en-

ergy consumption.

Partial Matching. This case corresponds to a par-

tial matching of the actual state to a single, or more,

event(s). We need to distinguish the two alternatives

(single or multiple) as the processing is different for

each case.

When the actual state partially matches a single

event e, a new sensing plan is derived from the actual

ICSOFT 2023 - 18th International Conference on Software Technologies

684

sensing plan R by including all the sensors belonging

to the features with high conﬁdence in event e. The

sensors to be included will run with high frequency

(2 minutes) since they are assigned high conﬁdence

H. The new sensing plan is calculated as follows:

R = R ∪ {(S

, 2)|∀i (S

, H) ∈ ||e||}

If this is still not enough to fully conﬁrm the event

partially matching the actual state, the sensing plan

will be upgraded to include the sensors belonging to

the features with medium conﬁdence in event e. Such

sensors will run with medium frequency (4 minutes)

since they are assigned medium conﬁdence M.

R = R ∪ {(S

, 4)|∀i (S

, M) ∈ ||e||}

If the partial matching situation persists, the ac-

tual sensing plan will be expanded to include all the

sensor, being Off, belonging to the features with low

conﬁdence in event e. This will lead to either conﬁrm

the actual event e (good matching case), or disapprove

the event e (no-match case). Thus, the actual sensing

plan R is updated as follows:

R = R ∪ {(S

, 8)|∀i (S

, L) ∈ ||e||}

When the actual state partially matches multiple

events, it is necessary to ﬁnd a way to decide which

event is actually ongoing without the need to activate

all sensors as that would not be cost effective. This is

done by calculating the difference between the events

partially matching the actual state. If two events e

and e

partially match the actual state, then we calcu-

late the set of features differentiating the two events

and e

as follows: D = ||e

|| − ||e

||.

This difference is then used to update the sensing

plan R, to be able to decide which of the events is

actually ongoing, as follows:

R = R ∪ D

Depending on how many features differentiate the

two events e

and e

, low conﬁdence ones can be ex-

cluded from the new sensing plan as to not turn on all

sensors and only operate the sensors leading to dis-

tinguish among the two events. A recursive approach

can be adopted in a similar way as for single partial

matching case.

5.2 Optimization of Sensing Plans

Optimizing a sensing plan consists in tracking the ac-

tual state and matching it to an expected event, or po-

tentially ending with a good match with an event even

though it is not the expected one. Following the actual

state, the optimization alternates between the differ-

ent cases (good, partial, no-match) mentioned earlier

while tuning the sensor frequencies accordingly.

Mainly, when the conﬁrmation of an event occurs

consecutively the number of sensors to use to track

such an event will be reduced. The ﬁrst time an event

is identiﬁed, no optimization will be done. On the

second consecutive conﬁrmation, all sensors assigned

to that conﬁrmed event with low conﬁdence will be

turned off. Similarly, on the fourth consecutive ap-

proved occurrence of the expected event all sensors

assigned to features with medium conﬁdence will be

turned off. On any consecutive occurrence conﬁrma-

tion (good match) beyond this, no optimization will

be done.

6 IMPLEMENTATION AND

EXPERIMENTAL RESULTS

The proposed system language, functionality, occu-

pancy analysis and optimization algorithm have been

implemented in C++. The implementation has been

made bottom-up where the basic constructs are im-

plemented as classes, thereafter such class types form

the variable types within other classes.

Since the number of features within an event and

the number of sensors in a feature are variable and

need to be known when instantiating the classes to

actual objects, we assigned a static array length for

each type that is going to be the maximum entities

and track the number of actual objects (for example

adding or deleting a sensor to a sensing plan) at run-

time.

To assess the performance achieved by our proac-

tivation algorithm, we have conducted an actual oc-

cupancy experiment analysis of a University building

(single) space for a 24-hour period. By performance,

we mean the decision accuracy, the percentage of

spared sensors and the energy saved due to some of

the sensors being turned off/stretched frequency part

of the time. The software implementation and data

gathered are available here

Figure 1 depicts the variable total number of sen-

sors being active at runtime, following the actual

state, for ﬁrst 9 hours among the 24-hour period. The

analysis for the last 15 hours is omitted as the only

event recognized is Empty, where the number of sen-

sors operating is static (2 sensors) and only the sen-

sors with high conﬁdence to that event are active i.e.,

namely motion and light sensors. This has saved con-

https://e.pcloud.link/publink/show?code=

kZw0sjZVvNS4RWIuoRSHhpTioRyMj3QF6NX

A Knowledge-Based Proactive Intelligent System for Buildings Occupancy Monitoring

685

Figure 1: Number of sensors active at runtime.

siderable energy amount, that would have been con-

sumed by 6 sensors over 15 hours if the proactivation

is not used.

A single sensor sampling data every 2 minutes

(2.5 minutes in (Perra et al., 2021)) consumes in av-

erage 0,13 kW/h, this amounts to 0,0043 kW per sam-

ple, if we disregard the slight differences of the en-

ergy consumption between the different sensor types.

Given that our use case operates 8 sensors, for 24

hours period the total energy consumption for data

sampling when using an occupancy analysis algo-

rithm that samples data statically every 2 minutes is:

p = 8 × 24 × 0, 13 = 24.96kw/day.

The total number of samples would be then 5760.

The test run of our proactivation algorithm for the

same setup and same data accumulated to 3942 sam-

ples, thus saving 31% of the data sampling compared

to the naive algorithm, while achieving the same ac-

curacy and state knowledge level. This means that

our intelligent occupancy analysis algorithms has led

to save 7,89kw/day compared to the static frequency

algorithm, such as in (Perra et al., 2021; Aftab et al.,

2013; Beck et al., 2021), that is mostly adopted in

many occupancy monitoring solutions. This savings

come from sensors only. The gain in energy saving

is far higher one comparing the proactivation solution

to the monitoring solutions where sensors are perma-

nently actively, such in (Ortiz Perez et al., 2018; Abra-

ham and Li, 2014). In fact, the real-time knowledge

synthesized by the proposed solution can be used as

a ground to optimize other energy consumption loads

such as lights, heating and ventilation.

The matching between expected events and input

data as well as sensing plan optimization percentage

have also been analyzed. The results are depicted in

Figure 2. This is to showcase the algorithm trade-

off between reliability, having more sensors active to

identify whether an event is actually ongoing, and the

performance optimization achieved by having fewer

sensors active to save on energy.

As one can see from that ﬁgure, the sensing plan

optimization (orange) kicks in a little while after the

Figure 2: Expected event matching and sensing plan opti-

mization.

event match (blue) has reached 100%. The sensing

plan optimization also does not go to 100% right away

but spends a couple of iterations at 50% ﬁrst. In

fact, the higher the event matching is, the higher the

sensing plan optimization will be. This is justiﬁable

as having the expected event conﬁrmed many times

leads to lower the sensors needed to track the contin-

ual availability of that event.

7 CONCLUSION

This paper proposed a proactivation system, that is

an intelligent knowledge-driven real-time occupancy

monitoring solution. The proposed analysis enables

to tune the sensors frequency on-the-ﬂy following the

actual state so that to reduce data sampling and energy

consumption.

The core idea is that rather than collecting large

amounts of sensor data to perform occupancy analysis

post hoc, we adopted a knowledge-driven approach

where we proactively identify the minimal data rel-

evant to the actual state following the semantics of

the expected activities. The proposed contribution has

been mathematically modeled and an early proof-of-

concept prototype has been implemented in C++.

Our solution has been tested and compared to re-

lated occupancy analysis alternatives where it outper-

formed by reducing the sensors energy consumption

with up to 31%.

As a future work, we plan to conduct an exten-

sive experiment on actual use cases with larger set of

events. Another future work would be to study the

complexity and schedulability of the proposed algo-

rithm on the actual multicore platform used for de-

ployment (Boudjadar et al., 2014).

REFERENCES

Abraham, S. and Li, X. (2014). A cost-effective wireless

sensor network system for indoor air quality mon-

ICSOFT 2023 - 18th International Conference on Software Technologies

686

itoring applications. Procedia Computer Science,

34:165–171.

Aftab, M., Chau, S., and Armstrong, P. (2013). Smart air-

conditioning control by wireless sensors: An online

optimization approach. In Proceedings of the 4th ACM

International Conference on Future Energy Systems,

pages 225–236.

Ahmad, J., Larijani, H., Emmanuel, R., Mannion, M.,

and Javed, A. (2021). Occupancy detection in non-

residential buildings–a survey and novel privacy pre-

served occupancy monitoring solution. Applied Com-

puting and Informatics, 17.

Azimi, S. and O’Brien, W. (2022). Fit-for-purpose: Mea-

suring occupancy to support commercial building op-

erations: A review. Building and Environment, 212.

Beck, M. M., Boudjadar, J., and Chougui, Y. (2021). En-

ergy efﬁcient real-time calibration of wireless sensor

networks for smart buildings. In Advances in Model

and Data Engineering in the Digitalization Era.

Boudjadar, J., David, A., Kim, J. H., Larsen, K. G., Nyman,

U., and Skou, A. (2014). Schedulability and energy

efﬁciency for multi-core hierarchical scheduling sys-

tems. In International conference on Embedded Real

Time Systems and Software ERTS.

Boudjadar, J. and Khooban, M. H. (2020). A safety-driven

cost optimization for the real-time operation of a hy-

brid energy system. In Proceedings of the 27th Inter-

national Conference on Systems Engineering.

Boudjadar, J. and Tomko, M. (2022). A digital twin setup

for safety-aware optimization of a cyber-physical sys-

tem. In In Proceedings of the 19th International Con-

ference on Informatics in Control, Automation and

Robotics.

Elkhoukhi, H., NaitMalek, Y., Berouine, A., Bakhouya, M.,

Elouadghiri, D., and Essaaidi, M. (2018). Towards

a real-time occupancy detection approach for smart

buildings. Procedia Computer Science, 134:114–120.

Jiang, J., Wang, C., Roth, T., Nguyen, C., Kamongi, P.,

Lee, H., and Liu, Y. (2022). Residential house occu-

pancy detection: Trust-based scheme using economic

and privacy-aware sensors. IEEE Internet of Things

Journal, 9(3).

Jiang, W. and Yin, Z. (2015). Human activity recognition

using wearable sensors by deep convolutional neural

networks. In Proceedings of the 23rd ACM Interna-

tional Conference on Multimedia.

Lasla, N., Doudou, M., Djenouri, D., Ouadjaout, A., and Zi-

zoua, C. (2019). Wireless energy efﬁcient occupancy-

monitoring system for smart buildings. Pervasive and

Mobile Computing, 59:101037.

Luo, X., Lam, K. P., Chen, Y., and Hong, T. (2017). Perfor-

mance evaluation of an agent-based occupancy simu-

lation model. Building and Environment, 115.

M, J., A, K., MZ, A., SS, W., and MA, H. (2021). Iot-based

occupancy monitoring techniques for energy efﬁcient

smart buildings. Turkish Online Journal of Qualitative

Inquiry, 12-3.

McKenna, E., Krawczynski, M., and Thomson, M. (2015).

Four-state domestic building occupancy model for en-

ergy demand simulations. Energy and Buildings, 96.

Minoli, D., Sohraby, K., and Occhiogrosso, B. (2017).

Iot considerations, requirements, and architectures

for smart buildings-energy optimization and next-

generation building management systems. IEEE In-

ternet of Things Journal, 4.

Ortiz Perez, A., Bierer, B., Scholz, L., W

ollenstein, J., and

Palzer, S. (2018). A wireless gas sensor network to

monitor indoor environmental quality in schools. Sen-

sors, 18(12).

Pan, S., Bonde, A., Jing, J., Zhang, L., Zhang, P., and Noh,

H. Y. (2014). Boes: Building occupancy estimation

system using sparse ambient vibration monitoring. In

Proceedings of SPIE - The International Society for

Optical Engineering, volume 9061.

Perra, C., Kumar, A., Losito, M., Pirino, P., Moradpour,

M., and Gatto, G. (2021). Monitoring indoor people

presence in buildings using low-cost infrared sensor

array in doorways. Sensors, 21.

Pradeep Kumar, H. (2016). Multi-sensor-based occupancy

monitoring for energy efﬁcient smart buildings based

on internet of things. ProQuest Dissertations and The-

ses.

Rai, S., Wang, M., and Hu, X. (2015). A graph-based

agent-oriented model for building occupancy simu-

lation. In Proceedings of the Symposium on Agent-

Directed Simulation.

Rault, T., Bouabdallah, A., and Challal, Y. (2014). Energy

efﬁciency in wireless sensor networks: A top-down

survey. Computer Networks, 67.

Salimi, S. and Hammad, A. (2019). Critical review and re-

search roadmap of ofﬁce building energy management

based on occupancy monitoring. Energy and Build-

ings, 182.

Salimi, S. and Hammad, A. (2020). Sensitivity analysis

of probabilistic occupancy prediction model using big

data. Building and Environment, 172:106729.

Seghezzi, E., Locatelli, M., Pellegrini, L., Pattini, G.,

Di Giuda, G. M., Tagliabue, L. C., and Boella, G.

(2021). Towards an occupancy-oriented digital twin

for facility management: Test campaign and sensors

assessment. Applied Sciences, 11.

Sun, K., Zhao, Q., and Zou, J. (2020). A review of building

occupancy measurement systems. Energy and Build-

ings, 216.

Yang, J., Santamouris, M., and Lee, S. E. (2016). Review

of occupancy sensing systems and occupancy model-

ing methodologies for the application in institutional

buildings. Energy and Buildings, 121:344–349.

Zhang, H.-B., Zhang, Y.-X., Zhong, B., Lei, Q., Yang, L.,

Du, J.-X., and Chen, D.-S. (2019). A comprehen-

sive survey of vision-based human action recognition

methods. Sensors, 19.

Zhang, W., Wu, Y., and Calautit, J. K. (2022). A review

on occupancy prediction through machine learning for

enhancing energy efﬁciency, air quality and thermal

comfort in the built environment. Renewable and Sus-

tainable Energy Reviews, 167:112704.

A Knowledge-Based Proactive Intelligent System for Buildings Occupancy Monitoring

687