Non-Invasive People Counting in Smart Buildings: Employing Machine

Learning with Binary PIR Sensors

Azad Shokrollahi

1 a

, Fredrik Karlsson

, Reza Malekian

1 b

, Jan A. Persson

1 c

and Arezoo Sarkheyli-H

agele

Internet of Things and People Research Center, Department of Computer Science and Media Technology,

Malm

o University, 211 19 Malm

o, Sweden

Sony Network Communications, 215 32 Lund, Sweden

{azad.shokrollahi, jan.a.persson, reza.malekian, arezoo.sarkheyli-haegele}@mau.se, fredrik.karlsson@sony.com

Keywords:

Smart Buildings, Occupancy Information, People Counting, Binary PIR Sensors, Machine Learning,

Non-Invasive Sensors.

Abstract:

People counting in smart buildings is crucial for the efﬁcient management of building systems such as energy,

space allocation, efﬁciency, and occupant comfort. This study investigates the use of two non-invasive binary

Passive Infrared (PIR) sensors for estimating the number of people in seven ofﬁce rooms with different people

counting intervals. Previous studies often relied on sensor fusion or more complex signal-based PIR sensors,

which increased hardware costs, raised privacy concerns, and added installation complexity. Our approach

addresses these limitations by utilizing fewer sensors, reducing hardware costs, and simplifying installation,

making it scalable and ﬂexible for different room conﬁgurations, while also ensuring high consideration of

privacy. Additionally, binary PIR sensors are typically part of smart building systems, eliminating the need for

additional sensors. We employed several machine learning methods to analyze motion detected by binary PIR

sensors, improving the accuracy of people counting estimates. We analyzed important features by extracting

event count, duration, and density from sensor data, along with features from the room’s shape, to estimate the

number of people. We used different machine learning models for estimating the number of people. Models

like Gradient Boosting, XGBoost, MLP, and LGBM demonstrated superior performance for their strong ability

to handle complex, non-linear relationships in sensor data, high-dimensional datasets, and imbalanced data,

which are common challenges in people counting tasks using PIR sensors. These models were evaluated

using performance metrics such as accuracy and F1-score. Additionally, the results show that features such as

passage events and the number of detected events, combined with machine learning algorithms, can achieve

good accuracy and reliability in people counting.

1 INTRODUCTION

Smart buildings integrate technology and infrastruc-

ture to enhance operational efﬁciency and improve

the quality of life for people by automating processes

and making decisions based on data (Alsafery et al.,

2023; Jamali et al., 2024; Liu et al., 2023; Chaudhari

et al., 2024; Natarajan et al., 2023). A key feature

of smart buildings is the use of advanced technology,

particularly the incorporation of Internet of Things

(IoT) sensors, to gather data on occupancy informa-

tion. This data is essential for various applications,

including optimizing energy use by adjusting systems

https://orcid.org/0000-0006-2237-3010

https://orcid.org/0000-0002-2763-8085

https://orcid.org/0000-0002-9471-8405

like lighting, heating, ventilation, and air condition-

ing based on occupancy. Other key applications in-

volve enhancing occupant comfort and security, as

well as improving building management, such as op-

timizing resource utilization and providing guidance

within smart ofﬁce environments.(Zhao et al., 2022).

However, the accuracy of this data can be affected by

challenges such as optimal sensor placement and pri-

vacy concerns (Ramzan et al., 2024). As IoT technol-

ogy continues to advance and the demand for smart

building solutions grows, there is an increased focus

on improving methods for generating accurate occu-

pancy data. This includes addressing privacy con-

cerns and improving data accuracy (Chaudhari et al.,

2024).

Occupancy information, including detecting pres-

ence, people counting, locating, activity detection,

394

Shokrollahi, A., Karlsson, F., Malekian, R., Persson, J. A. and Sarkheyli-Hägele, A.

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors.

DOI: 10.5220/0013141800003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 394-405

ISBN: 978-989-758-737-5; ISSN: 2184-433X

tracking, and identiﬁcation, is crucial for the manage-

ment of smart buildings, each presenting unique ben-

eﬁts and different applications in the smart building

context (Shokrollahi et al., 2024). People counting is

one of the levels that can be used in different applica-

tions in smart buildings. Determining the number of

people in a given space is essential in numerous types

of applications and is widely applied in various appli-

cations (Zhao et al., 2022). Accurate people counting

or estimation may result in substantial improvements

in energy management, space utilisation, security and

so on (Zhai et al., 2024; Gao et al., 2024).There are

three main kinds of approaches for people counting in

a given space: non-invasive, vision-based, and semi-

invasive sensors. Every category has unique char-

acteristics and is used in certain applications. Envi-

ronments that place a high value on privacy are es-

pecially inclined towards using non-invasive sensors,

since they do not collect sensitive personal data. On

the other hand, vision-based sensors use cameras and

image processing technology to identify and quan-

tify individuals, offering a signiﬁcant amount of accu-

racy but possibly giving rise to concerns over privacy.

Semi-invasive sensors achieve a compromise between

the two, providing both modest privacy and exten-

sive data collection. Today, non-invasive sensors are

becoming more popular in a variety of applications

as people become more concerned about their per-

sonal privacy. Additionally, advancements in machine

learning approaches have enabled these sensors to in-

fer more information from the sensor data than was

previously possible (Alsafery et al., 2023; Li et al.,

2024). Tao Li et al. discussed different types of sen-

sors used for people counting, along with occupancy

prediction algorithms such as statistical stochastic al-

gorithms, heuristic algorithms, nature-inspired meta-

heuristic algorithms, and hybrid algorithms. They

also noted that most research or implemented cases

focus on ofﬁce and residential buildings for people

counting applications (Li et al., 2024).

The rest of the paper is organized as follows: Sec-

tion 2 discusses related work, providing an overview

of existing methods and approaches in people count-

ing using various sensors and techniques. Section 3

presents the methodology used for non-invasive peo-

ple counting with binary PIR sensors. Section 4

presents the results, discussing the performance of

various machine learning models. Finally, Section

5 concludes the paper, summarizing the key ﬁndings

and implications of the research.

2 RELATED WORK

Vision-based approaches for people counting utilize

RGB cameras and computer vision techniques with

machine learning for precise occupancy estimation

(Zhai et al., 2024; Alhawsawi et al., 2024; Wang et al.,

2023). Navarro et al. used these methods for accurate

people counting (Navarro et al., 2022). Alishahi et

al. combined cameras and WiFi for occupancy pre-

diction (Alishahi et al., 2022). Brazauskas et al. de-

veloped ”Cerberus,” a privacy-preserving system us-

ing ceiling-mounted cameras and facial recognition,

offering cost-effective indoor monitoring (Brazauskas

et al., 2024). Regarding semi-invasive sensors, Sahoo

et al. conducted research using a privacy-respecting

thermal camera. The method was assessed against

several categorization algorithms in both sparse and

dense scenarios (Sahoo and Lone, 2023). The Ther-

mal Sensor Data-driven Occupancy Estimation Sys-

tem (TODOS) was created by Rajabi et al. as the

most advanced way to count people. It uses deep con-

volutional neural networks to achieve accuracy rates

of 98% to 100% (Rajabi et al., 2023). Naser et al.

employed the Thermal Sensor Array (TSA) architec-

ture to address challenges arising from diverse sensor

placements and environmental conditions. This archi-

tecture utilizes a deep convolutional encoder-decoder

network combined with an adaptive boosting tech-

nique, effectively segmenting human presence and

calculating occupancy with a maximum accuracy of

100% from sensor locations (Naser et al., 2020).

With regard to non-invasive sensors, using CO2

sensors is more popular for people counting in build-

ings (Risuleo et al., 2015; Tomokazu, ; Jiang et al.,

2016). The study conducted by Kim et al. used ma-

chine learning algorithms to predict occupancy levels

using data on CO2 concentration and ventilation sys-

tem information. The models achieved an accuracy of

0.9180 (Kim et al., 2023). Liang et al. also showed

how useful it is to use LSTM models along with cheap

CO2 sensors (Liang et al., 2024). In another study,

Lu et al. covered non-intrusive occupancy estima-

tion in buildings by combining deep learning with fea-

ture engineering. Their network used data from sen-

sors measuring environmental factors such as temper-

ature, CO2, and TVOC, as well as human interactions

with building components like windows and air con-

ditioning systems (Lu, 2024). Wang et al. used en-

vironmental sensors and Wi-Fi technology to predict

ofﬁce occupancy using kNN, SVM, and ANN mod-

els. They determined that the ANN model was the

most accurate. The maximum recorded occupancy

was 14 persons, and the integration of Wi-Fi and en-

vironmental data enhanced the reliability of predic-

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors

395

tions (Wang et al., 2018). Furthermore, Amayri et

al. used sensors like motion detectors, CO2 concen-

tration sensors, and power consumption monitors to

assess room occupancy at different levels, such as

Level 1 ((= 0), (> 0)), Level 2 ((= 0), (> 0 ≤ 3),

(> 3)), Level 3 ((= 0), (> 0 ≤ 2), (> 2 ≤ 4), (> 4)),

Level 4 ((= 0), (> 0 ≤ 1), (> 1 ≤ 2.2), (> 2.2 ≤ 3.2),

(> 3.2)), and Level 5 ((= 0), (> 0 ≤ 1), (> 1 ≤ 2),

(> 2 ≤ 3), (> 3 ≤ 4)). Their work utilized machine

learning algorithms, notably decision trees and ran-

dom forests, producing an average estimation error

of 0.18 to 0.19 (Amayri et al., 2016). In another

signiﬁcant study, Dobrilovic et al. employed Multi-

Layer Perceptron Regression (MLPR) with Wireless

Sensor Networks (WSN) data, including light, tem-

perature, sound, CO2, and PIR motion sensors, to

estimate room occupancy with high accuracy (up to

98.20%). Their research focused on optimizing the

MLPR model for various sensor combinations in a 24-

square-meter space (Dobrilovic et al., 2023). To esti-

mate room occupancy, Mao et al. developed a non-

invasive sensor fusion technique that included tem-

perature, CO2, sound, and PIR sensor data. They

tested models such as Random Forest, SVM, XG-

Boost, and Multilayer Perceptrons and found that

Random Forest performed the best. Their technique

demonstrates how machine learning may improve en-

ergy efﬁciency by precisely counting room occupants

using non-intrusive sensors (Mao et al., 2023). Hob-

son et al. conducted a study to investigate cost-

effective techniques for estimating the number of peo-

ple in buildings. They used sensor fusion from Wi-

Fi access points, CO2 sensors, PIR motion detectors,

and electrical load metres. The researchers created

models using both multiple linear regression and ar-

tiﬁcial neural networks. They discovered that Wi-Fi

device counts were very useful, with average R2 val-

ues ranging from 80.1% to 83.0%. The model that

performed the best was multiple linear regression,

primarily because it is both transferable and simple

(Hobson et al., 2019). In continuity of this work, the

research conducted by Kumari et al. employs a two-

layer approach utilising a range of sensors (including

PIR motion detectors, CO2 sensors, plug loads, light-

ing loads, electricity use, and Wi-Fi access points)

and machine learning models, speciﬁcally the Light

Gradient Boosting Machine (Light GBM), to attain a

99% accuracy and F1-score in forecasting indoor oc-

cupancy (Kumari et al., 2024).

One of the widely used sensors for occupancy in-

formation is the PIR sensor, due to its affordability

and unobtrusive nature, requiring no pre-existing in-

frastructure. PIR sensors are primarily used for mo-

tion detection and are often combined with other sen-

sors in sensor fusion systems to provide data for dif-

ferent levels of occupancy information. There are two

types of PIR sensors based on their output: signal-

based and binary-based PIR sensors. Binary PIR sen-

sors detect the presence or absence of motion through

the use of a straightforward on/off signal. These de-

vices are inexpensive, simple to implement, and re-

duce power consumption, making them appropriate

for simple motion detection and cost-sensitive appli-

cations. Nevertheless, their capacity to provide com-

prehensive motion data is limited. Signal-based PIR

sensors, conversely, generate an analogue signal that

ﬂuctuates in response to infrared intensity. This en-

ables them to offer comprehensive data pertaining to

the sizes, speed, and direction of objects in motion.

The ongoing signal processing of these sensors results

in increased power consumption, complicated con-

ﬁguration, and setup requirements. Although signal-

based sensors have a greater power consumption and

higher cost, they are well-suited for applications that

demand careful and precise motion analysis (Shokrol-

lahi et al., 2024; Yun et al., 2023; Ngamakeur et al.,

2023; Umutoni et al., 2023). Several research papers

have used signal-based PIR sensors to count individ-

uals using various approaches. Zhang et al. presented

number intervals for counting with a single PIR sen-

sor and achieved 85% accuracy for adjacent intervals.

They classiﬁed the number of occupants into intervals

as follows: (0), (1–2), (3–4), (5–6), (7–9), (10–12),

(13–15), (16–18), (19–22), up to (87–96), effectively

categorizing large groups (Zhang et al., 2023). Tsou

et al. used a PIR sensor array and CNNs to achieve

92.75% accuracy (Tsou et al., 2020). Yang et al.

obtained 99.5% accuracy with four sensors in room

corners using a neural network (Yang et al., 2020).

Raykov et al. employed a single sensor and inﬁnite

hidden Markov models (iHMM) to achieve 99% ac-

curacy (Raykov et al., 2016). For people counting us-

ing binary-based PIR sensors, Wahl et al. employed

pairs of PIR sensors placed strategically to count in-

dividuals. Inward-facing PIR sensors detect people

entering a room, while outward-facing PIR sensors

detect people leaving, providing for exact movement

monitoring across the company. These sensors were

placed across an ofﬁce ﬂoor to establish a dispersed

network, assuring complete coverage while reduc-

ing installation and maintenance expenses using solar

power (Wahl et al., 2012). Binary PIR sensors have

been widely utilized for people counting due to their

simplicity and cost-effectiveness. Hitiyise et al. in-

troduced a method using two binary PIR sensors at

entrance areas, with counts increasing when the outer

sensor is activated ﬁrst and decreasing when the in-

ner sensor is triggered ﬁrst, achieving improved accu-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

396

racy without machine learning (Hitiyise et al., 2016).

Similarly, Udrea et al. developed a system using two

sensors to detect motion within a limited time frame,

though it faced challenges in distinguishing between

simultaneous movements (Udrea et al., 2022). A real-

life experiment was done by Masciadri et al. in an

apartment with eight PIR sensors and a contact sensor.

They used a directed acyclic graph (DAG) to track

movements and estimate occupancy with 86.78% ac-

curacy for up to six people (Masciadri et al., 2022).

Additionally, Wang et al. introduced a linear Gaus-

sian dynamic model to convert raw sensor data into

structured vectors, enhancing the accuracy of predic-

tions in multi-resident tracking with about 80% ac-

curacy in the TM004 dataset. These methods high-

light the effectiveness of binary PIR sensors in vari-

ous occupancy detection and people counting scenar-

ios (Wang and Cook, 2020).

Most previous work has utilized PIR sensors pri-

marily for motion detection purposes, such as turn-

ing lights on and off, and these sensors are now com-

monly available for lighting control and occupancy

detection in most modern buildings. Regarding the

use of PIR sensors for people counting, the common

approach has been to incorporate them as part of sen-

sor fusion systems or employ signal-based PIR sen-

sors. However, while binary-based PIR sensors are

known for their simplicity and cost-effectiveness in

people counting, none of the existing studies have uti-

lized motion counting for people counting within a

room. Only one prior study has employed machine

learning for people counting based on binary PIR sen-

sors. This study highlighted the potential of using ad-

vanced machine learning algorithms to enhance accu-

racy by analyzing motion events detected inside the

room and extracting new features. In earlier research,

binary PIR sensors were mainly deployed in pairs at

doorways to detect entry and exit movements, trig-

gering based on whether individuals were entering or

leaving (Hitiyise et al., 2016). This method, how-

ever, has a notable limitation: if multiple people enter

the room simultaneously, the sensors may only trig-

ger once, counting all individuals as a single entry.

Additionally, there is no sensor inside the room to ac-

count for individuals after they have entered, leading

to further inaccuracies in people counting. To address

these issues, our approach uses one PIR sensor at the

entrance and another for counting motion within the

room and using advanced machine learning.

This research is based on Sony Nimway system,

which is designed to optimize ofﬁce space usage and

enhance workplace efﬁciency through advanced sen-

sor technology and data analysis. Based on that, we

introduce a novel approach by using one PIR sensor at

the entrance and another for counting motion within

the room. This approach not only improves the ac-

curacy of detection but also offers several advantages

over previous binary PIR sensor methods for people

counting by using machine learning. Our method re-

duces overall hardware costs and simpliﬁes the instal-

lation process by requiring fewer sensors. By lever-

aging machine learning techniques and extracting dif-

ferent features based on sensor data, room layout, and

time-based patterns, we enhance the accuracy of peo-

ple counting. This approach enables us to count mul-

tiple individuals in a room, addressing the limitations

of previous methods that could miscount group en-

tries as a single event. Furthermore, this approach

can be adapted to different room sizes and conﬁgu-

rations without the need for multiple entry/exit sen-

sors, making it more scalable and ﬂexible for vari-

ous smart building applications. Improved accuracy,

enhanced privacy protection, and reduced hardware

requirements lead to better system performance and

increased user satisfaction in smart building environ-

ments. Our approach ensures a more efﬁcient, secure,

and user-friendly solution that meets the needs of both

building managers and occupants. Additionally, This

method improves the reliability of people counting

compared to traditional binary sensor methods. While

our approach may not always provide exact counts, it

offers better overall accuracy and ﬂexibility for vari-

ous smart building applications.

3 METHODOLOGY

Based on the Sony Nimway system (Corporation,

2024), this study utilized data from this smart ofﬁce

solution to optimize ofﬁce space usage and enhance

workplace efﬁciency. The Nimway system’s fea-

tures—room booking, way ﬁnding, occupancy moni-

toring, and desk management—aid in effective ofﬁce

space management. It provides real-time insights into

occupancy and utilization, reducing unused space,

improving productivity, and enhancing workplace sat-

isfaction, making it a valuable tool for smart building

management.

In this study, based on Nimway data collected

from seven rooms within a building over three years,

each room was equipped with two PIR sensors: a PIR

motion sensor inside the room and a PIR passage sen-

sor at the entrance. The passage sensor at the entrance

detects movement through the doorway without the

ability to detect the direction of the movement, while

the motion sensor inside the room monitors activity

within the room. Each sensor has distinct ﬁelds of

view, as depicted in Figure 1.

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors

397

Figure 1: Motion Sensor and Passage Sensor.

The motion sensor is designed to cover an entire

room or a speciﬁc section of it. For optimal perfor-

mance, the motion sensor is mounted at an angle and

a height of 2.5 meters, allowing it to cover a room up

to 4 x 6 meters. To prevent detection of movement

outside the door, the sensor is positioned on the same

wall as the door, ideally placed as close to the center

of the wall as possible. In contrast, the passage sensor

is also mounted at a height of 2.5 meters and covers a

door or passage up to 2 meters wide. It is positioned

to look directly down above the door to ensure accu-

rate detection. Figure 2 shows the side and top views

of each sensor.

Figure 2: Side and top views of the motion sensor.

The primary objective of our study is to count the

number of people between each passage sensor event

to provide a more accurate and reliable estimate of

room occupancy, as illustrated in Figure 3. In this

ﬁgure, each ”P” represents a passage sensor event,

and each ”M” represents a motion sensor event. The

number of people (GT) corresponds to the number of

individuals between each passage sensor event, ver-

iﬁed using a camera. Our approach involves using

both non-invasive passage and motion sensors to send

event data to a gateway at speciﬁc time periods. If

there is an event during these periods, the sensors

transmit the data; otherwise, they wait for the next

time period. Due to this difference in timing, the in-

tervals between events detected by passage and mo-

tion sensors vary. Additionally, it is likely that more

events will be recorded during these time periods for

both passage and motion sensors. The gateway col-

lects and aggregates the data from both sensors and

then sends it to the cloud for further processing, lever-

aging the cloud-based nature of the Sony Nimway

system. Moreover, regarding occupancy status, if the

sensor inside the room detects movement, it indicates

that someone is there, and we consider the room oc-

cupied. Conversely, if the sensor at the entrance de-

tects movement but the sensor inside does not, it sug-

gests that the person has left the room and the room is

unoccupied, which is extracted based on sensor data.

Using both sensors together helps accurately detect if

the room is occupied or not.

Based on Figure 3, we extracted various features

from the sensor datasets. ”Duration” represents the

time between each passage sensor event, and ”Event

Count” is the total number of motion events recorded

between each passage sensor. ”Density” is calculated

as the duration divided by the event count. The ”Max

Motions Time” indicates the maximum time differ-

ence between each motion sensor event received by

the gateway (not the time between individual motion

events). Additionally, we count the number of pas-

sage events that occur before and after each duration

until the room becomes unoccupied, referred to as

”Passage Event Before” and ”Passage Event After,”

respectively. Moreover, as shown in Figure 4, we uti-

lized map data to extract features for different rooms,

such as length, width, and ﬂoor area. Another feature

is the room size, which indicates the number of seats

inside the room. Additionally, we extract time-based

features, such as ”Weekday” (e.g., day of the week

like Saturday) and ”Category” (e.g., non-ofﬁcial time,

ofﬁcial time, and lunch time).

After data collection, our data preprocessing

pipeline involved several critical steps to prepare the

data for modeling. After feature extraction, the data

undergoes encoding and scaling to standardize it for

the machine learning models. To address the problem

of imbalanced data, as illustrated in Figure 5, which

relates to the frequency of people counting, we cre-

ated different target levels based on threshold mov-

ing, as shown in Table 1. By using threshold moving,

we not only solve the problem of imbalanced data but

also create different levels for people counting. This

approach helps to balance the data and categorizes oc-

cupancy into various levels, thereby improving the

accuracy of people counting. The prepared data is

then split either randomly (80% training, 20% test-

ing) or based on room conﬁgurations to create training

and testing datasets. Various validation techniques,

such as hand-out, k-fold, and stratiﬁed sampling, are

employed to ensure reliable model evaluation. The

model training process involved comprehensive hy-

perparameter optimization using grid search for mod-

els such as Gradient Boosting, MLP, and XGBoost.

Key parameters, including learning rates, maximum

depths, and feature importance, were ﬁne-tuned to en-

hance model performance. Once trained, the model is

evaluated for its accuracy and reliability in counting

the number of people.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

398

Figure 3: People counting structure.

Figure 4: Preprocessing, training, testing, and evaluation in machine learning.

Figure 5: Frequency of people counting.

Table 1: Classiﬁcation Levels and Corresponding Thresh-

olds.

Level Classes Thresholds

1 12 (0), (1), ..., (10), (11 or higher)

2 7 (0), (1, 2), (3, 4), (5, 6), (7, 8), (9, 10), (11 or higher)

3 7 (0), (1, 2), (3 to 5), (6 to 8), (9 to 15), (16 to 20), (20 or higher)

4 6 (0), (1, 2), (3 to 5), (6 to 8), (9 to 12), (13 or higher)

5 5 (0), (1 to 3), (4 to 6), (7 to 10), (11 or higher)

6 5 (0), (1), (3 to 5), (6 to 10), (10 or higher)

7 4 (0), (1 to 4), (5 to 10), (11 or higher)

By employing these levels, we improve the robust-

ness and accuracy of our occupancy detection model,

making it ideal for smart building applications. This

approach helps optimize space usage, enhance energy

efﬁciency, and improve occupant comfort, aligning

with the goals of the Nimway system.

4 RESULTS

In this section, we ﬁrst present the results of an anal-

ysis of feature correlations to understand their rela-

tionships and signiﬁcance. Following this, we evalu-

ate the performance of various algorithms by examin-

ing their Accuracy and F1 scores. We also compare

the accuracy results using k-fold and stratiﬁed k-fold

cross-validation techniques. Finally, we discuss the

importance of certain features as determined by the

algorithms

In the previous section, we extracted various fea-

tures based on sensor data, map data, and time-based

data. Using these extracted features and considering

the different levels of people counting, we created a

correlation heatmap, as shown in Figure 6. The cor-

relation heatmap reveals the linear relationships be-

tween various features used in predicting occupancy

levels, indicating their importance for level predic-

tion. Key features such as ”passage event before”

and ”passage event after” show strong correlations

with the exact count (0.47 and 0.41, respectively),

indicating their critical role in capturing the dy-

namics of entry and exit events. Spatial features

like ”area” (0.3), ”size” (0.31), and ”length” (0.3)

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors

399

are moderately correlated with the exact count, of-

fering insights into the room’s capacity and po-

tential occupancy limits. Although ”event count”

has a lower correlation (0.16), it still contributes

valuable information about motion frequency within

the space.”max motions time” and ”density” exhibit

lower correlations with the exact count (-0.056 and -

0.041, respectively), suggesting they capture unique

aspects of occupancy not covered by the more sig-

niﬁcant features. ”Duration” has a very weak corre-

lation with the exact count (0.011), suggesting that

while it provides temporal context, it may not be a

strong predictor on its own. Other features, such as

”category” and ”weekday,” show negligible correla-

tions with the exact count (-0.051 and 0.042, respec-

tively), indicating they have minimal impact on pre-

dicting occupancy levels. Overall, the heatmap pro-

vides a visual representation of how different features

relate to each other and to each level of people count-

ing, helping to identify the most signiﬁcant predictors

for occupancy levels.

In our study, we implemented 15 different ma-

chine learning algorithms to analyze 7 levels of oc-

cupancy counting with different people counting in-

tervals. The data was split using different strate-

gies: ﬁrst, by considering one room as the test dataset

and another as the training dataset; second, by ran-

dom splitting (80% training, 20% testing); and third,

by using k-fold and stratiﬁed k-fold cross-validation

techniques. The results obtained from these differ-

ent strategies were closely comparable. Therefore,

we focused on presenting the results based on the

random splitting approach, as shown in the Table

2. The algorithms evaluated include SVM, Logistic

Regression, Ridge Regression, K-Nearest Neighbors

(KNN), Gradient Boosting (GB), AdaBoost, Bag-

ging, Extra Trees Classiﬁer, Decision Tree, Random

Forest, Naive Bayes, LightGBM (LGBM), CatBoost,

XGBoost, and Multi-Layer Perceptron (MLP). These

models were assessed for their accuracy and F1 macro

scores across different levels of occupancy counting.

Based on the detailed results presented in Table 2,

Gradient Boosting (GB) and XGBoost consistently

demonstrate superior performance across various lev-

els of occupancy counting. XGBoost, in particular,

achieved a high accuracy of 0.753 and a strong F1

score of 0.367 on Level 3, while GB showed similar

high accuracy (0.752) and a higher F1 score (0.493)

on the same level, indicating their robustness in both

correct predictions and balancing precision and re-

call. LightGBM (LGBM) and CatBoost also stand

out, especially in higher levels, with LGBM reach-

ing an accuracy of 0.748 and an F1 score of 0.363

on Level 3, and CatBoost achieving an accuracy of

0.75 and an F1 score of 0.484. This showcases their

efﬁciency in handling complex classiﬁcations effec-

tively. Random Forest also shows solid performance,

particularly with high accuracy (0.746) and a good F1

score (0.523) on Level 7, reﬂecting its reliability in

making accurate predictions while maintaining a bal-

ance between precision and recall. In contrast, mod-

els like K-Nearest Neighbors (KNN) perform moder-

ately well, particularly on Level 7, with an accuracy

of 0.705 and an F1 score of 0.593, indicating reason-

able but not top-tier performance. Bagging and Extra

Trees also provide moderate results, with Extra Trees

achieving an accuracy of 0.746 and an F1 score of

0.405 on Level 3. AdaBoost shows variability in per-

formance, performing better on some levels but gener-

ally lagging behind the top models. Conversely, Naive

Bayes consistently underperforms, with notably low

accuracy (0.282) and an F1 score of 0.242 on Level

3, indicating poor model performance overall. Sim-

ilarly, Ridge Regression shows lower effectiveness,

especially on Level 6, with an accuracy of 0.284

and an F1 score of 0.203. Support Vector Machine

(SVM) and Logistic Regression (LR) exhibit mixed

results, with some moderate performances but gener-

ally lower compared to other models. For instance,

SVM achieves an accuracy of 0.685 and an F1 score

of 0.456 on Level 7, and Logistic Regression shows

moderate performance but less effectiveness in han-

dling complex and detailed occupancy data. Multi-

Layer Perceptron (MLP) also demonstrates competi-

tive performance, achieving strong results particularly

at higher granularity levels, such as an accuracy of

0.744 and an F1 score of 0.501 on Level 3, and an ac-

curacy of 0.533 and an F1 score of 0.493 on Level 6.

MLP’s neural network architecture allows it to cap-

ture intricate patterns in the data, making it suitable

for applications requiring high precision and reliabil-

ity.

Overall, the superior performance of XGBoost,

Gradient Boosting, LightGBM, CatBoost, and MLP

highlights their capability in handling detailed and

high-accuracy occupancy counting tasks. Random

Forest remains a strong contender, particularly for ap-

plications requiring a balance of accuracy and model

interpretability. In contrast, simpler models like

KNN, Bagging, and Extra Trees provide moderate

performance, suitable for less complex applications.

Naive Bayes and Ridge Regression, along with Lo-

gistic Regression, generally underperform, especially

for more detailed and complex occupancy counting

levels. SVM, while showing better performance than

some simpler models, still lags behind the top per-

formers, indicating its limitations in handling the

most detailed classiﬁcations. The table suggests that

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

400

Figure 6: Feature Correlation Heatmap for different Occupancy Levels.

the choice of model and level signiﬁcantly impacts the

prediction accuracy and balance between precision

and recall, making it crucial to select the right model

and level for predicting occupancy levels in smart

buildings. Additionally, the choice of machine learn-

ing algorithm for people counting in smart buildings

should consider trade-offs between accuracy, com-

plexity, resource requirements, and the speciﬁc needs

of the application. High-accuracy models like Gra-

dient Boosting, XGBoost, LightGBM, CatBoost, and

MLP are ideal for detailed tasks but require more

computational resources. SVM offers high accuracy

with minimal scaling needs but is resource-intensive.

KNN requires signiﬁcant computational resources for

large datasets. Random Forest and Deep Learning

models excel in accuracy and resilience but need sub-

stantial data and computational power. Simpler mod-

els like Naive Bayes and Ridge Regression are suit-

able for less complex scenarios where quick deploy-

ment and interpretability are prioritized. In addi-

tion, the performance across different levels reveals

that levels with fewer classes, such as Level 7, gen-

erally see better accuracy and F1 scores. Level 7,

which has the fewest classes (4), and Level 3, with

7 classes, achieve better and similar accuracy across

most models, indicating their effectiveness for occu-

pancy counting tasks. Level 7 beneﬁts from broader

intervals that capture fewer classes, leading to higher

accuracy, while Level 3 also performs well with its

speciﬁc interval distribution. Following these, Level

5, with 5 classes, balances detail and simplicity ef-

fectively, making it suitable for moderately detailed

occupancy counting. Level 4, with 6 classes, pro-

vides detailed counting without being overly com-

plex. Level 2, despite having 7 classes, shows reason-

able accuracy but loses effectiveness as the intervals

become broader. Finally, Level 6, designed to han-

dle imbalanced data with 5 classes, tends to perform

less effectively due to its simpliﬁed classiﬁcation and

broader intervals.

The choice of intervals is crucial in people count-

ing, as shown by the distribution of people counts in

Figure 5. The frequency distribution shows a high

number of lower counts (0-3 people), which decreases

as the number of people increases. This pattern high-

lights the importance of selecting appropriate inter-

vals for each level so that the models can effectively

distinguish between different occupancy levels. Well-

chosen intervals, like those in Levels 7, 3, and 5, cap-

ture detailed occupancy patterns more accurately, re-

sulting in higher precision. On the other hand, poorly

chosen intervals, such as those in Level 6, simplify the

task but may miss ﬁner distinctions, affecting overall

performance. Levels 4 and 2 offer a balance, provid-

ing reasonable accuracy without being overly com-

plex. This supports the observation that the distribu-

tion of people counts can impact results if not appro-

priately managed by choosing the right intervals and

number of classes. Moreover, the choice of intervals

may depend on the speciﬁc application, such as secu-

rity systems, energy management, and space alloca-

tion systems like Nimway. For security systems, ﬁner

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors

401

Table 2: Comparison of Accuracy and F1 Macro Scores for Various Models.

Model Level1 Level2 Level3 Level4 Level5 Level6 Level7

Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1

SVM 0.285 0.107 0.474 0.180 0.684 0.289 0.526 0.299 0.597 0.306 0.395 0.329 0.685 0.456

LR 0.322 0.127 0.431 0.144 0.653 0.259 0.488 0.245 0.552 0.248 0.414 0.343 0.650 0.362

Ridge 0.280 0.062 0.467 0.110 0.691 0.243 0.467 0.233 0.598 0.225 0.284 0.203 0.685 0.310

KNN 0.349 0.207 0.510 0.356 0.701 0.427 0.554 0.440 0.616 0.488 0.461 0.427 0.705 0.593

GB 0.381 0.234 0.558 0.406 0.752 0.493 0.609 0.464 0.664 0.520 0.523 0.503 0.748 0.592

AdaBoost 0.332 0.157 0.478 0.279 0.461 0.343 0.549 0.371 0.628 0.418 0.498 0.409 0.706 0.530

Bagging 0.354 0.209 0.512 0.329 0.717 0.401 0.560 0.343 0.623 0.429 0.474 0.401 0.717 0.540

Extra Trees 0.365 0.196 0.534 0.306 0.746 0.405 0.584 0.378 0.654 0.426 0.496 0.405 0.744 0.514

Decision Tree 0.309 0.179 0.451 0.330 0.654 0.343 0.499 0.333 0.564 0.330 0.430 0.367 0.654 0.303

Random Forest 0.379 0.219 0.545 0.341 0.746 0.350 0.596 0.411 0.654 0.451 0.511 0.428 0.746 0.523

Naive Bayes 0.187 0.111 0.213 0.176 0.282 0.242 0.236 0.243 0.247 0.282 0.228 0.242 0.287 0.325

LGBM 0.382 0.215 0.554 0.537 0.748 0.363 0.611 0.474 0.660 0.537 0.521 0.363 0.750 0.537

CatBoost 0.395 0.258 0.565 0.425 0.750 0.484 0.619 0.462 0.665 0.522 0.527 0.484 0.753 0.612

XGBoost 0.386 0.226 0.558 0.364 0.753 0.367 0.613 0.421 0.661 0.500 0.517 0.367 0.756 0.612

MLP 0.398 0.235 0.563 0.394 0.744 0.501 0.618 0.504 0.672 0.611 0.533 0.493 0.730 0.624

intervals with more classes may be necessary to accu-

rately monitor and respond to changes in occupancy.

In contrast, energy management systems may beneﬁt

from broader intervals that simplify classiﬁcation and

focus on larger occupancy trends to optimize HVAC

systems, leading to signiﬁcant cost savings. Regard-

ing space allocation systems like Nimway, an accu-

racy level of around 70-75% for people counting is

sufﬁcient for ensuring efﬁcient room utilization and

reducing the cost of unused spaces. This accuracy

aids in better planning and enhances workplace pro-

ductivity and satisfaction. By emphasizing the practi-

cality and applicability of these models in real-world

scenarios, our study demonstrates that even moderate

accuracy levels can yield signiﬁcant beneﬁts in smart

building management, validating the effectiveness of

machine learning models in this domain.

To ensure robust evaluation of our models, we

employed K-Fold, Stratiﬁed K-Fold cross-validation

methods, and random splitting (handout) for divid-

ing the test and train data. The results, shown in Ta-

ble 3 for Level 7, indicate that K-Fold and Stratiﬁed

K-Fold yield very similar performance metrics, with

nearly identical accuracy and F1 Macro scores. This

suggests that the data distribution is balanced and the

models are robust. Additionally, the handout method

results, as shown in Table 2, are close to those from K-

Fold and Stratiﬁed K-Fold cross-validation, conﬁrm-

ing that all three methods provide comparable and re-

liable estimates of model performance.

Regarding feature importance, as shown in Fig-

ure 7 we have chosen ﬁve algorithms that perform

better. The feature importance scores for these mod-

els, including Gradient Boosting, AdaBoost, Ex-

tra Trees Classiﬁer, Decision Tree, Random Forest,

and XGBoost Classiﬁer, indicate the signiﬁcance of

each feature in predicting occupancy levels. ”pas-

sage event before” is the most critical feature for all

of them, highlighting its crucial role in determin-

ing the number of people based on previous pas-

sage events. ”event count” is also notably important

across models, especially in XGBoost. Other sig-

niﬁcant features include ”duration,” which measures

the time between events and shows high importance

in Extra Trees and Decision Tree models, and ”den-

sity,” reﬂecting the event concentration, which is im-

portant in Random Forest and Decision Tree models.

Lesser but still relevant features like ”area,” ”size,”

and ”length” provide spatial context with varying im-

portance across models. Overall, temporal and event-

based features are paramount for accurate occupancy

predictions, while spatial and categorical features also

contribute to the models’ predictive power. These

insights guide the reﬁnement of models by focusing

on the most impactful features, ensuring improved

performance in smart building management applica-

tions.It’s important to note that a feature’s low impor-

tance in the correlation heatmap but high importance

in the feature importance plot, such as ”event count,”

indicates that the feature contributes to complex, non-

linear relationships or interactions with other features

that the model uses to improve its predictions.

Comparing this work with previous research in

smart building people counting reveals signiﬁcant ad-

vancements. Historically, many studies have relied

on sensor fusion, often raising privacy concerns due

to their invasive nature.In contrast, our approach uti-

lizes only two non-invasive, simple binary PIR sen-

sors: a passage sensor located at the entrance and

a motion sensor inside the room, both of which are

PIR sensors with different ﬁelds of view. Unlike pre-

vious efforts that used two passage binary PIR sen-

sors outside and inside the room, triggering when

someone goes outside and decreasing the count, this

work integrates both passage and motion sensors to

count multiple people inside the room. Additionally,

we enhance the system’s capabilities by introducing

multiple levels of occupancy counting and employing

advanced machine learning algorithms like Gradient

Boosting, XGBoost, and LightGBM, which have not

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

402

Table 3: Comparison of Accuracy and F1 Macro K-Fold and Stratiﬁed K-Fold Cross-Validation Methods.

Metric / Model SVM LR RC KNN GB AdaBoost Bagging Extra Trees DT RF NB LGBM CatBoost XGBoost MLP

Accuracy (KFold) 0.657 0.615 0.546 0.717 0.719 0.657 0.699 0.731 0.644 0.722 0.296 0.734 0.739 0.734 0.719

Accuracy (StratiﬁedKFold) 0.659 0.615 0.546 0.718 0.719 0.655 0.700 0.731 0.642 0.720 0.295 0.733 0.738 0.734 0.720

F1 Macro (KFold) 0.548 0.526 0.312 0.686 0.683 0.625 0.662 0.698 0.612 0.687 0.289 0.702 0.708 0.702 0.684

F1 Macro (StratiﬁedKFold) 0.554 0.526 0.312 0.686 0.683 0.626 0.663 0.698 0.609 0.686 0.289 0.702 0.707 0.702 0.682

Figure 7: Features important.

been widely used in prior studies. This innovative

combination not only respects user privacy but also

boosts the accuracy and adaptability of the occupancy

detection system, offering a more comprehensive so-

lution to smart building management.

5 CONCLUSION

This study presented a novel approach to people

counting in smart buildings, utilizing a minimal

setup of binary passive infrared sensors integrated

with advanced machine learning techniques. Our

method, focused on two strategically placed sensors

per room, overcomes traditional challenges associ-

ated with complex installations and high hardware

costs. By analyzing data collected over three years

from seven different ofﬁce rooms, we demonstrated

that even simple binary PIR sensors, when coupled

with sophisticated machine learning algorithms like

Gradient Boosting, XGBoost, and LightGBM, can

yield accurate and reliable occupancy estimations.

The machine learning models employed successfully

interpreted the sensor data to predict the number of

people with acceptable accuracy. Our results con-

ﬁrmed that fewer sensors can still provide reliable oc-

cupancy data, provided that the data is processed with

effective machine learning strategies. The feature im-

portance analysis highlighted the signiﬁcant roles of

passage events and motion detection patterns in en-

hancing the accuracy of people counting. Moreover,

by categorizing occupancy into multiple levels, we

further improved the accuracy and robustness of our

models, making them well-suited for various appli-

cations in smart building management. Furthermore,

our approach emphasizes scalability, ﬂexibility, and

privacy, making it adaptable to various room conﬁg-

urations and sizes. This adaptability, along with the

reduced cost and installation simplicity, makes our

method a practical solution for real-world applica-

tions in smart building management. Future research

explores several promising directions. Integrating

non-invasive sensors, such as CO2 or sound sensors,

with booking data offers an opportunity to enhance

people counting accuracy by combining real-time en-

vironmental insights with contextual usage patterns.

Employing real-time data processing techniques en-

ables dynamic adaptation to changing environments

and user behaviors, while large-scale validation stud-

ies across diverse building types ensure broader ap-

plicability and reliability. Additionally, hybrid ap-

proaches that combine binary PIR sensors with low-

resolution vision-based systems balance privacy, cost,

and accuracy. By addressing these directions, future

work advances the capabilities of smart building tech-

nologies, delivering enhanced energy efﬁciency, opti-

mized space utilization, and improved occupant com-

fort.

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors

403

ACKNOWLEDGEMENTS

This work is supported by the Knowledge Founda-

tion (Stiftelsen f

or kunskaps- och kompetensutveck-

ling) for the project titled Intelligent and Trustworthy

IoT Systems under Grant No. 20220087-H-01.

REFERENCES

Alhawsawi, A. N., Khan, S. D., and Ur Rehman, F. (2024).

Crowd counting in diverse environments using a deep

routing mechanism informed by crowd density levels.

Information, 15(5):275.

Alishahi, N., Ouf, M. M., and Nik-Bakht, M. (2022). Us-

ing wiﬁ connection counts and camera-based occu-

pancy counts to estimate and predict building occu-

pancy. Energy and Buildings, 257:111759.

Alsafery, W., Rana, O., and Perera, C. (2023). Sensing

within smart buildings: A survey. ACM Computing

Surveys, 55(13s):1–35.

Amayri, M., Arora, A., Ploix, S., Bandhyopadyay, S., Ngo,

Q.-D., and Badarla, V. R. (2016). Estimating occu-

pancy in heterogeneous sensor environment. Energy

and Buildings, 129:46–58.

Brazauskas, J., Jensen, C., Danish, M., Lewis, I., and

Mortier, R. (2024). Cerberus: Privacy-preserving

crowd counting and localisation using face detection

in edge devices. In Proceedings of the 7th Interna-

tional Workshop on Edge Systems, Analytics and Net-

working, pages 25–30.

Chaudhari, P., Xiao, Y., Cheng, M. M.-C., and Li, T. (2024).

Fundamentals, algorithms, and technologies of occu-

pancy detection for smart buildings using iot sensors.

Sensors, 24(7):2123.

Corporation, S. (2024). Nimway smart ofﬁce by sony. Ac-

cessed: 2024-12-30.

Dobrilovic, D., Bogdan, R., Ognjenovic, V., and Marcu, M.

(2023). Analyses on usage of mlp regression with wsn

data for predicting room occupancy. In 2023 IEEE

19th International Conference on Intelligent Com-

puter Communication and Processing (ICCP), pages

131–136. IEEE.

Gao, M., Souri, A., Zaker, M., Zhai, W., Guo, X., and Li, Q.

(2024). A comprehensive analysis for crowd counting

methodologies and algorithms in internet of things.

Cluster Computing, 27(1):859–873.

Hitiyise, E., Ntagwirumugara, E., Habarurema, W.,

Ingabire, W., and Gasore, G. (2016). Building oc-

cupancy monitoring based on microcontroller and pir

sensors. Int. J. Appl. Eng. Res, 11:10414–10419.

Hobson, B. W., Lowcay, D., Gunay, H. B., Ashouri, A., and

Newsham, G. R. (2019). Opportunistic occupancy-

count estimation using sensor fusion: A case study.

Building and environment, 159:106154.

Jamali, M., Davidsson, P., Khoshkangini, R., Ljungqvist,

M. G., and Mihailescu, R.-C. (2024). Specialized in-

door and outdoor scene-speciﬁc object detection mod-

els. In Sixteenth International Conference on Machine

Vision (ICMV 2023), volume 13072, pages 201–210.

SPIE.

Jiang, C., Masood, M. K., Soh, Y. C., and Li, H. (2016). In-

door occupancy estimation from carbon dioxide con-

centration. Energy and Buildings, 131:132–141.

Kim, J., Bang, J., Choi, A., Moon, H. J., and Sung, M.

(2023). Estimation of occupancy using iot sensors and

a carbon dioxide-based machine learning model with

ventilation system and differential pressure data. Sen-

sors, 23(2):585.

Kumari, P., Reddy, S., and Yadav, R. (2024). Indoor occu-

pancy detection and counting system based on boost-

ing algorithm using different sensor data. Building

Research & Information, 52(1-2):87–106.

Li, T., Liu, X., Li, G., Wang, X., Ma, J., Xu, C., and Mao,

Q. (2024). A systematic review and comprehensive

analysis of building occupancy prediction. Renewable

and Sustainable Energy Reviews, 193:114284.

Liang, X., Shim, J., Anderton, O., and Song, D. (2024).

Low-cost data-driven estimation of indoor occupancy

based on carbon dioxide (co2) concentration: A multi-

scenario case study. Journal of Building Engineering,

82:108180.

Liu, Z., Guo, Z., Chen, Q., Song, C., Shang, W., Yuan, M.,

and Zhang, H. (2023). A review of data-driven smart

building-integrated photovoltaic systems: Challenges

and objectives. Energy, 263:126082.

Lu, C. (2024). Enhancing real-time nonintrusive occupancy

estimation in buildings via knowledge fusion network.

Energy and Buildings, 303:113812.

Mao, S., Yuan, Y., Li, Y., Wang, Z., Yao, Y., and Kang,

Y. (2023). Room occupancy prediction: Exploring

the power of machine learning and temporal insights.

arXiv preprint arXiv:2312.14426.

Masciadri, A., Lin, C., Comai, S., and Salice, F. (2022).

A multi-resident number estimation method for smart

homes. Sensors, 22(13):4823.

Naser, A., Lotﬁ, A., and Zhong, J. (2020). Adaptive ther-

mal sensor array placement for human segmentation

and occupancy estimation. IEEE Sensors Journal,

21(2):1993–2002.

Natarajan, K. S., Balu, S., and Mangottiri, V. (2023). Smart

and sustainable infrastructure for future energy and

environmental management. Environmental Science

and Pollution Research, 30(44):98993–98994.

Navarro, R. C., Ruiz, A. R., Molina, F. J. V., Romero, M.

J. S., Chaparro, J. D., Alises, D. V., and Lopez, J. C. L.

(2022). Indoor occupancy estimation for smart utili-

ties: A novel approach based on depth sensors. Build-

ing and Environment, 222:109406.

Ngamakeur, K., Yongchareon, S., Yu, J., and Islam, S.

(2023). Passive infrared sensor dataset and deep

learning models for device-free indoor localization

and tracking. Pervasive and Mobile Computing,

88:101721.

Rajabi, H., Ding, X., Du, W., and Cerpa, A. (2023). To-

dos: Thermal sensor data-driven occupancy estima-

tion system for smart buildings. In Proceedings of

the 10th ACM international conference on systems for

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

404

energy-efﬁcient buildings, cities, and transportation,

pages 198–207.

Ramzan, F., Sartori, C., Consoli, S., and Reforgiato Recu-

pero, D. (2024). Generative adversarial networks for

synthetic data generation in ﬁnance: Evaluating statis-

tical similarities and quality assessment. AI, 5(2):667–

685.

Raykov, Y. P., Ozer, E., Dasika, G., Boukouvalas, A., and

Little, M. A. (2016). Predicting room occupancy with

a single passive infrared (pir) sensor through behavior

extraction. In Proceedings of the 2016 ACM inter-

national joint conference on pervasive and ubiquitous

computing, pages 1016–1027.

Risuleo, R. S., Molinari, M., Bottegal, G., Hjalmarsson, H.,

and Johansson, K. H. (2015). A benchmark for data-

based ofﬁce modeling: challenges related to co2 dy-

namics. IFAC-PapersOnLine, 48(28):1256–1261.

Sahoo, S. R. and Lone, H. R. (2023). Occupancy count-

ing in dense and sparse settings with a low-cost ther-

mal camera. In 2023 15th International Conference

on COMmunication Systems & NETworkS (COM-

SNETS), pages 537–544. IEEE.

Shokrollahi, A., Persson, J. A., Malekian, R., Sarkheyli-

agele, A., and Karlsson, F. (2024). Passive infrared

sensor-based occupancy monitoring in smart build-

ings: A review of methodologies and machine learn-

ing approaches. Sensors, 24(5):1533.

Tomokazu, T. Cooperative distributed demand control by

environmental sensor network-estimating the number

of people by co2 concentration. In The IEEE Interna-

tional Conference on Industrial Informatics (INDIN

2008).

Tsou, P.-R., Wu, C.-E., Chen, Y.-R., Ho, Y.-T., Chang, J.-

K., and Tsai, H.-P. (2020). Counting people by using

convolutional neural network and a pir array. In 2020

21st IEEE International Conference on Mobile Data

Management (MDM), pages 342–347. IEEE.

Udrea, I., Simion, N. D., Alionte, C. G., Ionut, V., Ghe-

orghe, V. F. K., and Petrache, S. (2022). Counting

versus detection in an fm application that deals with

rooms reservation. Journal of Eastern Europe Re-

search in Business and Economics, Norristown, PA,

USA.

Umutoni, R. M., Ogore, M., Hanyurwimfura, D., and

Nsenga, J. (2023). Integrating analog pir sen-

sor telemetry with tinyml inference for on-the-edge

classiﬁcation of moving objects. In International

Congress on Information and Communication Tech-

nology, pages 405–415. Springer.

Wahl, F., Milenkovic, M., and Amft, O. (2012). A

distributed pir-based approach for estimating people

count in ofﬁce environments. In 2012 IEEE 15th

International Conference on Computational Science

and Engineering, pages 640–647. IEEE.

Wang, K., Wang, Y., Ren, R., Zou, H., and Shao, Z. (2023).

Transformer-based feature aggregation and stitching

network for crowd counting. IEEE Access.

Wang, T. and Cook, D. J. (2020). smrt: Multi-resident track-

ing in smart homes with sensor vectorization. IEEE

transactions on pattern analysis and machine intelli-

gence, 43(8):2809–2821.

Wang, W., Chen, J., and Hong, T. (2018). Occupancy pre-

diction through machine learning and data fusion of

environmental sensing and wi-ﬁ sensing in buildings.

Automation in Construction, 94:233–243.

Yang, T., Guo, P., Liu, W., Liu, X., and Hao, T. (2020). En-

hancing pir-based multi-person localization through

combining deep learning with domain knowledge.

IEEE Sensors Journal, 21(4):4874–4886.

Yun, J., Kim, D., Kim, D. M., Song, T., and Woo, J. (2023).

Gan-based sensor data augmentation: Application for

counting moving people and detecting directions us-

ing pir sensors. Engineering Applications of Artiﬁcial

Intelligence, 117:105508.

Zhai, W., Gao, M., Guo, X., Zou, G., Li, Q., and Jeon, G.

(2024). Scale attentive aggregation network for crowd

counting and localization in smart city. ACM Trans-

actions on Sensor Networks.

Zhang, X., Zhou, T., Kokogiannakis, G., Xia, L., and Wang,

C. (2023). Estimating the number of occupants and

activity intensity in large spaces with environmental

sensors. Building and Environment, 243:110714.

Zhao, L., Li, Y., Liang, R., and Wang, P. (2022). A state

of art review on methodologies of occupancy esti-

mating in buildings from 2011 to 2021. Electronics,

11(19):3173.

Non-Invasive People Counting in Smart Buildings: Employing Machine Learning with Binary PIR Sensors

405