Application of AI-based Image Processing for Occupancy Monitoring

in Building Energy Management

Andreas Reichel

, Jens Döge

, Dirk Mayer

and Jan Bräunig

Fraunhofer Institute of Integrated Circuits IIS, Division Engineering of Adaptive Systems EAS,

Münchner Str. 16, 01187 Dresden, Germany

Keywords: Vision-based User Recognition, Machine Learning, Building Energy System.

Abstract: Smart Buildings enable significant savings in energy and CO2 emissions by model-predictive methods. The

building users have a considerable influence on the energetic building management. On the one hand, they

dictate the comfort parameters to be set. On the other hand, they generate internal thermal gains through their

presence, affect humidity, consume oxygen and produce carbon dioxide. The more precisely the user behavior

is known, the more precisely and resource-efficiently the room climate control can be adapted to this user

behavior. In this paper, an intelligent vision-based sensor concept is proposed and tested that is capable to

estimate occupancy and activity inside a building. The contribution initially concentrates on functional

buildings, since here, compared to residential buildings, there is an even greater need for use-oriented room

air conditioning, including savings potential.

1 INTRODUCTION

Buildings, and particularly their energy systems, are a

significant contributor to the world-wide CO

emissions. Besides conventional measures like

enhanced insulation of walls or windows, digital

technologies are an enabler towards greener buildings.

These smart buildings integrate data from external

sources (e.g. weather forecasts) and sensors with

powerful analytics and control systems to optimize

energy consumption. One of the most important

parameters is occupancy.

Traditionally, building automation systems use

sensors to measure temperature, humidity and CO

concentration in order to draw conclusions about the

indoor climate. Such systems usually provide very

inaccurate data on user behaviour, which makes user-

adapted energy-optimised building control with

advanced control strategies hardly feasible.

However, improvements in detection of user

behaviour are still under development. Two questions

need to be answered: technical possibilities and data-

law constraints.

https://orcid.org/0000-0002-3971-1585

https://orcid.org/0000-0002-2891-984X

https://orcid.org/0000-0002-4972-6529

https://orcid.org/0000-0002-7282-723X

This paper aims to provide an overview of

integrating a vision-based system into a smart building

data infrastructure.

To this end, general system architectures for smart

buildings are considered, with a focus on the

distribution of data analysis to the edge of the network.

Potential scenarios for the integration of a smart

sensor for occupancy detection into building

automation are discussed. Finally, the set-up and first

results from a feasibility study at a real building are

presented and directions for further research are given.

2 CURENT STATE OF THE ART

2.1 Smart Building Technologies and

System Architectures

Smart buildings integrate digital building control

systems and networked building automation. Using

multiple, distributed sensors for the building energy

system enables advanced building control schemes. In

Reichel, A., Döge, J., Mayer, D. and Bräunig, J.

Application of AI-based Image Processing for Occupancy Monitoring in Building Energy Management.

DOI: 10.5220/0011080600003203

In Proceedings of the 11th International Conference on Smar t Cities and Green ICT Systems (SMARTGREENS 2022), pages 139-146

ISBN: 978-989-758-572-2; ISSN: 2184-4968

139

contrast to preset schedules, the energy system can be

adapted to physical parameters, weather forecast or

occupancy. This can lead to significant energy savings

of 24% (King, 2017). The gathered sensor information

can also be applied to fault detection and diagnosis.

In summary, the main hardware components of a

smart building are:

 Sensors to acquire physical parameters

 An internal network to connect smart

devices and control systems as well as an

internet connection

 Centralized or decentralized computing

capabilities

This infrastructure is complemented by analytics

and control algorithms to extract relevant information

from sensor data, classify the state of components,

predict time series for relevant parameters and, last

but not least, implement control actions. Obviously,

these applications are very well-suited for machine

learning algorithms.

Findings of a recent review study show that most

control architectures are cloud-based (Yaïci, 2021).

This enables using powerful machine learning

algorithms, e.g. deep learning, that require high

efforts for training. However, such an approach needs

a reliable data connection from the building to the

cloud; also, data privacy might be a concern.

An alternative to this is shifting parts of the data

analytics and control to the edge of the network

(Figure 1). Due to computing power and a network

connection being available in nearly every building

automation device, this enables a variety of system

architectures. In most cases, data analytics and

control functions are implemented on a unit that also

serves as a gateway between the sensor-actuator-

network and the internet (Curto Fuentes, 2021). In a

Figure 1: System architecture for deployment of AI to the

edge and the sensor.

building automation system, this would be the central

building automation control platform (Rinaldi, 2020).

Such an approach has several advantages. The

distribution of computing power enhances the

scalability of the system. The control system is more

robust, which is particularly relevant in case of

extreme weather events: If the connection to the cloud

is lost, external information (e.g. weather forecast) is

missing, but the control system is still operating.

Furthermore, the building owner is not forced to share

raw data with a remote cloud server and can

implement higher privacy standards.

Still, smart functions can be shifted further

towards the edge of the network by using smart

sensors (Figure 1). The concept of a smart sensor is

well known for three decades (Najafi, 1991). Their

main characteristics are:

 Analog preprocessing of the input signals

(amplification, filtering)

 Analog to digital conversion

 Bidirectional communication possibility

 Protocol based communication interface and

a unique identification (network address)

 Possibility to implement data analysis

algorithms, e.g. compressed sensing, auto-

calibration and self-diagnosis

For building automation applications, the digital

networking capabilities alleviate the installation of a

scalable sensor network that can be spread also on

larger facilities. Compression of the acquired signals

can be useful, if a low-bandwidth wireless network is

used and a large number of sensors have to be

connected.

In certain applications, the sensor can also be shut

down into sleep mode, enabling long-term operation

on batteries or even energy harvesting. This is

especially relevant when the sensors are wireless,

which is an attractive solution for retrofitting of

existing buildings.

There are several potential energy sources in

buildings to power energy harvesting systems:

Floor vibrations can transmit power to

piezoelectric oscillators, that also can be used as a

sensor measure to detect occupancy behavior (Jung,

2018). Also, indoor solar panels have been

investigated for powering wireless sensors, e.g.

temperature and light sensors (Fraternali, 2018). Last

but not least, radio frequency (RF) energy harvesting

should be mentioned. (Bjorkqvist, 2018) developed a

harvester that is able to generate power in the micro-

Watt range from surrounding Wifi and cellular

networks in a building.

SMARTGREENS 2022 - 11th International Conference on Smart Cities and Green ICT Systems

140

Usually, sensor platforms detect a variety of

physical parameters. For instance, the BiB (Building

in Briefcase) platform presented by Weekly (Weekly,

2018) integrates temperature, humidity, ambient

light, CO2, infrared motion detection, and inertial

measurement units in a wireless platform. Another

example for a wireless multi-sensor platform was

developed and tested by De Donno et al. (De Donno

2018), acquiring humidity, temperature, CO

, air

quality, and occupancy by passive infrared sensors.

Yet, power measurements showed that the device

would operate just about 1 month until a battery

replacement is necessary, i.e. a significant

maintenance effort would be necessary.

Especially for the case of occupancy detection or

other perhaps sensitive personal data, preprocessing

directly at the sensor avoids transmission of raw data

to remote devices, thus increasing the privacy of the

acquired data.

2.2 Inclusion of User Behavior in

Building Automation

For some years there has been an effort to use model

predictive control strategies (MPC) in the field of

building automation, which can save up to 40% of the

energy used compared to conventional control

(Serale, 2018). These include, for example, model-

predictive zone and individual room controllers

(Paschke, 2018; Kelman, 2011; Gwender, 2010), the

design of which, however, requires knowledge of a

prediction model for the air comfort parameters. This

can be generated automatically using recorded

measurement data (Paschke, 2019), but the number of

people present in a room remains unknown. Since air

comfort parameters are significantly influenced by

the number of people, measuring room occupancy is

needed for model identification and thus for the

design of an MPC control strategy.

The recording of high-resolution measurement

data from building occupancy as well as control and

operation of technical systems with the help of

decentralized sensor technology has become an

important part of building control technology and the

planning process in the last 10 years due to the

increasing spread of bus systems and IoT applications

in buildings. In this way, engineers are given a tool to

cope with the increasing complexity of efficient

energy supply systems. In this context, the

implementation of multi-year monitoring campaigns

for system optimization after commissioning has

become established in selective research and

innovation projects.

2.3 Determination of User Behavior

Several options exist to determine the occupancy and

user behavior in-situ as an input to model predictive

building control systems.

A well-established technology are CO

sensors, that

directly acquire the quality of the air inside a room as

an important control input to the ventilation.

A major difficulty here is that measurements are

usually taken at only one or very few individual

points. This is particularly problematic in large rooms

(lecture halls, gymnasiums...), as very strong local

differences can occur, which can negatively influence

the functionality of the control system and thus the

energy consumption of the building if the sensor is

placed in an unfavorable position. Further

disadvantages of conventional CO

sensor technology

include:

 No direct statement about the number of

people present.

 Delayed/indirect measurement of human

activity

 Relatively high susceptibility to interference

 High calibration effort for CO

sensor

technology, comparably strong sensor drift

However, intelligent sensing can enhance the

performance of such systems: The number of

occupants can be estimated using an observer that

processes the current CO

concentration by a physical

model (Jin, 2015). Another approach is the fusion of

a CO

and a light sensor (Huang, 2017).

Besides CO

, also other modalities are suitable to

measure occupancy. (Ahmad, 2020) gives a review

on several technologies, including passive infrared

sensors, ultrasound or wireless network connections.

As demonstrated in (Naylor, 2017), integration of

with other sensor modalities like Wi-Fi tracking

by machine learning based data fusion can drastically

increase the precision of the occupancy estimation.

The classic optical approach is based on simple

presence or activity sensing using passive IR sensors

(PIR, 2020). Here, one to three individual point

sensors per sensor head with upstream Fresnel optics

detect the movement of people through the room

based on the change in their thermal signatures. This

approach can be used very well to activate light in

small rooms, but even transferring it to larger rooms

is not possible due to the limited resolution. Likewise,

people who do not move "disappear", as they are no

longer detected by the system without a change in

their whereabouts in the room.

Thus, this sensor technology is very inaccurate,

especially in large rooms, and is neither able to

provide information about the actual number of

Application of AI-based Image Processing for Occupancy Monitoring in Building Energy Management

141

people in a certain area of the building, nor

information about their level of activity. It is also not

possible to reliably detect the absence of people, so a

predefined period of inactivity in the room is usually

used as an equivalent.

Alternatively, there are surveillance cameras in

security-critical building sections that transmit or

record real images and thus technically provide the

possibility to also derive information about the actual

number of people and their activity. However, the

computer-aided storage and automatic analysis of

image data must usually be carried out in compliance

with strict data protection guidelines.

Doppler radar-based solutions (Radar, 2020) try

to circumvent this disadvantage, as they react much

more sensitively to movements. Since these sensors

also only provide integral information about larger

areas, they are also not suitable for detecting specific

numbers of people.

According to the state of the art, the use of

systems based on area scan cameras is currently the

only possibility to solve the resolution problem,

however, especially in connection with the automatic

evaluation of the image data, it rightly places very

high demands on data protection. Thus, the output

and storage of real image data and their external

processing is a fundamental problem that is difficult

to solve technically and, moreover, also has to

contend with considerable acceptance problems

among users.

In certain areas with increased security

requirements, such as the gate environment of

airports, autonomous 3D camera systems based on a

stereo arrangement (Stereo 2020) are sometimes used

to count and track people. These systems usually have

a very high accuracy, but can only be used in quite

small areas - usually up to 5m x 5m at typical room

heights - and have to be installed and set up in a

complex way. Due to the two camera heads required

for 3D reconstruction and the high computing effort,

these systems are also quite expensive. The data

protection aspect is the same as with classic camera

solutions.

To prevent using raw image data, that can include

private information, encryption methods have been

proposed that shuffle the image data in the regions

containing persons (Ahmad 2020). Yet, such

algorithms should be implemented directly at the

pixel sensor. This way, an intelligent sensor is

realized that does not transmit privacy-relevant

information.

3 SOLUTION APPROACH AND

FIRST FEASIBILITY STUDY

For the real-world demonstrator the lecture hall center

“Bergstraße” of the TU Dresden was chosen and

equipped. First, in a standard lecture hall, user

behaviour can be recorded with sufficient accuracy

over a certain period of time and processed with

regard to the activity level.

As a key component for the vision-based

occupancy sensor, the vision system-on-chip (VSoC)

depicted in Figure 2 has been used (Döge, 2015).

Figure 2: Chip microphotograph of the Vision System-on-

Chip (VSoC).

This VSoC is capable to analyze image data

directly on the sensor chip, extract information that

can no longer be personalized and output only the

features required for further processing. Based on

this, the embedded image processing system performs

the next sensor-external processing steps only on pre-

selected image descriptors - such as specific gradient

and texture features or local activity levels, and

calculate the approximate number of people in

regions of increased activity. The implemented

activity and occupancy measurement system thus

takes special account of the data protection required

for public buildings.

The auditorium observed is shown in Figure 3 top.

Due to its size and geometry, it has to be observed

from two overlapping perspectives (see Figure 3

bottom) in order to achieve a complete coverage of

the room with the used optics on the one hand and on

the other hand to avoid that the effective subject size

in pixels varies too much out of the camera

perspective (see Figure 4 right). By choosing a height

of the vision system's viewpoint of about 5m and

splitting it into two views for two installed vision

SMARTGREENS 2022 - 11th International Conference on Smart Cities and Green ICT Systems

142

systems, the minimum number of pixels per

monitored seat could be slightly improved. However,

the resolution in the rear areas of the room is still too

low for detection and classification of the head-

shoulder region, which is common according to the

state of the art. Dynamic and static methods were

investigated as a basis for occupancy analysis, but all

of them basically work without object detection,

which is potentially problematic in terms of data

protection.

Figure 3: Photography of the lecture hall (top) positioning

of the vision systems (bottom).

In dynamic scene analysis, it is assumed that

people are generally not completely motionless.

Regardless of the lighting situation, the

corresponding regions are only learned as

background on the assumption of the absence of

movements over a certain period of time. Regions of

a defined dimension, in which deviations from the

current background model are detected, are by

definition foreground and thus people.

In static methods, a large number of images of the

unoccupied space are used to train the classifier. The

training data differ mainly in the possible illumination

situations that have been automatically recorded in

advance. An extension of the training data set by data

augmentation is also possible but was not necessary

in the presented example. The trained model of the

room is then used in operation by detecting deviations

from it as people.

Figure 4: Gradient image from the wall-mounted vision

system (left) and marked seats with derived occupancy data

(right).

As a basis for both approaches, various

descriptors feasible on the VSoC were evaluated,

such as:

 gradients and angles,

 histograms based thereon (Histogram-of-

Oriented-Gradients - HOG), and,

 Local Binary Patterns (LBP).

The static approach using principal component

analysis (PCA) based on gradient data proved to be

the most robust to the confounding effects of

illumination. For this purpose, a static background

model is initially trained by dividing a set of training

gradient images into rectangles and, for each

rectangle, calculating the first five principal

components. Using more than five components did

not improve performance any further. At runtime,

each rectangle’s gray-values are projected onto its

respective principal components and back into image

space. An activity map is then obtained from the

absolute difference between this reconstructed image

and the pre-trained background model. Binary

thresholding and morphological operations are used

to close holes and suppress false positives. Including

angular information provided very little added value,

but introduced a significant noise contribution due to

uncertainty in low-contrast areas.

Application of AI-based Image Processing for Occupancy Monitoring in Building Energy Management

143

Figure 5: Detected occupancy levels and estimated error.

The determined active regions were then

converted into an occupancy level using the spatial

resolution illustrated by the size of the seats in Figure

4 (right) as an example. Although this approach is

more of an estimate than a precise count, especially

when the room is very crowded with clusters of

people connected or overlapping, it does provide a

reliable indication that is considerably more accurate

than required in terms of the air turnover and thermal

load of the people present. In particular, considering

random samples at low occupancy levels (less than 10

people), the estimated occupancy differs from the true

occupancy by up to 20–25%. Empty rooms were

consistently detected as empty, though. At higher

occupancy levels (50–100 people), the relative

estimation error was consistently below 5%. For

reference, an example sequence of detected

occupancy and roughly estimated error is depicted in

Figure 5. It can be clearly observed that students start

trickling into the lecture hall at around 04:27 pm, with

an even larger group entering at around 04:35 pm. As

expected, the estimated number of people fluctuates

only mildly until 06:10 pm, when lecture time ends

and students start leaving the lecture hall rather

hastily. A more in-depth evaluation would have

required either acquiring high-resolution gray-value

image data instead of gradient data or attending

events in the lecture hall in-person. The former was

discarded due to data protection concerns raised by

the proprietor of the building. The latter was not

possible because most events taking place at the time

were examinations with strict access control.

The vision-based occupancy sensor system and

the building management system operate without

feedback and independently of each other. The

occupancy percentage determined in the building area

being monitored is assigned to the available building

operating data and statements can then be derived on

comfort and energy optimization of the operational

management. The aerosol load in the rooms is also to

be reduced by adjusting the implemented air volume.

Machine learning algorithms are used for this

optimization task.

4 PERSPECTIVES

The operator of a single functional building or city

district with very heterogenous building usage (e.g.

lecture hall, classrooms and seminar rooms, sports

hall, shopping center, exhibition hall) receives

information about when and where exactly how many

people are in a quarter/building area at what activity

level (walking, sitting, physical work, playing sports,

etc.). For each usage scenario, the system provides

statements on optimal building operating parameters

and target specifications that can be applied manually

by the system operator. As a result, the energy

demand (and thus the CO

emissions) for the

building/district area under consideration is reduced

while the required comfort parameters are ensured

through use-specific system operation and

minimisation of aerosol dispersion.

In the future, the system can be expanded in such

a way that, after a pilot phase, manual tracking of the

operating parameters can be rolled out to other

buildings/districts and finally integrated fully

automatically into the building management system.

An OPC-UA interface to commercial building

management systems (BMS) and energy

management systems (EMS) systems, which is

operated together with these systems as a value-added

service, is ideal for this purpose. The operator of a

building gains potentially comprehensive benefits

through such a scenario (Focus on energy and cost

savings):

 Long-term forecast

 model- and user-specific optimisation

 Communication on necessary maintenance

 Increase in system availability through wear

prediction adapted to the usage behaviour

The algorithms implemented at the VSoC level

have been tested in a lecture hall with fixed furniture

positions and limited activity of the occupants. Future

research should include improvements towards

rooms with movable furniture (resulting in varying

background images and greater variations in positions

and activities of the occupants.

In addition to the determination of the occupancy

levels of buildings, the presented vision technology

can also be used in a variety of other applications, e.g.

for the analysis of walkways in stores for the

evaluation of interest in presented advertising

measures or for the demand-driven regulation of

ventilation and air-conditioning of booths at trade

fairs. In particular, the large dynamic range of the

sensor with a linear-logarithmic characteristic makes

it predestined, e.g. in combination with other

SMARTGREENS 2022 - 11th International Conference on Smart Cities and Green ICT Systems

144

wavelength ranges, for outdoor use for movement

sensing in public areas (Reichel, 2019).

Due to the high compression rate of the intelligent

sensor from raw images to the estimated number of

occupants, a wireless connection is a possible

improvement. This would facilitate the retrofit

installation of the sensors at hard-to-reach positions

in existing buildings. However, in turn an energy

management on the sensor platform will have to

implemented to enable a long-term operation on

batteries. Even more ambitious would be the

development of a low power system that could be

supplied by energy harvesting technologies.

5 CONCLUSIONS

In this paper, a concept for using machine learning

methods on sensor level to provide a privacy-oriented

vision-based occupancy detection has been proposed

and tested in a first feasibility study.

The article makes it clear that there is great

potential in the application of vision technology in the

building sector with regard to energy saving.

Furthermore, the approach provides an important

building block for intelligent neighbourhood

monitoring, also with the aim of recording the

utilisation concepts in the neighbourhood and

deriving further savings potential from them.

In the perspective, in addition to CO

-compatible

modes of operation, it is also necessary to find ways

of reducing the aerosol load individually when there

is a high density of people in order to reduce the

potential for viral contagion.

ACKNOWLEDGEMENTS

On the basis of the TGM.plus project (SAB

Sächsische Aufbaubank), among others in

cooperation with the partner of the SIB (Staatsbetrieb

Sächsisches Immobilien- und Baumanagement),

initial results on image processing-based presence

detection could be elaborated: The aim of the project

was to develop a value-added module for existing

building automation systems. Its application enables

a reduction of resource consumption in technical

building operation through model-based forecasts of

energy consumption and the derivation of appropriate

measures. The basic knowledge created there makes

a significant contribution to this paper.

The authors gratefully acknowledge the ENLIGHT

project (http://www.enlight-project.eu/), funded by

ENIAC and the German Federal Ministry of

Education and Research, for the financial support of

the development of the Vision System-on-Chip.

Last not least the authors wish to thank Dr. Jürgen

Haufe, Chief Scientist at EAS, for initiating the work

and acting as a mentor for the research team.

REFERENCES

King, J. & Perry, Christopher. (2017). Smart Buildings:

Using Smart Technology to Save Energy in Existing

Buildings. Report A1701. American Council for an

Energy-Efficient Economy.

Yaïci, W., Krishnamurthy, K., Entchev, E., & Longo, M.

(2021). Recent Advances in Internet of Things (IoT)

Infrastructures for Building Energy Systems: A

Review. Sensors, 21(6), 2152.

https://doi.org/10.3390/s21062152

Rinaldi, S., Bellagente, P., Ciribini, A. L. C., Tagliabue, L.

C., Poli, T., Mainini, A. G., Speroni, A., Blanco

Cadena, J. D., & Lupica Spagnolo, S. (2020). A

Cognitive-Driven Building Renovation for Improving

Energy Efficiency: The Experience of the ELISIR

Project. Electronics, 9(4), 666.

https://doi.org/10.3390/electronics9040666

Curto Fuentes, Monika (Hrsg.). (2021). AI at the Edge 2021

Whitepaper. VDI/VDE-IT. https://www.smart-

systems-integration.org/system/files/document/

EPoSS-White-Paper-AI%20print.pdf

Najafi, K. (1991). Smart sensors. Journal of

Micromechanics and Microengineering, 1(2), 86–102.

https://doi.org/10.1088/0960-1317/1/2/002

Jung, B. C., Huh, Y. C., & Park, J.-W. (2018). A Self-

Powered, Threshold-Based Wireless Sensor for the

Detection of Floor Vibrations. Sensors, 18(12), 4276.

https://doi.org/10.3390/s18124276

Fraternali, F., Balaji, B., & Gupta, R. (2018). Scaling

configuration of energy harvesting sensors with

reinforcement learning. Proceedings of the 6th

International Workshop on Energy Harvesting &

Energy-Neutral Sensing Systems, 7–13.

https://doi.org/10.1145/3279755.3279760

Bjorkqvist, O., Dahlberg, O., Silver, G., Kolitsidas, C.,

Quevedo-Teruel, O., & Jonsson, B. L. G. (2018).

Wireless Sensor Network Utilizing Radio-Frequency

Energy Harvesting for Smart Building Applications

[Education Corner]. IEEE Antennas and Propagation

Magazine, 60(5), 124–136.

https://doi.org/10.1109/MAP.2018.2859196

Weekly, K., Jin, M., Zou, H., Hsu, C., Soyza, C., Bayen, A.,

& Spanos, C. (2018). Building-in-Briefcase: A

Rapidly-Deployable Environmental Sensor Suite for

the Smart Building. Sensors, 18(5), 1381.

https://doi.org/10.3390/s18051381

De Donno M. and O'Flynn B. (2018). Innovative Low

Power Multiradio Sensing and Control Device for Non-

Intrusive Occupancy Monitoring. In Proceedings of the

7th International Conference on Smart Cities and

Application of AI-based Image Processing for Occupancy Monitoring in Building Energy Management

145

Green ICT Systems - SMARTGREENS,

ISBN 978-989-758-292-9 pages 173-181. DOI:

10.5220/0006692601730181

Jin, M., Bekiaris-Liberis, N., Weekly, K., Spanos, C., &

Bayen, A. (2015). Sensing by Proxy: Occupancy

Detection Based on Indoor CO2 Concentration. The 9th

International Conference on Mobile Ubiquitous

Computing, Systems, Services and Technologies

(UBICOMM'15)

Huang, Q., & Mao, C. (2017). Occupancy Estimation in

Smart Building Using Hybrid CO2/Light Wireless

Sensor Network. Journal of Applied Sciences and Arts,

1(2). https://opensiuc.lib.siu.edu/jasa/vol1/iss2/5

Naylor, S., Gillott, M., & Herries, G. (2017). The

development of occupancy monitoring for removing

uncertainty within building energy management

systems. 2017 International Conference on

Localization and GNSS (ICL-GNSS), 1–6.

https://doi.org/10.1109/ICL-GNSS.2017.8376240

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

Javed, A. (2020). Occupancy detection in non-

residential buildings – A survey and novel privacy

preserved occupancy monitoring solution. Applied

Computing and Informatics, 17(2), 279–295.

https://doi.org/10.1016/j.aci.2018.12.001

Serale, G.; Fiorentini, M.; Capozzoli, A.; Bernardini, D.;

Bemporad, A. (2018). Model Predictive Control (MPC)

for Enhancing Building and HVAC System Energy

Efficiency: Problem Formulation, Applications and

Opportunities. Energies 11(3), 631

Paschke, F.; Franke, M.; Haufe, J. (2017) Modellprädiktive

Einzelraumregelung auf Basis empirischer Modelle. In:

Gebäudetechnik in Wissenschaft & Praxis Vol. 138,

Nr.6, S. 468-476

Kelman, A.; Borrelli, F. (2011). Bilinear Model Predictive

Control of a HVAC System Using Sequential Quadratic

Programming. In: IFAC Proceedings Volumes Vol. 44,

Nr. 1, S. 9869-9874

Gwender, M.; Gyalistras, D.; Oldewurtel, F.; Wirth, K.;

Lehmann, B.; Stauch, V.; Sagerschnig C. (2010).

Prädiktive Gebäuderegelung mithilfe von Wetter- und

Anwesenheitsvorhersagen. Resultate des Projekts

OptiControl

Produkt PIR Sensor, Firma Busch-Jäger GmbH,

https://www.busch-jaeger.de/produktloesung/

bewegungs-praesenzmelder, Datum: 25.02.2020

Podukt Radar-Sensor, Firma Gira GmbH,

https://www.gira.com/en/gebaeudetechnik/produkte/au

tomatische-lichtsteuerung/beruehrungslose-

lichtschalter/sensotec.html, Datum: 25.02.2020

Produkt Stereo-Kamera, Firma Xovis AG,

https://www.xovis.com/products/detail/pc2/, Datum:

25.02.2020

Paschke, F.; Zaiczek, T., Röbenack, K. (2019) :

Identification of room temperature models using k-step

PEM for Hammerstein systems. In: Proceedings of the

23rd International Conference on System Theory,

Control and Computing, S. 320 – 325

Döge, J.; Hoppe, C.; Reichel, P.; Peter, N. (2015):

Megapixel HDR Image Sensor SoC with Highly

Parallel Mixed-Signal Processing. In Proc.

International Image Sensor Workshop (IISW).

Reichel, A.; Peter, N.; Döge, J.; Priwitzer, H.; Kasper, A.,

Ludwig, M., Ziems, B. (2019) Data fusion of multi-

spectral cameras on a low-power processing platform

for self-sufficient outdoor operation, Proceedings

Volume 11144, Photonics and Education in

Measurement Science, SPIE

SMARTGREENS 2022 - 11th International Conference on Smart Cities and Green ICT Systems

146