You are on page 1of 15

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/369568091

Thermal reality capturing with the Microsoft HoloLens 2 for energy system
analysis

Article  in  Energy and Buildings · March 2023


DOI: 10.1016/j.enbuild.2023.113020

CITATIONS READS

0 21

4 authors, including:

Tristan Kinnen Christoph Blut


RWTH Aachen University RWTH Aachen University
4 PUBLICATIONS   1 CITATION    18 PUBLICATIONS   38 CITATIONS   

SEE PROFILE SEE PROFILE

Joerg Blankenbach
RWTH Aachen University
111 PUBLICATIONS   831 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

As-built BIM View project

DigiPark View project

All content following this page was uploaded by Tristan Kinnen on 30 March 2023.

The user has requested enhancement of the downloaded file.


Energy & Buildings 288 (2023) 113020

Contents lists available at ScienceDirect

Energy & Buildings


journal homepage: www.elsevier.com/locate/enb

Thermal reality capturing with the Microsoft HoloLens 2 for energy


system analysis
T. Kinnen ⇑, C. Blut, C. Effkemann, J. Blankenbach
Geodetic Institute and Chair for Computing in Civil Engineering & Geo Information Systems, RWTH Aachen University, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Using digital methods to analyze the often highly complex structures of technical building equipment
Received 4 January 2023 (TBE) holds great potential for identifying energy savings. For digitizing TBE, we evaluated the
Revised 17 March 2023 Microsoft HoloLens 2 (MHL2), since it promises to be a mobile and cost-effective solution and found that
Accepted 23 March 2023
the MHL2 can be a suitable instrument for as-built acquisition, depending on the use-case’s accuracy
Available online 27 March 2023
requirements. The depth sensor of the MHL2 operates in a low single-digit centimeter accuracy range,
while the position tracking inaccuracy is more significant. Furthermore, we extended the acquisition sys-
Keywords:
tem with an external thermal camera to generate both, RGB and thermal point clouds. Additionally, we
Microsoft HoloLens 2
As-built acquisition
built a custom 3D calibration setup that allows us to calibrate both cameras in parallel and achieve
Thermographic point clouds promising accuracies for the level of detail of the resulting RGB and thermal point clouds and thus for
Technical building equipment the energy system analysis approach.
Sensor calibration Ó 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND
Co-registration license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction that is comprehensible to humans, a geometric, topological and


semantic as-built acquisition of the TBE is an essential first step.
The digitalization in the construction sector today is mainly The comprehension of topological relationships requires a collec-
characterized by the BIM method. BIM is already regularly being tion of geometric information with as little concealment as possi-
used for planning new buildings and across a building’s life cycle ble. A common method for this purpose is the use of terrestrial
[1]. Generally, BIM requires a continuous data flow. In particular, laser scanner (TLS) or mobile laser scanners (MLS). For enabling
after construction has finished, a consistent documentation of the the derivation of semantic information, RGB and thermographic
built situation (as-built) for creating an accurate digital twin and image data can be utilized by including them into the point cloud.
ensuring its up-to-dateness is crucial. Recently, energy-related The additional acquisition of thermographic data of the TBE fur-
topics have come into focus, for example, concepts that enable thermore permits the individual components to be categorized
the identification of potential for energy savings. The technical on the basis of the detection of heat and cold flows. As a result,
building equipment (TBE) is a key factor in the improvement of the systematic structure of the respective systems can be recon-
the energy efficiency of a building [2]. However, many construction structed and energy-related vulnerabilities in the plant’s system
projects that were planned with BIM have encountered various can be identified simultaneously. In general, data acquisition by
issues in which the built situation deviated from the planning means of scanning or imaging sensor technology is comparatively
model [34]. One of the reasons is that TBE typically is highly com- simple and can be performed rapidly nowadays. However, the
plex and the BIM method still has shortcomings in the field of instruments needed are either expert systems (e.g., total stations),
building services, for instance, the missing integration of aspects expensive (e.g., geodetic laser scanner, mobile laser scanners), or
such as building automation (BAS) and operational optimizations not well qualified for many applications (e.g., drone in indoor envi-
[1]. Accordingly, linking the advantages of digital working methods ronments). In contrast, the mixed reality (MR) glasses ‘‘Microsoft
and energetic system optimization of building technologies entails HoloLens 200 (MHL2) represent a potentially cost-effective and
a considerable amount of hidden potential. promising alternative, since the device already includes the neces-
In order to process the amount of information and communica- sary hardware, such as a depth sensor and a RGB camera for cap-
tion data concerning the complex structures of the TBE into a form turing colorized point clouds. Furthermore, it offers the
advantages of a head-mounted system with a high degree of
mobility and a user-friendly handling.
⇑ Corresponding author.
E-mail address: kinnen@gia.rwth-aachen.de (T. Kinnen).

https://doi.org/10.1016/j.enbuild.2023.113020
0378-7788/Ó 2023 The Author(s). Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

List of Abbreviations

2D two-dimensional MR Mixed Reality


3D three-dimensional PX Pixel
ADK Automotive Development Kit RGB Red-Green-Blue
API Application Programming Interface TBE Technical Building Equipment
BAS Building Automation System TLS Terrestrial Laserscanning
FPS Frames per Second ToF Time of Flight
Hz Hertz UV Axis designation of a two-dimensional coordinate sys-
IMU Inertial Measurement Unit tem
MHL Microsoft HoloLens UWP Universal Windows Platform
MHL2 Microsoft HoloLens 2 WDP Windows Device Portal
mK Millikelvin
MLS Mobile Laserscanning
MP Megapixel

In this paper, we investigate and evaluate the application the transition from triangular meshes captured with MHL to
potential of the MHL2 for as-built acquisition, specifically for semantically enriched, topologically correct building models is still
energy-related TBE. For this, we analyzed its sensors and their a challenge and holds much potential for future research. The
interfaces. Based on this, we present an off-the-shelf system con- result of [9] are confirmed by [8] and [11] by recording an average
sisting of the MHL2 and an external thermal imaging camera for distance of the meshes of the MHL to the registered point cloud of a
parallel acquisition of depth, RGB, and thermal data. With this laser scanner in the low to mid-single-digit centimeter range. With
approach, we can deliver an easy-to-use and cost-effective mobile this achieved accuracy, the MHL could fulfill the requirements for
reality capture system that does not require on-site professionals. an as-built acquisition as well as for a subsequent rule-based
In addition, in order to form a basis for energy system analysis reconstruction of a semantically enriched and topologically correct
from the acquired data, we developed and evaluated a 3D setup model. [8] furthermore investigated the capabilities of the MHL
for calibration of RGB and thermal sensors. Finally, with the help regarding its localization (in the sense of pose determination) in
of this method, we were able to develop an automated method indoor environments as well as its spatial mapping of indoor
for generating point clouds colored with both RGB and thermal scenes. To analyze the stability of the distance measurement, a
data as a basis for the analysis of energy systems. wall with an initially fully cooled down MHL was captured for
100 minutes in order to compare the temporal variation of the
2. Related work resulting distance data relative to the first frame. The results
demonstrated that the MHL only achieves stable range measure-
Since Microsoft released the first generation of the MHL in 2016, ments after an operating time of more than 60 minutes. Evalua-
it has been continuously researched for its diverse potential appli- tions of the MHL2 in terms of suitability as an instrument for as-
cations in a wide range of fields. Researchers have already recog- built acquisition have not been published to date. However, first
nized the potential of the MHL for the acquisition of geometry results for the evaluation of the sensors of the MHL2 are shown
data of construction projects and investigated it under various by [12], whose investigations revealed an average deviation of
research focuses. 2 cm for its hand-tracking accuracy.
As shown in [5], a challenge with the first generation of MHL The function for creating point clouds colored with data from
was that the captured object or environment data was only avail- RGB images has become a standard feature of many commercial
able to the user in the form of meshes instead of point clouds. software products from laser scanner manufacturers. [5] and [7],
However, a mesh is a discretization of a geometric domain into on the other hand, presented two different methods to realize
small simple shapes [6]. Therefore, the process of generating a the generation of colored point clouds in real time with sensor data
mesh is always based on the existence of a quantity of geometric of the MHL. The method of [7] is based on the external acquisition
point data. Accordingly, the MHL scans its surroundings with the of geometric data by a higher quality depth sensor and an addi-
built-in depth sensor and initially generates a point cloud before tional RGB camera connected to a computer. Via a WiFi connection,
automatically converting it into a triangular mesh. As access to the data is transferred to the MHL in real time. The external RGB
the raw data of the MHL sensors is not provided, [7] titled the uti- camera acts as an interface for the mapping process by calibrating
lization of the MR glasses as a research tool as severely limited. with the RGB camera of MHL via a manually set up checkerboard.
Furthermore, the first generation of the MHL suffers from a lack This enables the data from the external depth sensor to be trans-
of high-quality hardware to enable accurate geometric data acqui- formed first to the coordinate system of the external RGB camera
sition, resulting in an inability to compete with the geometric and then to the coordinate system of the MHL through the existing
accuracy achieved by a TLS as demonstrated by [7–10]. In [9] the calibration. One advantage of this system is that higher quality
MHL was investigated for as-built acquisition by scanning an results can be achieved by using external sensors. However, the
empty apartment several times and comparing the generated system design limits mobility during the acquisition process by
meshes to a point cloud captured by a TLS. The presented results carrying external sensors. In contrast, the method developed by
of the paper show a marginal variance between different acquisi- [5] focuses exclusively on the data from the built-in sensors of
tion processes of an identical environment with the MHL. The the MHL by coloring the dynamic meshes in real time using the
observed mean variance between the acquired data of the MHL images from the built-in RGB camera. For this purpose, the tech-
and the data of the TLS amounted to 2.4 cm inside of a single room. nique of UV mapping is used, which projects a two-dimensional
With the addition of neighboring rooms and larger distances image onto the surface of a three-dimensional model for texture
between the surface to be captured and the depth sensor of the mapping as also shown by [13]. The abbreviation UV denotes the
MHL, the deviations increased. In conclusion, it was found that axes of the two-dimensional image. Consequently, the system first

2
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

creates a UV texture into which the RGB data are transferred. Then
the colored UV texture is mapped onto the mesh. To realize this
process in real time, an adaptive mapping scheme is used that dis-
tributes each triangle of the dynamic mesh evenly over the texture
and adapts to the new geometries without affecting existing color
data. Approaches to generate colored point clouds from MHL2 sen-
sor data have not yet been published. However, [14] presented an
end-to-end framework for transferring and interacting with 3D
point clouds to the MHL2 system.
The development of methods for the generation of thermo-
graphic point clouds are in principle documented in different con-
texts. [15] discussed the topic of how a 3D point cloud of the
environment with a TLS and a thermographic panorama taken with
a nodal point adapter can be linked together. [16] has shown a
Fig. 1. Overview of the sensors of the MHL2.
method to generate a 3D building model with color and thermal
information from the acquired data enabled by a handheld system
consisting of a ‘‘Microsoft Kinect” and a simple infrared camera. Table 1
[17], on the other hand, presented a fully automated method for List of sensors of the MHL2 and their functions [19].
generating point clouds colored by RGB and thermal data. In this Sensors Quantity Resolution Function
case, a laser scanner, an RGB camera and a thermal camera were
Grayscale visible light 4 – Head-Tracking
permanently mounted on a freely controllable robot. Therefore,
cameras
the advantage that all sensors are calibrated is given. However, Infrared cameras 3 – Eye-Tracking
the robot is not variable in height and can therefore only record Time-of-flight depth sensor 1 1 MP Depth acquisition
all data from a constant small distance above the ground, which Accelerometer, Gyroscope 1 – Orientation
and Magnetometer determination
means that parts of the environment to be captured could not be
(Inertial Measurement
recorded. Unit (IMU))
Consequently, some methods for generating thermal point RGB camera 1 8 MP Color image and
clouds in general do already exist. However, no procedure has Image, video recording
yet been developed that makes it possible to link thermographic 1080p30
Video
image recordings with the depth points recorded by the MHL2.
Furthermore, in contrast to its predecessor, the MHL2 has not yet
been analyzed for its potential for its application as an as-built
acquisition system. thermal information of their surroundings. For creating thermal
point clouds, it is necessary to integrate an additional external
sensor.
3. Material and methods The depth sensor of MHL2 operates in two different modes for
near and long-distance ranges. The near range mode is aimed to
In this chapter, we first introduce the sensor system of the track the hand movements of the user (Articulated Hand Tracking).
MHL2 and its technical data. Subsequently, we present our method Opposite is the long-distance range mode, which provides the geo-
for the sensory extension of the acquisition system and for the metric mapping of the environment. In both cases, the respective
automated sensor data fusion for the generation of colored point distance is determined via the Time of Flight (ToF). The arrange-
clouds. ment and the different lenses of the MHL2 camera sensors result
in different fields of detection of the environment. These overlap
3.1. Working with the MHL2 each other partly to a greater and partly to a lesser extent. A
detailed illustration of this is given in [9]. The fact that the detec-
The MHL2 MR glasses are a head-mounted and therefore hands- tion fields of the RGB camera and the depth sensor completely
free operable computer that creates a digital layer of information overlap provides an advantage by ensuring that a high amount of
within the real world, which allows users to interact with holo- pixel of a RGB image can be assigned to the acquired depth values.
grams that remain in a specific position and behave like physical Our approach of a co-registration for simultaneous mapping of
objects [18]. From a technical point of view, the constant acquisi- depth, RGB and thermal data requires an access to the raw data
tion of the spatial environment is fundamental for operating in from the MHL2 sensors. For the advancement of academic or
MR. Accordingly, the hardware of the MHL2 includes various sen- industrial research and development processes in thematic fields
sors that are integrated in the visor of the glasses. An overview of in which the MHL2 may be applied, Microsoft has published the
the sensors of the MHL2 and their functions is shown in Fig. 1 and option of a Research Mode. With this option, access to the raw data
Table 1. of the MHL2 sensors is provided by various application program-
This technical equipment enables real-time 3D mapping of the ming interfaces (APIs) and tools [20]. The research mode can be
immediate environment and self-localization within the acquired activated in the Windows Device Portal (WDP), which needs to
3D data [8]. The position of the MHL2 can be determined from be previously set up in the settings of the MHL2 under the options
the data of the four grayscale cameras and the inertial measure- of the developer mode.
ment unit (IMU). Therefore, it represents a highly potential system Since each sensor operates in an individual coordinate system,
for the as-built data acquisition in the construction sector and transformation matrices that include rotational as well as transla-
offers the possibility of direct visualization of the results and a tional parameters are generated for each acquired frame. There-
higher mobility level compared to other data acquisition systems. with, the different sensor data must be transformed into a
According to the sensor data, the built-in infrared cameras are common coordinate system. In this context, the left front grayscale
responsible for the eye tracking. Consequently, these cameras are camera serves as the reference system for the MHL2. The data can
focused on the eyes of the user and are not suitable for recording be subsequently transformed from the reference coordinate sys-

3
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

tem of the left front greyscale camera into world coordinates, if the
location of at least one position in the acquired area is defined. To
obtain access to the data provided by the research mode, we used
the StreamRecorder application from the HoloLens2ForCV1 reposi-
tory, which can be manually set to save the data of the ToF depth
sensor and the RGB camera. In this way, the data from the MHL2 sen-
sors can be recorded for a user-defined period of time, and subse-
quently be stored on the device. After the acquisition, the data can
be downloaded from the WDP. The data of the ToF depth sensor is
available as 2D depth images. So, the data package of an acquisition
process with the StreamRecorder application includes 2D images of
the depth camera and the RGB camera as well as all necessary geo-
metric values needed for the transformation of different local sensor
data into a common coordinate system, including transformation
matrices and calibration information of the RGB camera. Based on
this data structure and the user-friendly interface of the StreamRe-
corder application (Fig. 2), the MHL2 represents a promising acquisi-
tion system that does not require qualified personnel to operate.
Fig. 3. FLIR Boson ADK thermal camera.

3.2. System expansion interfaces


tion was paid to maximize the robustness and stability of the
As mentioned in 2.1, the infrared sensors of the MHL2 are not
mount for the FLIR Boson ADK for the co-registration. In addition,
suitable for capturing thermal data. Therefore, we have extended
a small tilt angle was added to increase the overlap between the
the system with the FLIR Boson ADK thermal camera (Fig. 3). The
FLIR Boson ADK and the depth and RGB camera of the MHL2.
technical data of the FLIR Boson ADK are listed in the following:
In the following, the individual components of our 3D printed
mount are illustrated in Fig. 4 and our resulting acquisition system
 Resolution: 640  512px
is presented in Fig. 5.
 Frame rate: 60 Hz
During the acquisition process, images from the depth camera
 Sensitivity: 50mK
and the RGB camera of the MHL2 are captured in parallel via the
 Aperture angle: 32°
StreamRecorder application, from which RGB point clouds can be
 IP67 case
generated in postprocessing. For the additional generation of ther-
mal point clouds, a parallel acquisition of thermal images is
3.2.1. Mounting the FLIR Boson ADK to MHL2 required. Unfortunately, the extension possibilities of the MHL2
Since we did not intend to limit the factor of the high amount of are limited, so that the USB-based FLIR Boson ADK cannot be con-
mobility that the MHL2 entails as a system for the data acquisition, nected directly. Therefore, we connected the FLIR Boson ADK to an
we followed the approach of attaching the FLIR Boson ADK thermal external computer and incorporated it into the system architec-
camera directly to the MHL2. However, in order to achieve high ture. The external computer is responsible for reading and trans-
quality thermographic point clouds, a constant offset and orienta- mitting the thermal images to the MHL2. For this, we developed
tion of the sensors of the MHL2 and the FLIR Boson ADK thermal a client application that runs on the external computer and a
camera over the time of the acquisition process is important. We Universal Windows Platform (UWP) server application for the
adapted and optimized a publicly available 3D-printed camera MHL2. The communication is based on the Transmission Control
mount to attach the thermal camera to the MHL2. Particular atten-

Fig. 2. User Interface of the StreamRecorder-App.

4
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 4. 3D-printed mount for attaching the FLIR Boson ADK thermal camera to the MHL2.

Fig. 5. MHL 2 with FLIR Boson ADK thermal camera.

Protocol. In this way, a signal is transmitted from the MHL2 to the 3.2.2. Calibration
external computer at the time when the user starts an acquisition Since the thermal camera is mounted on top of the MHL2, it is
process in the StreamRecorder application, initiating the parallel shifted vertically. To combine the thermal images with the depth
acquisition of thermographic image data. The system times of the images, it is necessary to determine the exact transformation of
external computer and MHL2 are synchronized, so that each depth, the thermal camera sensor and the depth camera sensor. This is
RGB and thermal image can be stamped with the acquisition time. done during a system calibration, before starting the capturing
These timestamps are based on the standard defined by Windows process. In general, camera calibrations can be performed image-
(one hundredth of a nanosecond). With this approach, data from based on the basis of corresponding feature points. The RGB cam-
the depth, RGB and thermal cameras can be uniquely mapped to era of the MHL2 is suitable as a reference in this context, since the
each other as part of a co-registration process. As a result, the color RGB images, in contrast to the depth images, enable an unambigu-
images can be combined with the depth data to generate colored ous detection of the feature points. Furthermore, the RGB camera is
point clouds. An example of the raw data of the depth, RGB and already calibrated to the depth camera as a part of the system of
thermal sensors for one frame is visualized in Table 2 for the acqui- the MHL2 and its calibration parameters are provided within the
sition of a heater. data of the StreamRecorder application. As part of the thermal
In order to use the FLIR Boson ADK thermal camera for captur- camera calibration, the intrinsic and extrinsic as well as the deriva-
ing thermal images intended for coloring point clouds, it is neces- tion parameters are determined. The intrinsic parameters include
 
sary to reconfigure the system settings for the color visualization the camera-specific parameters, such as the focal length f x ; f y
for each new data acquisition, as it is basically set to define the  
and the image principal point cx ; cy . The extrinsic parameters,
color scale always in relation to the warmest and coldest areas of  
on the other hand, describe the translational t x ; t y ; tz and rota-
a respective exposure field. This setting leads to the risk that the
tional ðRÞ offset between the RGB camera and the thermal camera
color representation of the same objects may differ in the different
expressed by a transformation matrix, which is composed by a
frames, resulting in incoherent colors in the point cloud. To fix this,
rotation matrix and a translation vector. For RGB cameras, black
the damping factor must be adjusted in a way that the color repre-
and white checkerboard patterns printed on paper are therefore
sentation is applicable for the current scene.
often used, as the corners of the squares can be detected easily
and robustly as feature points. However, simple paper based black

5
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Table 2
Raw data of the depth, RGB and thermal camera.

Depth camera RGB camera Thermal camera

and white checkerboard pattern that cannot generate any heat are subsequently be determined. To ensure that the most accurate
unsuitable for the calibration of thermal cameras, since the sensor parameters possible result from a calibration process, it is essential
reacts to thermal radiation. For this reason, we developed a calibra- that the positions of the various sensors are constant between the
tion system (Fig. 6). Light bulbs were inserted in an irregular man- respective images. For this reason, we placed the acquisition sys-
ner to serve as heat sources that can be detected by the thermal tem on a stationary surface for the duration of the recording.
camera. In addition, uniquely identifiable markers were placed
on the front side with the light bulbs located in their respective
centers. These centers were determined photogrammetrically in 3.3. Generation of colored point clouds
advance from the image set of a high-resolution single lens reflex
camera. The resulting positions for the centers of the markers serve The process for generating point clouds colored with both RGB
as ground truth for all calibration processes of the RGB camera of and thermal data starts with the selection of a specific image of
the MHL2 and the FLIR Boson ADK thermal camera. The respective each sensor. While the RGB camera of the MHL2 captures 30
marker centers can be determined from the images of the MHL2 images (or frames) per second (FPS), the frame rate of the depth
RGB camera in reference to the ground truth data within a calibra- camera is 1 to 5 FPS and is therefore not constant [21]. By relating
tion process. In the thermal images, the light bulbs are each repre- the respective scan times to the floor area of the acquired environ-
sented as heat sources in the form of dots. Since the centers of the ment of a scan process, inferences can be made that the amount of
markers and of the light bulbs are identical, the centers can be taken depth images depends on the movement speed of the MHL2
determined from the images of the thermal camera equivalent to and thus on the head movements of the user during the acquisition
the RGB images in reference to the ground truth. The extrinsic process. The data from different scans show that the frame rate of
and intrinsic calibration parameters of the thermal camera can the depth camera increases equivalently to a faster acquisition
speed. As a result, the number of sensor data used as a starting

Fig. 6. Setup for calibrating the FLIR Boson ADK thermal camera to the RGB camera of the MHL2.

6
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

point for the generation of colored point clouds is unequal, so that StreamRecorder software application at the start of each acquisi-
a selection must be made as to which color images are assigned to tion process, so that data can be transferred from the reference sys-
the respective depth images. As a solution to this problem, we tem of the left front grayscale camera via the world coordinate
selected the corresponding images of the color sensors based on system to the coordinate system of the RGB camera. Subsequently,
the known timestamps of the files. The reference element for this the 3D points are converted to the coordinate system of the RGB
operation is the grayscale image of the depth sensor. Based on camera with the transformation matrices provided by the MHL2.
the timestamp of this image, the respective timestamp with the Using the known extrinsic calibration parameters, which include
smallest temporal difference is determined in the lists of all times- the rotational and translational spatial offset between the MHL2
tamps of the data sets of the RGB camera and the thermal camera. RGB and the FLIR Boson ADK thermal camera, the depth data can
All following coordinate transformations are based on a pinhole be transferred in parallel to the coordinate system of the thermal
camera model as presented in [22] (Fig. 7). The overall schematic camera at this point. Continuing, the depth points are projected
process workflow of our mapping method is shown in Fig. 8. onto the two-dimensional image planes of the RGB and thermal
For transforming 2D image points ðu; v Þ to 3D coordinates camera using the respective intrinsic camera parameters. In order
ðX C ; Y C ; Z C Þ in the corresponding camera coordinate system can to optimize the mapping results, each frame of the FLIR Boson ADK
be performed by using equation (1) [23,24]: thermal camera is previously corrected for distortion with the help
2 3 2 32 3 of the determined distortion parameters of the calibration process.
u fx 0 cx XC In this manner, the pixels of the color sensor frames that are pre-
6 7 6 76 7
s4 v 5 ¼ 4 0 fy cy 54 Y c 5 ð1Þ sent in the temporally allocated viewing area of the depth sensor
1 0 0 1 Zc can be identified and each depth point can be within the intersect-
ing areas is thereupon assigned a color value from the respective
The transformation of 3D coordinates to world coordinates color sensor data.
ðX w ; Y w ; Z w Þ or to a different camera coordinate system, on the For generating a point cloud that has points are simultaneously
other hand, requires the multiplication with the extrinsic parame- assigned color values from the image data of the RGB and the FLIR
ters as shown in equation (2) [23,24]: Boson ADK thermal camera, we have followed the approach of
2 3 transferring the depth data and the respective assigned points from
2 3 2 3 Xw
x0 r 11 r12 r13 tx 6 7 the data of the color sensors to a common image plane. To maxi-
6 7 6 76 Y w 7 Xc Yc mize the accuracy of the representations of the object geometries
Z c 4 y0 5 ¼ 4 r 21 r22 r23 ty 56 7; with x0 ¼ and y0 ¼
4 Zw 5 Zc Zc in the point cloud, a virtual pinhole camera is generated whose
1 r 31 r32 r33 tz
1 projection center corresponds to the depth sensor of the MHL2.
ð2Þ This ensures that the highest possible amount of acquired depth
points is transferred to the point cloud. For the same reason, the
After assigning the sensor images for the respective frame, the image sizes are chosen to be identical to those of the grayscale
depth points are transformed from the 2D image plane into 3D images of the depth camera and the principal point of the image
points in the coordinate system of the depth sensor. In the follow- is set to correspond to the center of the image. The depth points
ing, the 3D depth points are converted to the coordinate system of are projected onto the image plane at first. Afterwards, the respec-
the RGB camera. For this purpose, the depth images are first pro- tive color images are transformed onto the image plane of the pin-
jected onto the reference system of the left front grayscale camera. hole camera based on the previous assignment of the RGB and
Afterwards, the points are transformed from this coordinate sys- thermal image points to the depth points. In the following, the
tem into world coordinates. The reference point for defining a resulting images for the depth, RGB and thermal data are stored.
world coordinate system is determined automatically by the

Fig. 7. Illustration of the relations of image plane, camera coordinate system and world coordinate system in the pinhole camera model [22].

7
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 8. Schematic representation of the mapping method.

This entire process is repeated for each existing depth image. For tances to the objects and constantly slower movements of the
the generation of the colored point clouds, the stored image pairs acquisition system in order to achieve higher accuracies of the
are finally superimposed. Because of the defined data structure, depth data. Especially in the field of TBE acquisition, the high
the assignment of color values from the RGB and thermal data amount of metal surfaces in combination with punctual artificial
can consequently be retrieved pixel by pixel for each depth point. lighting poses a challenge. Therefore, the examination of the
After the mapping, the depth points are transformed again into 3D MHL2 point clouds initially reveals a high scattering of points,
world coordinates and written into a point cloud file with their especially in the areas that represent the surfaces of objects. In
two-color values. addition, it is evident from the point cloud analysis that this scat-
tering increases for objects that could only be captured from a
4. Results greater distance due to the surrounding structure, so that signifi-
cant details and structures are displayed more inadequately within
In the following, we first present our results on the evaluation the point cloud. An example of this is illustrated in Fig. 9. The figure
of the MHL2 in terms of its usability for the as-built acquisition. presents the frontal and lateral view from the point cloud gener-
Afterwards, we summarize our critical investigation of our method ated from depth images taken by the MHL2 of four jacketed pipes
for calibrating the MHL2 with the FLIR Boson ADK thermal camera with control elements and an electronic distribution box, which
and its results. Lastly, we present our obtained results of generat- stood under direct lighting and could only be acquired from a dis-
ing colored point clouds resulting from sensor data fusion. tance of more than half a meter. In the frontal section of the point
cloud, some points are discernable in the free space between the
4.1. Evaluation if the MHL2 in terms of depth data acquisition four pipes, which is a result of the high number of reflections in
this area. Furthermore, the frontal and lateral sections indicate
In order to evaluate the application of the MHL2 as a tool for the almost imperceptible representation of the contours of the elec-
as-built building acquisition, the point cloud quality and the tronic distributor box and the four control elements in the point
achieved accuracies of the ToF depth sensor have been investi- cloud. Similarly, the cables located in this area are not represented
gated. First, a representative test environment was selected. As in the point cloud. Objects with a cross-sectional area smaller than
part of the energyTWIN2 project, a domestic hot water simulator one centimeter are generally only recognizable in the point clouds
provides access to all components of a TBE system so that developed of the MHL2 if they were captured from a very short distance.
methods for the optimization of the energy efficiency can be tested Opposite to this, the parts of the captured environment, which
and evaluated. The water simulator is located in a large-scale exper- were captured from a short distance and with a low level of occlu-
imental hall over three levels on a steel frame. The lowest and upper- sion, are represented with clearer structures in the point cloud, as
most levels each contain the main components of the TBE system, shown in Fig. 10 by the example of a black pipe equipped with a
while the middle level has two simulation rooms for bathrooms with valve. However, the point cloud was manually processed by filter-
the corresponding connections and fittings. As a reference acquisi- ing and removing the scatter points in the empty space in this case,
tion system, a terrestrial laser scanner (TLS) Riegl VZ-400 was used aiming to provide a distinct contrast. Nevertheless, the removed
in combination with an external RGB camera. According to the man- points and the scatter are significantly less existent compared to
ufacturer, the data acquired by the TLS VZ-400 has a geometric accu- the example from Fig. 9, so that the shape of the pipe and its valve
racy of a maximum deviation of five millimeters [26]. do not have to be estimated, but are clearly recognizable which
From the analysis of numerous acquisition series with the enables the determination of geometries as well as the derivation
MHL2, we have found that the risk of multiple reflections of the of topological relationships.
emitted rays significantly increase at longer distances to the Like mobile laser scanners, the MHL2 has advantages compared
objects to be scanned. Therefore, we recommend short scan dis- to a TLS of a time-efficient acquisition process and the ability to
acquire object geometry with little occlusion, which can simplify
2
the derivation of topological relationships. However, mobile acqui-
energyTWIN is a research project that investigates new sensor-based and AI-
based methods for the digital, BIM-based commissioning of technical systems in
sition methods often exhibit significant systematic drift [27]. This
building structures and their energetic system optimization [25]. drift, in addition to the scan duration and the distance of the acqui-

8
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 9. Example of the effect of larger acquisition distances in combination with the problem of multiple reflections.

Fig. 10. Prepared output display for the acquisition of TBE components from short distance with the MHL2.

sition path, is fundamentally dependent on its extend on the IMU. 4.2. Calibration of the MHL2 with the FLIR Boson ADK thermal camera
To investigate the drift of the MHL2, we processed a six-minute
scan with an acquisition path of 75 m and compared it to the The foundation for any successful fusion of camera sensor data
ground truth data from the TLS. Meanwhile, the MHL2 acquisition consists in calibrating the systems as accurately as possible. For
was performed without visiting a location twice. The point clouds this reason, we iterated the calibration process numerous times
of the MHL2 and the Riegl VZ-400 were registered only on the basis for the MHL2 and FLIR Boson ADK camera systems. Thereby, the
of prominent control points in the immediate vicinity of the start- number of recording sites differed in the individual implementa-
ing point of the MHL2 acquisition. Using this approach, an average tions. Generally, the calibration quality increases with the number
drift of 75 cm was detected at the end of the MHL2 acquisition path of images taken from different positions and angles.
by measuring distances of identical points in both point clouds as In order to analyze the consistency of the calibration parame-
shown in Fig. 11. In terms of the length of the acquisition path and ters determined by our method, we first investigated the devia-
the acquisition time, this corresponds to a drift of one centimeter tions of the certain parameters from different calibration
per meter and 0.2 cm per second, respectively. processes. Table 3 contains the specific intrinsic parameters of
To investigate the accuracy of the distance determination of the the FLIR Boson ADK camera from four different calibrations with
MHL20 s depth sensor, a comparison method was chosen that their respective standard deviation. Generally, intrinsic camera
excludes the influence of drift. For this reason, we examined a sec- parameters are always interdependent [28]. For this reason, and
tion of a wall plane taken immediately at the beginning of the due to the fact that, on the one hand, the bulbs in the calibration
acquisition process with the MHL2 and from a distance of less than setup are of different depths at each feature point, and on the other
two meters. In this case, the control points for matching the align- hand, the FLIR Boson ADK thermal imaging camera is equipped
ment and orientation of the MHL2 point cloud with the point cloud with the function of an automatic focus, occasional variations of
of the TLS were set in close proximity to the area from which the the determined intrinsic values in our calibration are unavoidable.
imaged wall section originated. Thus, a drift deviation can be Besides determining the intrinsic and extrinsic parameters, we
excluded and the detection accuracy of the depth sensor of the also identified the distortion factors within the calibration process
MHL2 can be examined. As a result of this observation, a maximum to correct each FLIR Boson ADK camera image prior to the mapping
deviation of two centimeters can be observed in the point cloud of process to the extent that the distortion minimally affects the
the MHL2 in the area of the wall surface as shown in Fig. 12. result. To illustrate the visual results of the mapping of an image
pair realized under the implementation of previously determined
9
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 11. Deviation of the point cloud of the MHL2 (yellow) and the point cloud of the VZ-400 (green) as a result of drifting of the pose along the acquisition path. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 3
Compilation of certain intrinsic camera parameters of the FLIR Boson ADK camera
from different calibration processes.

Calibration process Focal length Principal point Principal point


(X-Component) (Y-Component)
1 13.4524 0.5221 0.0784
+/- 0.0750 +/- 0.1678 +/- 0.1320
2 13.4186 0.0053 0.0113
+/- 0.0304 +/- 0.0249 +/- 0.0272
3 12.6655 0.0133 0.0081
+/- 0.0460 +/- 0.0263 +/- 0.0189
4 13.3668 0.0091 0.0517
+/- 0.0257 +/- 0.0313 +/- 0.0331

* All dimensions in [mm].

ginal deviations exist at the edges of the thermal image, which is


partly due to camera-specific factors.
Since the calibration board was photogrammetrically measured
in advance with an image composite of a high-resolution camera,
which provided the exact coordinates of the feature points, the
RGB image of the MHL2 is transformed into the same coordinate
system in the first step of the calibration. Subsequently, the devia-
tions of the image coordinates from the control points are calcu-
lated on the image plane. The result is the relative orientation of
the MHL2 as a transformation matrix. In the second step of the cal-
Fig. 12. Evaluation results of the accuracy of the MHL2 depth sensor (yellow)
ibration process, the thermal image is transferred to the MHL2
versus a TLS (red). (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.) coordinate system and the relative orientation of the FLIR Boson
ADK camera is determined as before by measuring the deviations
of the image coordinates of all control points of the thermal camera
calibration parameters is shown below in Fig. 13, that the image with those of the RGB camera of the MHL2 and calculating the best
coordinates of the FLIR Boson ADK camera and the RGB camera mathematical approximation to match them. This process is
of the MHL2 match particularly accurately in the central area repeated for each image pair of an acquisition location. By imagin-
around the principal point of the thermal image. However, mar- ing the connection between the projection center, the image plane
and the control points as a ray, the totality of all rays is bundled in

10
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 13. Visual representation of the calibration results between the MHL2 and the FLIR Boson ADK thermal camera.

the projection center. Fig. 14 schematically shows the bundle for the FLIR Boson ADK thermal camera. The use of the mean rela-
adjustment for the calibration of the sensor systems. tive orientation results in fewer degrees of freedom for the system.
To classify and evaluate the calibration results, it is necessary to As a result, the residual gaps of the bundle adjustment and conse-
consider the residuals in the image after bundle adjustment. How- quently the standard deviations of the image coordinates increase.
ever, since the determined deviations of individual image coordi- To evaluate the effects of this, some of the degrees of freedom were
nates are not constant within the analysis of an image pair and replaced with constants by assuming the known relative orienta-
the derived transformation matrix merely describes the best possi- tion of the MHL2 to be correct and using multiple image pairs from
ble mathematical approximation, residuals occur. To evaluate the a calibration process to determine the mean relative orientation of
calibration results, we analyzed these residuals, which can also the FLIR Boson ADK thermal camera. Subsequently, the thermal
be referred to as the standard deviation of the image coordinates. images were again projected into the coordinate system of the
Therefore, we determined the standard deviations in the direction high-resolution camera and the deviations to the control point cen-
of the two axes of the image plane separately for each RGB and ters were measured. Based on these data, we were able to deter-
thermal image pair during the calibration process. From these mine the standard deviations of the image coordinate similarly
determined data, we subsequently determined the mean values to the previous method. Thereby, the result of a total of 608 mea-
of the standard deviations in the direction of the two image axes surements was a standard deviation of the image coordinate of the
for each calibration process, in order to calculate the standard devi- FLIR Boson ADK thermal camera of 7.246 pixels or 0.087 mm.
ation of the image coordinate. The results of this method are
shown in Table 4 for four different calibration processes.
According to the determined data, both sensor systems point 4.3. RGB- and thermographic colored point clouds
out a minor standard deviation of the image coordinates from
the bundle adjustment. For the FLIR Boson ADK thermal camera, Generally, the results are significantly improved if the acquisi-
the maximum value amounted to less than two pixels for a pixel tion distance during the acquisition process is less than 2 m, as
size of 0.012 mm. For the MHL2, on the other hand, no maximum shown in the evaluation of the MHL2 for depth data acquisition
value of more than one pixel was determined for a pixel size of (Section 3.1). Especially for the coloring results in the point cloud,
0.05 mm. Furthermore, the results show that the accuracy of the we recommend constant and slow head movements with the
calibration method improves simultaneously with the number of acquisition system. Furthermore, in the comparison of different
image pairs. acquisition processes realized conclusively that the starting point
However, the presented results are based on a calculation that of the acquisition is of crucial importance for the color representa-
includes the specific relative orientations of the sensor systems tion of the thermal data due to the system configuration for the
of each image pair of a calibration. But for the subsequent sensor acquisition of absolute color values. The finest coloring results
data fusion and generation of colored point clouds, we used the were obtained when the acquisition process began with the cap-
mean relative orientation of all image pairs of a calibration process ture of a representative component of the TBE, which was charac-
terized by high heat or cold intensities relative to the other
11
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 14. Schematic illustration of the bundle adjustment process for the sensor calibration.

Table 4
Standard deviations of the image coordinates after bundle adjustment for images taken from a distance of between 1.50 and 2 m.

Calibration Number of Image Pairs Standard Deviation of the Image Standard Deviation of the Image
Coordinates (FLIR Boson ADK, Coordinates (MHL2,
pixel size = 0.012 mm) pixel size = 0.05 mm)
[mm] [Pixel] [mm] [Pixel]
1 4 0.0221 1.8403 0.0047 0.9477
2 9 0.0081 0.6741 0.0018 0.366
3 7 0.0079 0.6543 0.0047 0.9477
4 15 0.0065 0.5379 0.0015 0.2967

components of the systems. In compliance with this, we achieved clouds, which represent a promising basis for the collection of geo-
detailed color representations of the RGB and thermal camera data metric information as well as for the derivation of topological and
in the resulting point clouds. An example of the results is shown in semantic data.
Fig. 15 for an acquisition of a technical plant.
The figure shows that the representation of distinct structural
edges as well as the color representation is especially accurate 5. Discussion
for large objects. In addition, even smaller color gradients such as
isolated heat regions on the walls or reflected heat radiation on We developed a functional sensor system for the acquisition of
the neighboring objects are transferred to the 3D point cloud geometric, real color and thermal environmental data consisting of
unambiguously and identically to the original sensor data. How- the Microsoft MHL2 and the FLIR Boson ADK thermal camera,
ever, particularly objects with a comparatively small cross- implemented a method for the sensor data fusion for the genera-
section, such as valves or narrow pipes, occasionally reveal indis- tion of point clouds colored with RGB and thermal data and evalu-
tinct color gradients along the edges of the structure due to scat- ated the results. The use of this head-mounted sensor system for
tered color transitions between the various objects. Furthermore, as-built acquisition offers the advantage of a high degree of mobil-
chromaticity values are also assigned to points in actually empty ity, enabling low occlusion capture of objects. Moreover, the sys-
space, which is due to the previously mentioned problem of multi- tem provides the advantage of a user-friendly interface, which
ple reflections and the method of sensor data fusion on a two- eliminates the need for specialized personnel on site. The evalua-
dimensional level. Nevertheless, the acquisition system and the tion of the accuracy of the MHL2 depth camera revealed a maxi-
presented method enable the generation of colored 3D point mum deviation of 2 cm. Therefore, the MHL2 is not able to
compete with high-priced sensor systems, but it represents a
12
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

Fig. 15. Exemplary representation of the obtained colored point clouds of a TBE element.

promising and cost-effective alternative for use cases whose the realizable accuracy of the calibration setup. Considering the
requirements are compatible with the achievable accuracy of the averaged relative orientation for the determination of the standard
depth camera of the MHL2. However, the MHL2 shows a large drift deviations of the image coordinates of the FLIR Boson ADK thermal
in the course of longer recordings and paths. For this reason, it is camera, the value of the scatter increases into the range of a low
not recommended to use the MHL2 for large-scale acquisitions two-digit pixel number. Nevertheless, the degree of dispersion is
without splitting them into several scans. Furthermore, since the to be classified as marginal. However, the calculation of the aver-
presented acquisition system aims to collect data for the energetic aged relative orientation is based on the assumption that the
analysis of TBE components, which often have metallic surfaces known relative orientations of the MHL2 correspond to the ideal
and are placed under artificial light, multiple reflections may occur. values. Consequently, the relative orientation of the MHL2 needs
Due to this, the point cloud can partially exhibit strong noise, but to be verified initially for a more accurate evaluation of the calibra-
this is a typical challenge that acquisition systems must deal with. tion results. More generally, the calibration procedure also requires
Nevertheless, we have found that although manual processing of that the calibration parameters be determined prior to each new
the point cloud may be required for some components, most rep- acquisition process, since the intrinsic camera parameters cannot
resentations of the components in the point cloud show a sufficient be constant for the sensors in addition to the extrinsic camera
level of detail. Future work towards integrating an automated pro- parameters, as in the case of a total station, for example. On the
cess for point cloud processing, e.g., using neighborhood analysis to basis of our calibration method, we successfully implemented an
identify points in empty space, could counteract this issue. For the automated generation of 3D point clouds colored with RGB and
sensor data fusion of the depth, RGB and thermal camera, a custom thermal image data. The results of the colored representations in
calibration setup was developed to determine the intrinsic and the point clouds achieve a level of detail that serves for the recog-
extrinsic camera parameters of the FLIR Boson ADK thermal cam- nition of energetic saving potentials and topological correlations as
era in relation to the MHL2 system. By determining the standard well as for the derivation of semantic object data in particular for
deviations of the image coordinates of the RGB camera of the bigger object. The extent to which these subsequent analysis pro-
MHL2 and the FLIR Boson ADK thermal camera from the bundle cedures can be implemented in an automated process offers much
adjustment equation, the dispersion for both systems were shown potential for future investigation. For objects with particularly
to vary within a spectrum of a low single-digit pixel number. Since small cross-sections, such as valves or narrow pipes, the geometric
the calculation of these values includes the specific relative orien- and color representations are imperfect and therefore prospec-
tation of each thermal image and not the averaged relative orien- tively unusable for further analysis procedures. However, the over-
tation from all image pairs of a calibration process, these results all results are dependent on the speed of movement and the
primarily serve to prove the basic functionality and to classify distance to the objects during the acquisition process. The best

13
T. Kinnen, C. Blut, C. Effkemann et al. Energy & Buildings 288 (2023) 113020

results can therefore be achieved when the acquisition process is [5] S. Dong, T. Höllerer, Real-time re-textured geometry modeling using microsoft
hololens, in: 2018 IEEE Conference on Virtual Reality and 3D User Interfaces
constantly slow and the objects are acquired from a distance of
(VR), 2018, pp. 231–237.
0.5 to 1 m. Furthermore, the method of the sensor data fusion pre- [6] M. Bern, P. Plassmann, Mesh Generation, in: Handbook of Computational
sented in this paper is predicated on the availability of all the data Geometry, Elsevier, 2000, pp. 291–332.
of the acquisition process. The reason for this is that our system [7] M. Garon, P.-O. Boulet, J.-P. Doiron, L. Beaulieu, J.-F. Lalonde, Real-time high
resolution 3D data on the HoloLens, in: 2016 IEEE International Symposium on
setup enables access to the different data only after the acquisition Mixed and Augmented Reality (ISMAR-Adjunct), 2016, pp. 189–191.
process has been completed. Another interesting approach for [8] M. Weinmann, S. Wursthorn, M. Weinmann, P. Hübner, Efficient 3D Mapping
future investigations is therefore the modification of the system and Modelling of Indoor Scenes with the Microsoft HoloLens: A Survey, PFG-
Journal of Photogrammetry, Rem. Sens. Geoinformat. Sci. 89 (2021) 319–333.
setup for enabling a sensor data fusion between the data of the [9] P. Hübner, K. Clintworth, Q. Liu, M. Weinmann, S. Wursthorn, Evaluation of
MHL2 and the FLIR Boson ADK thermal camera in real-time. In HoloLens tracking and depth sensing for indoor mapping applications, Sensors
addition, an extended investigation regarding the possibility to 20 (2020) 1021.
[10] G. Loporcaro, L. Bellamy, P. McKenzie, H. Riley, Evaluation of Microsoft
capture and store numerical temperature values as well as their HoloLens augmented reality technology as a construction checking tool
integration into the generated thermal point cloud represents a (2019).
promising approach for future work. [11] K. Khoshelham, H. Tran, D. Acharya, INDOOR MAPPING EYEWEAR:
GEOMETRIC EVALUATION OF SPATIAL MAPPING CAPABILITY OF HOLOLENS,
International Archives of the Photogrammetry, Remote Sensing & Spatial
Data availability Information Sciences (2019).
[12] I. Soares, R.B. Sousa, M. Petry, A.P. Moreira, Accuracy and repeatability tests on
HoloLens 2 and HTC Vive, Multimodal Technol. Interact. 5 (2021) 47.
The data that has been used is confidential. [13] S. Kubo, Y. Iwasawa, M. Suzuki, Y. Matsuo, UVTON: UV Mapping to Consider
the 3D Structure of a Human in Image-Based Virtual Try-On Network, in:
Declaration of Competing Interest Proceedings of the IEEE/CVF International Conference on Computer Vision
(ICCV) Workshops, 2019.
[14] S. Teruggi, E. Grilli, F. Fassi, F. Remondino, 3D Surveying, Semantic Enrichment
The authors declare that they have no known competing finan- and Virtual Acces Of Large Cultural Heritage, in: ICOMOS/ISPRS International
cial interests or personal relationships that could have appeared Scientific Committee on Heritage Documentation (CIPA) 28th CIPA
Symposium ‘‘Great Learning & Digital Emotion”, 2021, pp. 155–162.
to influence the work reported in this paper. [15] S. Fiedler, P. Grümpel, G. Karl, S. Knoblach, 3D-Punktwolke mit
thermografischer Information, zfv-Zeitschrift für Geodäsie, Geoinformation
Acknowledgments und Landmanagement (2019) 144.
[16] S. Vidas, P. Moghadam, HeatWave: A handheld 3D thermography system for
energy auditing, Energ. Buildings 66 (2013) 445–460.
The authors gratefully acknowledge the financial support of the [17] D. Borrmann, A. Nüchter, M. Ðakulović, I. Maurović, I. Petrović, D. Osmanković,
German Federal Ministry for Economic Affairs and Climate Action J. Velagić, A mobile robot based system for fully automated thermal 3D
mapping, Adv. Eng. Inf. 28 (4) (2014) 425–440.
in the project ‘‘EnOB: EnergyTWIN - Energiediagnosestecker Digi-
[18] Microsoft, HoloLens 2: Technische Daten und Funktionen – Microsoft
taler Zwilling. Neue sensorgestützte und KI-basierte Methoden HoloLens 2, 2022. https://www.microsoft.com/de-de/d/hololens-2/
für die digitale, BIM-basierte Inbetriebnahme von technischen 91pnzzznzwcp?activetab=pivot:%C3%BCbersichttab (accessed 10 September
Anlagen in Hochbauwerken und deren energetische Systemopti- 2022).
[19] Microsoft, HoloLens 2 hardware, 2022. https://learn.microsoft.com/en-us/
mierung” (reference number 03EN1026). hololens/hololens2-hardware (accessed 15 September 2022).
[20] D. Ungureanu, F. Bogo, S. Galliani, P. Sama, X. Duan, C. Meekhof, J. Stühmer, T.J.
References Cashman, B. Tekin, J.L. Schönberger, others, Hololens 2 research mode as a tool
for computer vision research, arXiv preprint arXiv:2008.11239 (2020).
[21] Microsoft, HoloLens-Forschungsmodus - Mixed Reality, 2022. https://learn.
[1] R. Becker, C. Blut, C. Emunds, J. Frisch, D. Heidermann, T. Kinnen, A. Lenz, M.
microsoft.com/de-de/windows/mixed-reality/develop/advanced-concepts/
Möller, Pauen, Nicolas, T. Rettig, D. Schlütter, Wenthe, Matthias, J.
research-mode (accessed 23 September 2022).
Blankenbach, G. Bleimann-Gather, J. Fütterer, J. Jungedeitering, C. van Treeck,
[22] D.J. Yeong, G. Velasco-Hernandez, J. Barry, J. Walsh, Sensor and sensor fusion
BIM-assisted, automated processes for commissioning in building services
technology in autonomous vehicles: A review, Sensors 21 (2021) 2140.
engineering, in: Proceedings of the 39th International Symposium on
[23] Z. Zhang, A flexible new technique for camera calibration, IEEE Trans. Pattern
Automation and Robotics in Construction, International Association for
Anal. Machine Intell. 22 (2000) 1330–1334, https://doi.org/10.1109/
Automation and Robotics in Construction (IAARC), 2022.
34.888718.
[2] P. de Wilde, W. Tian, Identification of key factors for uncertainty in the
[24] G. Bradski, A. Kaehler, Learning OpenCV: Computer vision with the OpenCV
prediction of the thermal performance of an office building under climate
library, O’Reilly Media Inc, 2008.
change, Build. Simul. 2 (2009) 157–174, https://doi.org/10.1007/s12273-009-
[25] J. Blankenbach, G. Bleimann-Gather, energyTWIN - accessed 5 October 2022
9116-1.
Project Website (2022). https://energytwin.org/.
[3] R. Becker, E. Lublasser, J. Martens, R. Wollenberg, H. Zhang, S. Brell-Cokcan, J.
[26] Riegl, RIEGL VZ-400 Datasheet.
Blankenbach, Enabling BIM for Property Management of Existing Buildings
[27] J. Bureick, S. Vogel, I. Neumann, D. Diener, H. Alkhatib, Georeferenzierung von
Based on Automated As-is Capturing, in: Proceedings of the 36th International
unmanned aerial systems über laserscannermessun-gen und 3d-
Symposium on Automation and Robotics in Construction (ISARC), Banff, AB,
gebäudemodelle, DVW-Schriftenreihe 96 (2019) 63–74.
Canada, International Association for Automation and Robotics in Construction
[28] C. Ricolfe-Viala, A. Esparza, The Influence of Autofocus Lenses in the Camera
(IAARC), 2019.
Calibration Process, IEEE Trans. Instrum. Meas. 70 (2021) 1–15, https://doi.org/
[4] F. Bosché, M. Ahmed, Y. Turkan, C.T. Haas, R. Haas, The value of integrating
10.1109/TIM.2021.3055793.
Scan-to-BIM and Scan-vs-BIM techniques for construction monitoring using
laser scanning and BIM: The case of cylindrical MEP components, Autom.
Constr. 49 (2015) 201–213.

14

View publication stats

You might also like