41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

3D reconstruction of a bridge with concrete damage

classification using deep learning
Christopher Núñeza, Marck Regalado a, Luis Huaypara, Antonio Bedóna, Jordan Romanía, Oscar
Solísa, Karol Agredaa and Micaela Cárdenasb

a
Faculty of Civil Engineering, National University of Engineering, Perú
𝑏
Faculty of Civil Engineering, National University of Trujillo, Perú
E-mail: christopher.nunez.v@uni.pe, marck.regalado.e@uni.pe, luis.huaypar.a@uni.pe, antonio.bedon.r@uni.pe,
jordan.romani.c@uni.pe, oscar.solis.g@uni.pe, karol.agreda.e@uni.pe and t534000120@unitru.edu.pe

Abstract In recent years, notable bridge collapses attributed to

The classification of concrete damage in bridges various factors have occurred, including the Morandi
poses challenges, characterized by time-consuming, cable-stayed bridge in Genoa, Italy (2018), the Florida
hazardous, and often subjective inspection methods. International University Pedestrian Bridge (2018), and
Recognizing the need for efficient damage the Nanfang’ao steel single-arch bridge in Taiwan (2019)
identification and the creation of 3D models for [4]. Taiwan, with over 28,000 bridges, faces challenges
maintenance purposes, this paper introduces an due to natural disasters like earthquakes, typhoons, and
innovative approach to the inspection of reinforced rainstorms, along with its unique geographic features,
concrete bridges. The proposed methodology such as mountainous and stream-filled regions. Over the
involves 3D reconstruction of a bridge, coupled with past decades, several bridge collapses have been
a concrete damage classification system based on attributed to various scour issues, including general
severity. Notably, the analysis ensures objectivity scour and local scour [3].
through the implementation of deep learning for Concrete structures are increasingly experiencing
classifying concrete damage in UAV-captured deterioration due to various factors such as aging,
images. A noteworthy aspect of this research is that, increased traffic loads, loadings from earthquakes and
in the training models, a precision of over 90% is extreme weather conditions (wind, ambient vibrations)
achieved for each type of concrete damage. This [4,5,7–9], such deterioration is usually caused by
methodology serves as a valuable contribution to inadequate or untimely maintenance [1] Similarly, the
automating and streamlining concrete bridge prevalent structural configuration employed for bridges
inspections, aiming to reduce costs and enhance has been the multispan design with simply supported
efficiency throughout its life cycle. Reinforced Concrete (RC) girders. These girders are
interconnected by transverse beams and are supported
Keywords – by piers and abutments at the ends, facilitating
UAV; Bridge; Concrete damage; SHM; CNN construction phases [1]. Consequently, while the
emergence of cracks in concrete bridges is a crucial
indicator of structural performance, it doesn't
1 Introduction necessarily signify imminent collapse; however, it
In contemporary society, bridges hold significant frequently results in diminished structural integrity
environmental significance by establishing vital [8,10]. Additionally, concrete damaged impact various
connections between various localities. They play a aspects of structures, including bearing capacity,
crucial role in fostering economic and cultural stiffness, energy absorption capacity, and resistance to
development, among other societal factors [1–3]. reinforcement corrosion [8,9,11]. Therefore, timely
Therefore, preserving their structural integrity, safety, detection and measurement of damaged elements are
and functionality is paramount throughout their lifespan, crucial for making informed decisions regarding
especially for emergency situations such as resource necessary repairs and maintenance [8]. However,
transportation and rescue operations [3,4]. conventional inspection methods often prove laborious,
Currently, numerous countries, particularly those in time-consuming, and capital-intensive. Especially, in
Europe and North America, confront the deterioration of the case of large span bridges, traditional methods are
an extensive infrastructure inventory that has surpassed not effective for rapid full-field monitoring and hence a
its originally planned service life [1]. In the United radical monitoring approach is most needed [12].
States, the report states that 7.5% of bridges are During recent decades, ensuring life safety and the need
categorized as structurally deficient, predominantly to reduce inspection costs have emerged as the top
falling below the established standards [5]. In fact, a priorities for practicing engineers and researchers [5].
minimum of one-third of the over 600,000 bridges in the Compared to the traditional contact methods, most
United States feature a concrete superstructure or of the noncontact Structural Health Monitoring (SHM)
wearing surface [6]. methods have improvement in the convenience and

722
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

efficiency of structural inspection and concrete damage context of concrete damage classification.
[9]. SHM is a data processing approach that employs The paper is organized as follows: In Section 2, we
technology to offer early signals of disruption and the delve into the Background and Related Studies. Section
progression of damages and deterioration to avert 3 provides the methodology for obtaining the 3D
potentially hazardous results to a specific structure, reconstructed case study with the classification of the
which is useful for enhanced decision-making [1,7]. The damaged concrete using deep learning. The findings of
majority of existing SHM system consists of various our study are presented in Section 4, which covers the
sensors and accelerometers [7,13,14]. Results. Finally, Section 5 presents the scientific
Consequently, an increasingly number of SHM contribution and conclusions of this paper.
applications with noncontact means have been
developed and applied for the monitoring and inspection
of concrete damage in a variety of critical concrete 2 Background and Related studies
structures [9]. The deployment of Unmanned Aerial
Vehicles (UAV) for civil infrastructure monitoring is a 2.1 Concrete Damage classification
relatively recent development, with only a limited During the last three decades, there has been notable
number of practical case studies conducted for expansion in the utilization of high-strength concrete
industries, monuments, and other civil structures. applications in bridge construction [16]. The five most
Typically, UAVs are equipped with an image prevalent damage, as outlined in Table 1, encompass
acquisition system, and the captured data are manually
cracks, corrosion, efflorescence, spalling, and exposed
processed [1,7,12]. While this approach is convenient, it
steel reinforcement [17] . To establish a ranking system
involves labor-intensive efforts in quantifying and for the magnitude of primary damages in reinforced
analyzing the acquired data [12]. concrete, reference is made to Hüthwohl et al. [18] and
In this paper, we propose a methodology for the Highways England [19] particularly its document "CS
classification of damage in the concrete components of 450 Inspection of Highway Structures." This is
a bridge, employing Deep Learning (DL) based method
undertaken with the objective of formulating a model to
for image processing. DL serves as an automated categorize concrete damages based on their severity.
solution for predicting and classifying data, with the Hence, we have identified three magnitudes for each
Convolutional Neural Network (CNN) emerging as a concrete damage, serving as benchmarks to evaluate
prevalent and straightforward method, particularly well- bridge inspection damage classification methods: (1)
suited for image classification tasks. CNN excels in Not Found (No color), (2) Moderate Damage (Orange
processing and categorizing information presented in
color), and (3) Critical Damage (Red color). In instances
image formats, making it a widely adopted approach in where multiple damage types coexist, the color
the realm of deep learning [15]. The images utilized in corresponding to the highest magnitude will be applied.
this analysis were captured by an Unmanned Aerial As such, in the present article, neither the location of the
Vehicle (UAV). This approach leverages advanced damage nor the combination of types of damage in
algorithms to categorize and analyze the identified
concrete were taken into consideration when calculating
damaged elements, contributing to a nuanced
the magnitude. The focus was solely on the damage
understanding of structural integrity. The integration of itself in a 288x288-pixel image. Additionally, due to the
UAV technology and CNN method in the proposed variability in image quality, it hinders the analysis of
methodology offers a sophisticated and efficient means combinations of damage in concrete.
of assessing the condition of concrete elements in bridge
infrastructure. This method extracts features of different
abstract levels and maps raw pixel intensities of the
2.2 Relevant studies
crack patch into a feature vector through several fully The important aspect of the research presented in
connected layers. All convolutional filter kernel this paper lies in the utilization of computer vision to
elements are trained from the data in a supervised identify defects and damages, thereby establishing an
fashion, learning from the labeled set of examples. This objective classification process.
approach not only leverages advanced aerial imaging
capabilities but also employs CNN to automatically
extract and analyze features from the captured images,
enhancing the precision of structural assessment in the
Table 1 Damage scale for reinforced concrete.
Damage Scale (Color reference)
Damage
No damage (No color) Moderate Damage (Orange) Major Damage (Red)
Corrosion No signs of corrosion attack Moderate corrosion attack Major corrosion attack
No signs of cracks or difficult Cracks less than 1 mm Cracks more than 1 mm (easily
Crack
to detect visually (difficult to detect visually) visible)
No signs of efflorescence
Efflorescence Moderate efflorescence attack Major efflorescence attack
attack
Exposed Bars No exposed bars Moderately exposed bars Fully exposed bars
Spallation No spalls Minor deep spalls exposing Collapsed

723
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

Mansuri & Patel [20] devised an automated visual Table 2 DJI Phantom 4 V2. technical specification.
inspection system for defect detection in heritage
Unmanned Aerial DJI Phantom 4 V2
structures, leveraging artificial intelligence through an
Vehicle (UAV)
R-CNN (Faster Region-based Convolutional Neural
Network) object detection model. The inspection Dimension (mm) 289.5x289.5x196
accuracy of this model demonstrated optimal detection Weight (kg) 1.375
precision, reaching 91.58%, particularly in identifying Max flight time Approx. 30 minutes
three damage types: "spalling," "exposed bricks," and Hovering flying accuracy
Vertical: ±0.1m
"cracks." (with GPS and vision
Horizontal: ±0.3m
Zhao et al. [21] conducted a three-dimensional system)
reconstruction based on images captured by unmanned Max. speed (mph) 45
aerial vehicles for the monitoring and inspection of 5472x3648
Photo resolution
dams, focusing on the identification of damages in the pixels
obtained images. Subsequently, non-contact optical Sensor 1” CMOS
Camera
measurements were performed for disaster prevention. Pixels 20 Mega
Wang et al. [1] suggested a UAV-based method to Remote range (km) 8
promptly evaluate seismic risk in bridges. Their
methodology encompassed the acquisition of aerial
photogrammetric data and the automated extraction of
geometric features, subsequently integrated into
structural models to assess seismic risk in relation to
capacity-demand. The feasibility of their approach was
substantiated through a case study conducted on an
Italian bridge, thus contributing valuable insights to the
field of seismic risk assessment for infrastructure.
Nappo et al. [22] proposed the utilization of
Unmanned Aerial Vehicles (UAVs) for the semi-
automatic detection and classification of damages in
asphalt-paved roads affected by landslides. Leveraging
3D models and 2D images derived through UAV-based Fig. 1. UAV used for the data acquisition
photogrammetry, the approach aimed to overcome the
limitations associated with traditional visual inspections. 3.2 Data acquisition
The developed semi-automatic procedure quantitatively The task of data acquisition includes (i) site
identified and classified longitudinal and transverse prechecking of the bridge and surrounding area, (ii)
cracks in the pavement, presenting a swift, systematic, flight plan drafting, and (iii) on-site data collection.
and objective alternative to conventional field surveys. Firstly, an examination of the area should be conducted
Applied in the Province of Como, Northern Italy, the to consider multiple factors such as the complexity of
results underscored the methodology's utility for road the surroundings and the visibility of the structure (e.g.,
management, providing maps of damage hotspots, presence of obstacles/vegetation around the structure),
pavement damage detectors, criteria based on the the accessibility of the area to identify suitable take-off
International Roughness Index (IRI), and road damage and landing points, and flight restrictions according to
severity maps. local regulations (e.g., the presence of no-fly zones, the
highest permissible altitude for flying, etc.). Then, the
planning of the flight path exerts the most significant
3 Methods and Implementations influence on data quality, as it pertains to factors such as
lighting conditions, camera angles, offset distances, the
3.1 Selected UAV flight pattern chosen, and the level of overlap between
The aerial survey was performed using a commercial images [1]. Additionally, to explain the relationship
quadrotor, Phantom 4 V2. (DJI, China). This portable between camera angle and distance, the terminology
UAV has a built-in GPS that is used for way-point ground sampling distance (GSD) is referred to the ratio
navigation and dataset geotagging. The drone was between the measure of an actual object to a pixel size
deployed for visual inspections of various structural [1,23,24], which is used here to describe the image
components across four reinforced concrete bridges. quality.
Figure 1 illustrates the implementation of the proposed Our work is motivated by a project aiming at the
methodology outlined in this paper for bridge inspection damage classification in concrete bridge elements
utilizing a drone. The specifications of both the vehicle within a millimeter (1 mm) of accuracy, which is
and the camera, outlined in Table 2, play a crucial role considered adequate following previous research such
in the image processing carried out by the deep learning as Chen et al. [23] who show a relationship between
model. GSD and Working Distance (WD) for the DJI phantom
4 UAV with respect to the tilt angle and offset distance.

724
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

The overlap (OV) represents the percentage of the

object covered in two consecutive frames, depending on
the distance to the center of two consecutive photos.
The value of OV is one of the parameters for
determining the fly path in the vertical and horizontal
directions. Even though overlapping rates are seldom
documented and seem to be chosen based on empirical
observations, in Chen et al. [23] it is recommended 60%
±5% for endlap and 30% ±15% for sidelap. In this
work, a 60% OV vertical and 30% OV horizontal
direction should be considered in the to ensure the
survey quality.
After establishing the Ground Sample Distance
(GSD), overlap (OV), an appropriate working distance
(WD), and tilt angle that align with the research
objectives for image acquisition, we then move forward Fig. 3. Image acquisition process of the structural
to elucidate the procedure for capturing imagery of the column of a concrete bridge.
structural elements to create 3D models, as detailed in
the subsequent subsection.
3.2.1 Super-structure
Linear trajectories parallel to the longitudinal
direction of the bridge and outer beams were employed,
adjusting the inclination angle to ensure comprehensive
coverage of the structural beam (See Fig. 2).

Fig. 4. Image acquisition process of concrete bridge

abutments.

3.3 Imagen processing algorithms

The images were captured at a resolution of
5472x3648 pixels from a dataset comprising four
Fig. 2. Image acquisition process of a concrete bridge distinct bridges. The selection of these bridges was
structural beam. based on a non-probability sampling approach,
considering their proximity to the designated study area,
3.2.2 Sub-structure The training dataset incorporates images from three of
these bridges, totaling 848 images. Additionally, to
In the case of the pillars (see Fig. 3), linear
enhance the performance of our training model, we have
trajectories are executed, aligning with the lateral
incorporated another dataset provided by Hüthwohl et al.
elevation profile of the bridge at varying heights and
[18]. As part of the validation dataset, we have selected
employing different camera pitch and yaw angles
the Iniche bridge (refer to Fig. 5) as a case study,
(including upward orientation for mapping beneath the
yielding a total of 142 images.
deck). Where deemed safe, additional spiral (or point of
interest) flights along the height of each pillar are
conducted to ensure comprehensive coverage of all
surfaces. As for the abutments (see Fig. 4), a parallel
survey will be conducted along the primary faces,
adjusting both the height and inclination angles for a
thorough inspection.

Fig. 5 Case study: Iniche bridge. Chiclayo, Perú.

725
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

This implies that a higher number of neurons require Table 3 Number of samples used in the training model.
more parameters to be learned, increasing both the
Training Samples
learning time and complexity of the model. To mitigate Concrete
the size of the CNN model, the original image is resized No Moderate Major
Damage
to a smaller image, specifically 228x228 pixels. damage Damage Damage
Subsequently, 384 images, each measuring 228x228 Corrosion 544 690
pixels, were generated for every original image of Cracks 4609 6998
dimensions 5472x3648 pixels. Regarding the image Efflorescence 2180 1196 1094
quality, it is important that the images have a good Exposed Bars 289 766
resolution, such as HD or 4K, since the reduced image Spalling 1310 1592
used to analyze bridge damage (228 x 228 pixels) may 3.3.3 Data Classification
present noise that generates incorrect image treatment in
the model. Additionally, the normalization of the pixels In this section, a function is developed to manage
should be between 0 and 1, following the image the loading, resizing, and normalizing each grayscale
binarization criterion based on Elizondo et al. [25]. image to fit the required input dimensions of the model.
From the case study, 142 images of 5472x3648 pixels
3.3.1 Data Preparation were obtained from which only 89 images were selected
In this paper, we employ a deep convolutional neural for the classification process.
network based in Zhang et al., [26]. In the data The subsequent iteration through each image file
preparation phase, we implemented a code snippet using involves making predictions using the loaded model.
the TensorFlow Keras API. This code segment focuses Additionally, contours are detected in the original image,
on organizing and structuring the input data for model and based on the predicted class, they are highlighted
training. Subsequently, the images are reshaped into the with semi-transparent colors to emphasize the severity
appropriate format for TensorFlow, ensuring of concrete damage.
compatibility by adjusting the dimensions to (number of Notably, the code incorporates error handling to
images, width, height, channels). The reshaped data is ensure the successful loading and preparation of images.
then stored in an HDF5 file format using the h5py Furthermore, contours are drawn on images to visually
library, providing an efficient and compressed represent the detected damage patterns, contributing to a
representation of the dataset. more comprehensive analysis. The entire process is
geared towards automating the classification of concrete
3.3.2 Data Training damage, making it a valuable tool for efficient structural
In the data training phase, we utilized the health assessment.
TensorFlow Keras API to construct a Convolutional
Neural Network (CNN) architecture for the 3.4 3D Reconstruction
classification of concrete damage. The input data, stored The initial step involves the application of
in an HDF5 file format, is loaded, and pre-processed, colorization to highlight the severity of the damage in
including resizing the images to a standardized each 228x228-pixel image. Subsequently, a meticulous
dimension of 228x228 pixels and normalizing pixel merging process is undertaken to reconstruct these
values between 0 and 1. The CNN model is designed as images to their original resolution of 5472x3648 pixels.
a sequential stack of layers, starting with a This merging process is pivotal, as resizing each
convolutional layer with 16 filters, followed by max- image independently would result in the loss of critical
pooling, batch normalization, and dropout for georeferencing information. This information, including
regularization. This pattern is repeated with additional Latitude, Altitude, Longitude, Focal length, orientation,
convolutional layers, each increasing the number of and other metadata, is embedded in the data provided by
filters. The final layer is a dense layer with softmax the drone for each image. Maintaining the integrity of
activation, representing the three categories of damage this georeferencing data is essential for the accuracy of
scale (See Table 1). During training, the model is fed the final 3D spatial reconstruction.
with the pre-processed images, aiming for 30 epochs Following the image merging, each reconstructed
with a batch size of 64 and a validation split of 20%. image is meticulously reassigned its corresponding
The trained model is then saved for subsequent use. metadata. This involves a comprehensive analysis and
This architecture combines convolutional and pooling adjustment to ensure that the geospatial information
layers with normalization and dropout techniques, aligns accurately with the reconstructed visual data.
demonstrating its potential for accurate concrete damage To achieve a technically robust 3D reconstruction,
classification in structural inspection applications. specialized software designed for photogrammetric
Regarding the number of images in the training model, processing of digital images is employed. Notably,
this is summarized in the following Table 3. As such, it Agisoft Metashape stands out as a prominent example
is important to note that from the set of photos, each of such software. This software employs advanced
image was manually labeled according to the concrete algorithms and techniques to process the merged images
damage. and generate precise 3D spatial data.

726
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

4 Results
Table 5. Classification results for concrete damage in a
4.1 Data Training 288x288 pixel example photograph

The key metrics during the training of the deep Damage

Concrete Example
learning model are summarized in Table 4. Classification
Damage image
(Level of confidence)
Table 4. Loss and accuracy of the training model
Training Validation
Concrete
Epoch Loss Loss
Damage
(Accuracy) (Accuracy)
1.0573 4.4956 Corrosion
1/30
(0.7799) (0.6442)
Corrosion
0.0928 3.3723
30/30
(0.9766) (0.5447)
Major (59.74%)
1.0725 9.1893
1/30
(0.7264) (0.1653)
Cracks
0.1149 0.6825
30/30
(0.9625) (0.8677)
1.5263 1.0956
1/30
(0.6063) (0.3747)
Efflorescence Cracks
0.1526 2.6873
30/30
(0.9452) (0.6756)
0.9333 2.0844
1/30
(0.8396) (0.6909) No color (100.00%)
Exposed Bars
0.0388 1.1794
30/30
(0.9884) (0.8779)
1.4644 1.9797
1/30
(0.6148) (0.4435)
Spalling
0.1393 2.9182
30/30
(0.9542) (0.5280) Efflorescence

4.2 Data Classification (228x228

pixels)
Given the extensive dataset, we present a table Moderate (98.06%)
displaying the classification of select images featuring
damaged concrete along with their corresponding
confidence levels.
In Table 5, the classification results of concrete
damage for a photograph are illustrated. It can be
observed that, in the case of the example, after Exposed Bars
processing, only corrosion, efflorescence, and spalling
damage are noticeable. Similar results are obtained for
all other photographs based on the level of damage they No color (88.44%)
exhibit.

4.3 3D Reconstruction

After the classification process, the 228x228 pixels

colored images are combined with their counterparts Spalling
from the original image, which is 5472x3648 pixels
(See Fig. 6). Subsequently, the metadata from the
original image is transferred to the reconstructed image.
Major (99.88%)

727
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

Addressing the loss of georeferenced information

after resizing images is achieved through three-
dimensional reconstruction, where the original metadata
is transferred to the reconstructed images. This
highlights the importance of integrating geospatial data
for a more comprehensive and accurate assessment of
infrastructure.
The main limitations of this method were that it
does not assess the quality of the images due to their
reduction in size. Additionally, the classification of the
magnitude of the damage was partially subjective. Other
limitations included the acquisition of data, where
Fig. 6 Reconstructed image with concrete damage external conditions to the infrastructure, such as the
classification flow of a river, varied the quality of the images.

Ultimately, utilizing these 89 reconstructed images

from the case study, we proceed to generate a point References
cloud and perform the 3D reconstruction of the bridge
[1] Wang X, Demartino C, Narazaki Y, Monti G,
(See Fig. 7).
Spencer BF. Rapid seismic risk assessment of
bridges using UAV aerial photogrammetry. Eng
Struct. 2023 Mar 15;279:115589.
[2] Scattarreggia N, Salomone R, Moratti M, Malomo
D, Pinho R, Calvi GM. Collapse analysis of the
multi-span reinforced concrete arch bridge of
Caprigliola, Italy. Eng Struct. 2022 Jan
15;251:113375.
[3] Yau NJ, Tsai MK, Wang HL, Hung DM, Chen CS,
Hsu WK. Improving bridge collapse detection and
on-site emergency alarms: A case study in Taiwan.
Saf Sci. 2014 Dec 1;70:133–42.
[4] Mandirola M, Casarotti C, Peloso S, Lanese I,
Brunesi E, Senaldi I. Use of UAS for damage
inspection and assessment of bridge infrastructures.
International Journal of Disaster Risk Reduction.
2022 Apr 1;72:102824.
[5] Azimi M, Eslamlou AD, Pekcan G. Data-Driven
Fig. 7 3D reconstruction of the Iniche bridge with Structural Health Monitoring and Damage
concrete damage classification Detection through Deep Learning: State-of-the-Art
Review. Sensors [Internet]. 2020;20(10). Available
from: https://www.mdpi.com/1424-
5 Conclusions and discussions 8220/20/10/2778
The proposed methodology demonstrates [6] Dorafshan S, Thomas RJ, Maguire M. Comparison
efficiency in bridge inspections by combining three- of deep convolutional neural networks and edge
dimensional reconstruction with concrete damage detectors for image-based crack detection in
classification using deep learning. This suggests an concrete. Constr Build Mater. 2018 Oct
effective and automated alternative for assessing bridge 20;186:1031–45.
conditions, minimizing time and costs associated with [7] Payawal JMG, Kim DK. Image-Based Structural
traditional methods. Health Monitoring: A Systematic Review. Applied
The implementation of deep learning in concrete Sciences [Internet]. 2023;13(2). Available from:
damage classification ensures an objective approach. https://www.mdpi.com/2076-3417/13/2/968
The model's ability to accurately identify and categorize [8] Nyathi MA, Bai J, Wilson ID. Concrete Crack
defects in concrete from UAV-captured images suggests Width Measurement Using a Laser Beam and
a significant improvement in result objectivity Image Processing Algorithms. Applied Sciences
compared to conventional inspection methods. [Internet]. 2023;13(8). Available from:
Upon reviewing the outcomes derived from the https://www.mdpi.com/2076-3417/13/8/4981
228x228 pixel images, it is evident that a more [9] Zhang Z, Zhou H, Wang S, Xu C, Lv Y. Design
extensive dataset is essential for refining the training and Research of Low-Cost and Self-Adaptive
model. Also, due to the use of an external database, the Terrestrial Laser Scanning for Indoor
GDS has been variable which affects the quality of the Measurement Based on Adaptive Indoor
images and therefore affects the training database. Measurement Scanning Strategy and Structural

728
 41st International Symposium on Automation and Robotics in Construction (ISARC 2024)

Characteristics Point Cloud Segmentation. Han J, [22] Nappo N, Mavrouli O, Nex F, van Westen C,
editor. Advances in Civil Engineering [Internet]. Gambillara R, Michetti AM. Use of UAV-based
2022;2022:5681771. Available from: photogrammetry products for semi-automatic
https://doi.org/10.1155/2022/5681771 detection and classification of asphalt road damage
[10] Zoubir H, Rguig M, El Aroussi M, Chehri A, in landslide-affected areas. Eng Geol. 2021 Dec
Saadane R. Concrete Bridge Crack Image 5;294:106363.
Classification Using Histograms of Oriented [23] Siyuan C, F LD, Eleni M, Iman ZSM, Jonathan B.
Gradients, Uniform Local Binary Patterns, and UAV Bridge Inspection through Evaluated 3D
Kernel Principal Component Analysis. Electronics Reconstructions. Journal of Bridge Engineering
(Basel) [Internet]. 2022;11(20). Available from: [Internet]. 2019 Apr 1;24(4):05019001. Available
https://www.mdpi.com/2079-9292/11/20/3357 from: https://doi.org/10.1061/(ASCE)BE.1943-
[11] Li J, Li X, Liu K, Yao Z. Crack Identification for 5592.0001343
Bridge Structures Using an Unmanned Aerial [24] Atencio E, Plaza-Muñoz F, Muñoz-La Rivera F,
Vehicle (UAV) Incorporating Image Geometric Lozano-Galant JA. Calibration of UAV flight
Correction. Buildings [Internet]. 2022;12(11). parameters for pavement pothole detection using
Available from: https://www.mdpi.com/2075- orthogonal arrays. Autom Constr. 2022 Nov
5309/12/11/1869 1;143:104545.
[12] Sankarasrinivasan S, Balasubramanian E, Karthik [25] Esqueda Elizondo J, Palafox L. Fundamentos de
K, Chandrasekar U, Gupta R. Health Monitoring Procesamiento de Imágenes. 2005.
of Civil Structures with Integrated UAV and [26] Zhang L, Yang F, Daniel Zhang Y, Zhu YJ. Road
Image Processing System. Procedia Comput Sci. crack detection using deep convolutional neural
2015 Jan 1;54:508–15. network. In: 2016 IEEE International Conference
[13] Bao Y, Chen Z, Wei S, Xu Y, Tang Z, Li H. The on Image Processing (ICIP). 2016. p. 3708–12.
State of the Art of Data Science and Engineering
in Structural Health Monitoring. Engineering.
2019 Apr 1;5(2):234–42.
[14] Ye XW, Dong CZ, Liu T. A Review of Machine
Vision-Based Structural Health Monitoring:
Methodologies and Applications. Oddo CM, editor.
J Sens [Internet]. 2016;2016:7103039. Available
from: ttps://doi.org/10.1155/2016/7103039
[15] Permana SDH, Saputra G, Arifitama B,
Yaddarabullah, Caesarendra W, Rahim R.
Classification of bird sounds as an early warning
method of forest fires using Convolutional Neural
Network (CNN) algorithm. Journal of King Saud
University - Computer and Information Sciences.
2022 Jul 1;34(7):4345–57.
[16] Sagara A, Pane I. A Study on Effects of Creep and
Shrinkage in High Strength Concrete Bridges.
Procedia Eng. 2015 Jan 1;125:1087–93.
[17] Abubakr M, Rady M, Badran K, Mahfouz SY.
Application of deep learning in damage
classification of reinforced concrete bridges. Ain
Shams Engineering Journal. 2024 Jan
1;15(1):102297.
[18] Hüthwohl P, Lu R, Brilakis I. Multi-classifier for
reinforced concrete bridge defects. Autom Constr.
2019 Sep 1;105:102824.
[19] Highways England. CS 450 Inspection of highway
structures. In: 0.1.0. Design Manual for Roads and
Bridges; 2021.
[20] Mansuri L, Patel D. Artificial intelligence-based
automatic visual inspection system for built
heritage. Smart and Sustainable Built Environment.
2021 Feb 3;11:622–46.
[21] Zhao S, Kang F, Li J, Ma C. Structural health
monitoring and inspection of dams based on UAV
photogrammetry with image 3D reconstruction.
Autom Constr. 2021 Oct 1;130:103832.

729