Towards a Low-cost Vision System for Real-time Pavement Condition

Assessment

Kehinde Olufowobi and Nic Herndon

Department of Computer Science, East Carolina University, Greenville, U.S.A.

Keywords:

Pavement Distress Detection, Object Detection, Homography Estimation, Neural Networks.

Abstract:

Although advances in camera and sensing technology in the last decade helped propel the automation of pave-

ment distress detection and characterization, increased equipment acquisition and running costs limit access

to the most effective solutions. Furthermore, some of these advanced techniques require substantial human in-

volvement to process and analyze data correctly. We propose a cost-effective, end-to-end automated approach

to pavement condition assessment that employs a neural object detector to identify and measure instances

of pavement distress in real time from oblique two-dimensional imagery acquired using an unmanned aerial

vehicle. A state-of-the-art object detector architecture is applied to identify and localize pavement distress

instances in these images. Camera data, information about Street View image acquisition conditions, and the

principles of photogrammetry and planar homography are exploited to construct a mapping for translating

pixel distances to real-world distances. This capability is integrated into the neural network inference process

to derive an end-to-end system for real-time distress identiﬁcation and measurement.

1 INTRODUCTION

Pavement condition assessment enables cost-effective

decision-making associated with road maintenance.

These assessments are usually performed using spe-

cialized tools. Although the past decade yielded im-

provements in sensing and camera technology that

propelled the automation of pavement distress detec-

tion and characterization, this also led to increased

costs from the acquisition and operation of advanced

equipment. To reduce pavement condition assess-

ment costs, we propose to leverage advances in im-

age processing, which enable the deployment of cost-

effective, computationally efﬁcient resources in pave-

ment distress monitoring. We apply modern computer

vision techniques to the problem of detecting and

evaluating pavement distress from images taken us-

ing a camera mounted on an unmanned aerial vehicle

(UAV) ﬂying at a low altitude. Our objective is to in-

vestigate the effectiveness of YOLOv5 (Jocher et al.,

2021), a state-of-the-art object detector, for pavement

distress detection. We train and validate this neural

network on a revised Pavement Image Dataset (PID)

(Majidifard et al., 2019). Furthermore, we estimate

the potential asphalt patch area directly from image

https://orcid.org/0000-0001-9712-148X

pixels using techniques from traditional computer vi-

sion, but with some constraints. The contributions of

this study include: A network for low-cost, real-time,

end-to-end pavement distress detection and measure-

ment based on YOLOv5, and an approach to spa-

tially evaluate pavement distress from oblique pave-

ment imagery in a rapid manner using planar homog-

raphy. Figure 1 shows the output of this end-to-end

system.

2 RELATED WORKS

2.1 Pavement Distress Detection

In (Zhang et al., 2016), CrackNet, a convolutional

neural network (CNN) trained using raw image

patches, was used to determine whether an image con-

tained a crack or not. Building on this, CrackNet-

R (Zhang et al., 2019) used a recurrent neural net-

work (RNN) to accurately detect small cracks while

removing noise. Our approach involves using an ob-

ject detection algorithm to optimize the search space.

We ﬁrst extract discriminative features from an image

using a CNN. Then we generate regions of interest

within the image before ﬁnally detecting objects via

526

Olufowobi, K. and Herndon, N.

Towards a Low-cost Vision System for Real-time Pavement Condition Assessment.

DOI: 10.5220/0010785900003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 526-533

ISBN: 978-989-758-549-4; ISSN: 2184-4313

Figure 1: Real-time distress detection and measurement.

The bounding boxes identify individual pavement cracks in

the image and the measurements shown in square meters re-

ﬂect the true physical dimensions of the enclosed regions.

The central idea is to employ Google Street View images as

a proxy for oblique pavement imagery acquired by ﬂying a

UAV at a ﬁxed low altitude over a road section. These im-

ages are then used to train a neural object detector with a

built-in homography estimator to measure detected cracks.

This facilitates a low-cost automated pavement condition

assessment.

bounding box coordinates.

In (Mandal et al., 2018) YOLOv2 (Redmon and

Farhadi, 2017) was proposed to perform pavement

crack detection in automated pavement distress anal-

ysis systems. YOLOv3 (Redmon and Farhadi, 2018)

was adopted in (Nie and Wang, 2019) to achieve im-

proved real-time performance and better detection ac-

curacy. Sarmiento (2021) describes the deployment of

YOLOv4 (Bochkovskiy et al., 2020) in a dual demon-

stration of the effectiveness of deep learning for both

distress detection and segmentation. The efﬁcacy of

YOLOv5 for detecting concealed cracks in asphalt

pavement was recently demonstrated in (Li et al.,

2021). Our study exhibits the potential of YOLOv5.

We determine the optimal experiment conﬁgurations

and judge the model’s overall effectiveness for the

task of distress detection.

Fan et al. (2018) describes the deployment of

a CNN to learn pavement crack textures from raw

images by analyzing crack pixels within an image.

The problem was modeled as a multi-label classi-

ﬁcation problem. In (Jenkins et al., 2018), a U-

Net (Ronneberger et al., 2015) architecture was em-

ployed in implementing an encoder-decoder structure

for semantic segmentation of pavement images. Fur-

thermore, DeepCrack was proposed in (Zou et al.,

2019) as an end-to-end trainable network for auto-

matic crack detection and is based on the SegNet

(Badrinarayanan et al., 2017) encoder-decoder CNN

architecture.

The best-performing segmentation algorithms rely

on an encoder-decoder network architecture for which

a signiﬁcant reduction in resolution due to downsam-

pling is undesirable. Since the vast majority of pave-

ment distress image pixels (more than 95%) are back-

ground (Jenkins et al., 2018), low-quality input im-

ages can easily degrade model performance. These

segmentation techniques also rely on planar view im-

ages for more accurate views of distresses as com-

pared to oblique or wide-view images. However, pla-

nar view images generally require a more sophisti-

cated camera and mounting equipment.

2.2 Pavement Distress Characterization

In (Tan and Li, 2019) an algorithm was developed to

extract the geometric features of pavement distress

after they are automatically detected from road 3D

models constructed from UAV oblique photogramme-

try. Sarsam et al. (2015) describes the implementation

of CRP for evaluating pavement surface distresses to

validate the equivalency of the technique with visual

evaluation. The results of both the photogrammet-

ric and visual techniques were modeled and the pho-

togrammetric approach was found to be highly efﬁ-

cient with a Coefﬁcient of Determination value be-

tween 0.985 and 0.999 compared to the traditional ap-

proach.

Structure-from-motion (SfM) (Ozyesil et al.,

2017) techniques were used for pavement distress

analysis in (Inzerillo et al., 2018). Imagery acquired

using cameras at head level and higher altitudes were

analyzed to determine metric information about size,

shape, and position within the images. In (Zhang and

Elaksher, 2012) 3D reconstruction and measurement

of surface distresses on unpaved roads was performed

on UAV-acquired imagery. The acquired imagery was

subjected to photogrammetric orientation, resulting in

precise orientation parameters.

In general, the analyses furnished by the preced-

ing proposals rely on a combination of advanced and

manual imaging and 3D reconstruction techniques to

address questions regarding varied aspects of pave-

ment health. Most of these techniques were imple-

mented using commercial software. Our analysis fo-

cuses on the spatial quantiﬁcation of the degree of

damage via crack size measurement, to inform asphalt

patch requirements for the repair and maintenance of

a road section. Our proposal enables the aggregation

Towards a Low-cost Vision System for Real-time Pavement Condition Assessment

527

of such patch requirements across a long stretch of

road pavement and thus supports large-scale distress

analysis via straightforward and relatively inexpen-

sive means.

3 METHODS

3.1 Dataset

Our pavement assessment model is deﬁned on 2D

Street View data. The images are used to replicate the

outcome of ﬂying a drone with an onboard camera at

a low altitude to collect oblique pavement imagery.

On the one hand, this warrants that certain orientation

parameters be constrained – altitude, pitch, and ﬁeld

of view (FOV) have to be predetermined and ﬁxed for

the duration of the ﬂight. On the other hand, Street

View images offer the desirable advantages of ease of

acquisition and scale.

It is assumed that these orientation constraints can

be easily replicated for an arbitrary ﬂight. In partic-

ular, to successfully implement our proposal, during

image acquisition, a UAV should maintain an altitude

of 2.5 m, an onboard camera pitch of -70

◦

, and target

FOV should be 90

◦

, as shown in Figure 2. The camera

heading should align with an axis parallel to any pave-

ment centerline markings, and the axes of rotation are

to be kept constant. This guideline is informed by the

camera and image parameters speciﬁed in the Street

View Street API calls used to extract the images.

The primary dataset for our analysis is PID, a

manually-annotated collection of 7,237 640×640 im-

Figure 2: Prescribed orientation parameters. Note that ver-

tical FOV (v) is shown in this 2D diagram instead of hor-

izontal FOV (h). Both values are equivalent when deal-

ing with a 1:1 aspect ratio, as is our case. Note that

h = arctan(aspect ratio ∗ tan(v)) = arctan(tan(v)) = v.

ages of 22 different pavement sections that include

both interstate and US highways (Majidifard et al.,

2019). PID is designed to support simultaneous pave-

ment distress detection and density quantiﬁcation us-

ing various camera views. The images were extracted

using Street View Static APIs by specifying GPS co-

ordinates and camera and image parameters. For each

pavement section, the images were collected at 15-

meter intervals within predetermined start and end-

points.

Since the dataset was created for both distress

classiﬁcation and density determination, two different

views were collected at each coordinate point: pla-

nar view images taken at a pitch of -90

◦

and wide-

view images taken at -70

◦

. The wide-view images

were manually annotated by professionals for nine

distinct distress categories. These categories had been

determined as the most critical distresses that affect

pavement condition. The distribution of images and

bounding boxes across classes is shown in Figure 3.

To enable faster model training and inference,

variants of the same distress type were collapsed into

a single category, effectively reducing the total num-

ber of classes to be predicted from nine to four. Given

the sparsity of potholes across the entire dataset, we

were able to exclude this category from the scope of

our distress detection problem without compromising

the quality of the solution. The revised numeric dis-

tribution of the dataset is shown in Figure 4.

3.2 YOLOv5

YOLOv5 is a state-of-the-art, single-stage, real-time

object detector and one of the latest models in the evo-

lution of the YOLO algorithm. In general, YOLO

models are compact but highly performant, and this

makes them ideal candidates for real-time conditions

and on-device deployment environments.

Figure 5 shows the network architecture of

YOLOv5. The ﬁrst major component is a CNN back-

Figure 3: Sorted PID numeric distribution with overlapping

horizontal bars to show comparison between the number of

bounding boxes and images per class.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

528

bone for extracting image convolutional features at

different granularities from the input RGB image. For

its backbone, YOLOv5 incorporates Cross Stage Par-

tial Network (CSPNet) (Wang et al., 2020) into Dark-

net to create CSPDarknet.

The second major component is a series of lay-

ers for aggregating the image features called the neck.

The detection neck employed to boost information

ﬂow in YOLOv5 is PANet (Wang et al., 2019), a path

aggregation network that adopts the Feature Pyramid

Network (FPN) (Lin et al., 2017) structure to improve

low-level feature propagation.

The third major YOLOv5 component is the pre-

diction head which consumes features from the neck

to make bounding box and class predictions. This is

the ﬁnal layer in the network. To achieve multiscale

prediction, three feature maps of different sizes are

generated. This allows the model to handle variations

in object sizes.

3.3 Homography Estimation

If we use homogeneous coordinates to express the

point

P on an object plane and the corresponding

point ~p on an image plane, then the homography or

projective mapping can be expressed as:

~p = s · H ·

P (1)

where s is an arbitrary scale factor.

We use this result to compute the homography ma-

trix for mapping the road plane – and, by extension,

the pavement distress plane – onto the image plane in

our pavement distress measurement task. To calibrate

H we use correspondences between multiple points

on pavement distress images and control points in the

real world scene. This setup allows us to establish the

rotation and translation associated with the perspec-

tive projection.

Figure 6 illustrates this projective mapping be-

tween a 640×640 image and a quadrilateral depicting

the scope of the object plane of interest – that is, the

pavement view captured in the image. Ground dis-

tance measurements obtained using a reference im-

age show that |A

| = 9.85 m, |A

| = |B

| = 8.35

m, and |C

| = 4.25 m. Thus, the eight vertices

Figure 4: Numeric distribution of image dataset after revi-

sion.

Figure 5: Architecture of YOLOv5 from (Xu et al., 2021).

Features are extracted using CSPDarknet and fused using

PANet. The CSP Bottleneck is used to efﬁciently formu-

late image features. The prediction head generates 18×18,

36×36, and 72×72 feature maps and outputs the detection

results. Source: shorturl.at/cdhAW (licensed under CC BY

4.0).

(four for each plane) used to estimate homography H

are: A(0,0), B(639,0) , C(639, 639), D(0, 639), and

(0,8.35), B

(9.85,8.35), C

(7.05,0), D

(2.8,0).

Note that A, B, C, D are pixel coordinates while A

, C

, D

correspond to metric coordinates. The cam-

era orientation constraints we have deﬁned allow us to

estimate the image-plane-to-road-plane relation using

only homography – that is, without the need for ad-

ditional metadata or complex camera calibration. To

estimate the true size of abcd in the real world, we

simply compute the area of trapezoid a

. This is

given by:

Area of a

| + |c

∗ h (2)

where h is the height of the trapezoid, equal to the

perpendicular distance from |a

| to |c

3.4 Experimental Design

Our experiment was conducted using tools available

from the Utralytics GitHub repository

. We split the

modiﬁed dataset into training and validation sets as

follows: 5,426 training images and a validation set

of 1,811 images, or a 75-25 split, and 6,006 training

images and a validation set of 1,231 images, or an 83-

17 split.

We trained the four YOLOv5 models on the orig-

inal and revised versions of our dataset from scratch.

Training loss in YOLOv5 is a composition of box

loss, object loss, and classiﬁcation loss. The combina-

tion of these losses helps maximize the mAP metric.

https://github.com/ultralytics/yolov5.

Towards a Low-cost Vision System for Real-time Pavement Condition Assessment

529

Figure 6: View of bounding box abcd around a pavement distress as described by the perspective mapping between the road

plane and image planes. The true geometry of the planar view captured in image ABCD is reﬂected in trapezoid A

. Pla-

nar homography enables simultaneous comprehension of the location of both planes and effectively maps ABCD to A

thereby transforming abcd to a

. This task can be interpreted as artiﬁcially correcting for linear perspective in an image.

Input size was ﬁxed but we varied batch size, epochs,

optimizer, and batch normalization across the many

training experiments performed. The weight optimiz-

ers used were Stochastic gradient descent (SGD) and

Adam.

We compared the results obtained by training four

YOLOv5 models (small, medium, large, extra-large)

using batch sizes proportionate to the model sizes.

This allowed us to determine the most effective mod-

els for our dataset: the large and extra-large. We

focused on these two models and iterated through

more than twenty experiment conﬁgurations by vary-

ing batch size, dataset version and split, optimizer se-

lection, and training time.

We also observed the effect of ﬁne-tuning a pre-

trained model, batch normalization, and a combina-

tion of batch normalization and multi-GPU training.

Table 1 shows the experiment conﬁgurations adopted

in the tuning process. To realize a truly end-to-end

system, we integrated the distress measurement task

into the detection network’s inference process. The

code that estimates and applies the homography ma-

trix that maps the view in an image to its real-world

geometry is integrated into the YOLOv5 detection

pipeline.

Not only does our implementation enable real-

time detection of pavement distress, but the actual

area in square meters of the imaginary bounding box

enclosing each physical distress instance is also dis-

played as additional information alongside the bound-

ing box. Example outputs are shown in Figure 7. In-

put sources may be an image, video, or live video

stream.

4 EVALUATION

4.1 Training Metrics

The mAP, precision, recall, and F1 Score were em-

ployed in evaluating the performance of the trained

models on the validation set. The mAP was the pri-

mary index used to measure YOLOv5’s overall detec-

tion effectiveness per experiment conﬁguration. It is

deﬁned as:

mAP =

∑

i=1

AP(c

)

(3)

where C is the number of categories in the decision

space and AP(c

) is the average precision (AP) for a

given category, c

. AP is deﬁned as the area under the

precision vs. recall curve (PR Curve), where preci-

sion is the fraction of predictions that are correct and

recall is the fraction of the ground truth that the model

correctly predicts.

We evaluated each trained model on the validation

set using different IoU thresholds, via the mAP@0.5

and mAP@[.5, .95] metrics. IoU is a similarity metric

obtained by dividing the area of overlap or intersec-

tion between the predicted and ground truth boxes by

the area of their union. By deﬁnition, this is a value

between 0 and 1 that represents the quality of overlap

between the two boxes. The F1 Score measures the

balance achieved between precision and recall. Fig-

ure 8 shows the F1 Score and PR Curve plots for Ex-

periment 1.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

530

Table 1: Training experiment conﬁgurations. The result for Experiment 0 was obtained using hyperparameter values derived

from running a genetic algorithm for 90 generations, with Experiment 1 as the base scenario. The 83-17 dataset split was used

for Experiments 0, 1 and 2; all other runs were based on the 75-25 split. The top four mAP values are shown in bold font.

ID Model Batch Size Epochs Optimizer GPU Hours mAP@0.5 Classes Other

0 YOLOv5l 48 150 SGD 16.9 0.4657 4 –

1 YOLOv5l 48 150 SGD 16.9 0.4555 4 –

2 YOLOv5l 24 150 SGD 18.8 0.4459 4 –

3 YOLOv5l 48 150 SGD 9.5 0.4039 4 –

4 YOLOv5l 48 100 SGD 11.8 0.3984 4 Batch Normalization

5 YOLOv5x 72 200 SGD 18.6 0.3969 4 –

6 YOLOv5l 24 150 SGD 16.8 0.3948 4 –

7 YOLOv5l 48 100 SGD 11.8 0.3712 5 –

8 YOLOv5x 72 1000 Adam 93.6 0.3604 4 –

9 YOLOv5l 48 200 Adam 24.2 0.3545 4 Batch Normalization

10 YOLOv5l 48 500 Adam 33.1 0.3510 4 –

11 YOLOv5l 48 200 Adam 24.2 0.3501 4 –

12 YOLOv5l 48 100 SGD 11.6 0.3497 5 Pretrained

13 YOLOv5l 48 100 Adam 12.0 0.2988 5 –

14 YOLOv5l 48 200 SGD 16.7 0.2929 9 Batch Normalization

15 YOLOv5l 48 200 SGD 18.7 0.2919 9 –

16 YOLOv5x 72 500 SGD 47.4 0.2888 9 –

17 YOLOv5l 72 200 SGD 12.5 0.2861 9 Batch Normalization

18 YOLOv5l 72 500 SGD 30.9 0.2839 9 –

19 YOLOv5l 48 100 Adam 12.2 0.2822 5 Pretrained

20 YOLOv5m 16 100 SGD 9.7 0.2790 9 –

21 YOLOv5l 16 425 SGD 40.3 0.2777 9 –

22 YOLOv5x 12 200 SGD 59.6 0.2733 9 Batch Normalization

23 YOLOv5x 8 150 SGD 19.5 0.2710 9 Batch Norm + Parallelization

24 YOLOv5l 36 200 SGD 14.3 0.2687 9 –

25 YOLOv5m 8 50 SGD 10.9 0.2647 9 –

26 YOLOv5s 8 75 SGD 13.4 0.2562 9 –

27 YOLOv5x 4 200 SGD 67.1 0.2561 9 Batch Norm + Parallelization

28 YOLOv5s 16 75 SGD 10.3 0.2489 9 –

29 YOLOv5x 24 200 SGD 19.9 0.2396 9 Batch Norm + Parallelization

4.2 Discussion

Overall results show that the larger models are

more favorable for our problem context compared to

smaller models. For the same models, SGD achieved

better results compared to Adam optimizer. However,

we cannot assert that training the same models using

Adam optimizer will not eventually lead to better per-

formance if the training time is long enough.

Models trained on the reﬁned dataset generally

performed better. The only exceptions were models

trained using Adam optimizer, which did not show

any performance improvements within a reasonable

amount of time. Also, ﬁne-tuning a model pre-trained

on a previous dataset version was found to result in

worse performance than when the model was trained

from scratch on the reﬁned dataset.

The model appears to be more robust in detecting

block cracks compared to other categories, although

both transverse and longitudinal cracks are better rep-

resented in the dataset. This is explained by the high

intra-class variability associated with both the trans-

verse and longitudinal cracks (due to dataset revi-

sion), as compared with the other two categories. On

the contrary, the block crack class lends itself to low

intra-class variability and relatively balanced repre-

sentation, which the model is more amenable to. This

is not the case with the alligator crack class which has

both low representation and more complex features.

The lack of background images (images with no

cracks) in the dataset led to increased background

False Positives and False Negatives, the effect of

which impacted the model’s ability to predict classes

with complex features and low representation in the

dataset accurately enough.

4.3 Limitations

Our model suffers from a low-quality dataset and in-

sufﬁcient dataset size and scope. PID and the revised

versions we derived from it do not sufﬁciently address

the expectation of context variation. For example,

homogeneous highway scenery typiﬁes the views in

the training images, whereas the images we used dur-

ing inference included views characterized by slightly

more municipal scenes. This limits the model’s gen-

Towards a Low-cost Vision System for Real-time Pavement Condition Assessment

531

Figure 7: Example model outputs showing the real-world measurements for distress regions in square meters. Pavement

images were obtained from Google Street View panorama using a pitch of -70

◦

and FOV of 90

◦

Figure 8: F1 Score (left) and the Precision-Recall Curve (right) for Experiment 1. The maximum overall (blue curve) F1

value of 0.47 was achieved when prediction conﬁdence was 0.27. The mAP@0.5 score for the validation set is the area under

the blue PR Curve.

eralization ability as it is not able to correctly classify

street sidewalks as background.

To effectively estimate the size of a pavement

crack, we have relied on the assumption that image

acquisition would always be conducted under the con-

ditions of -70

◦

pitch, FOV of 90

◦

, and a camera height

of 2.5 m. However, this may not be the case in prac-

tice as drone footage may have been acquired using

the most convenient orientation conditions, depend-

ing on the nature of the permit secured and the road

situation at the point of conducting the exercise. This

limits the kind of pavement distress imagery we can

successfully apply our solution to.

5 CONCLUSIONS

This work proposes a cost-effective method for pave-

ment condition assessment that employs UAVs to ac-

quire road pavement imagery, and an object detector

to detect and measure instances of pavement distress

in oblique images. Although we proposed adopting

UAVs to acquire these images, we suggested an al-

ternative that enables the cost-effective collection of

data associated with initial modeling tasks. We pro-

posed the use of Google Street View images to serve

as a proxy for data collected using a UAV-mounted

camera. In line with this, we also introduced several

constraints that are required for the successful imple-

mentation of such a proposal.

Our proposal represents a preliminary effort in the

application of machine learning to automate costly or

challenging human-dependent processes. This work

can thus be extended by adapting the architecture of

YOLOv5 to the pavement distress data distribution to

create a robust network tailored for distress detection

in images acquired under unconstrained conditions.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

532

REFERENCES

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).

SegNet: A Deep Convolutional Encoder-Decoder Ar-

chitecture for Image Segmentation. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

39:2481–2495.

Bochkovskiy, A., Wang, C.-Y., and Liao, H. (2020).

YOLOv4: Optimal Speed and Accuracy of Object De-

tection. ArXiv, abs/2004.10934.

Fan, Z., Wu, Y., Lu, J., and Li, W. (2018). Automatic Pave-

ment Crack Detection Based on Structured Predic-

tion with the Convolutional Neural Network. ArXiv,

abs/1802.02208.

Inzerillo, L., Di Mino, G., and Roberts, R. (2018). Image-

based 3D reconstruction using traditional and UAV

datasets for analysis of road pavement distress. Au-

tomation in Construction, 96:457–469.

Jenkins, M., Carr, T. A., Iglesias, M. I., Buggy, T., and

Morison, G. (2018). A Deep Convolutional Neu-

ral Network for Semantic Pixel-Wise Segmentation

of Road and Pavement Surface Cracks. 2018 26th

European Signal Processing Conference (EUSIPCO),

pages 2120–2124.

Jocher, G., Stoken, A., Borovec, J., NanoCode012, Chaura-

sia, A., TaoXie, Changyu, L., V, A., Laughing,

tkianai, yxNONG, Hogan, A., lorenzomammana,

AlexWang1900, Hajek, J., Diaconu, L., Marc, Kwon,

Y., oleg, wanghaoyang0106, Defretin, Y., Lohia, A.,

ml5ah, Milanko, B., Fineran, B., Khromov, D., Yiwei,

D., Doug, Durgesh, and Ingham, F. (2021). ultralyt-

ics/yolov5: v5.0 - YOLOv5-P6 1280 models, AWS,

Supervise.ly and YouTube integrations.

Li, S., Gu, X., Xu, X., Xu, D., Zhang, T., Liu, Z., and

Dong, Q. (2021). Detection of concealed cracks from

ground penetrating radar images based on deep learn-

ing algorithm. Construction and Building Materials,

273:121949.

Lin, T.-Y., Doll

ar, P., Girshick, R., He, K., Hariharan, B.,

and Belongie, S. (2017). Feature pyramid networks

for object detection. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 2117–2125.

Majidifard, H., Jin, P., Adu-Gyamﬁ, Y., and Buttlar, W.

(2019). Pavement Image Datasets: A New Benchmark

Dataset to Classify and Densify Pavement Distresses.

Transportation Research Record, 2674:328–339.

Mandal, V., Uong, L., and Adu-Gyamﬁ, Y. (2018). Au-

tomated Road Crack Detection Using Deep Convo-

lutional Neural Networks. 2018 IEEE International

Conference on Big Data (Big Data), pages 5212–

5215.

Nie, M. and Wang, C. (2019). Pavement Crack Detection

based on YOLO v3. 2019 2nd International Con-

ference on Safety Produce Informatization (IICSPI),

pages 327–330.

Ozyesil, O., Voroninski, V., Basri, R., and Singer, A.

(2017). A survey of structure from motion. arXiv

preprint arXiv:1701.08493.

Redmon, J. and Farhadi, A. (2017). YOLO9000: Better,

Faster, Stronger. 2017 IEEE Conference on Computer

Vision and Pattern Recognition (CVPR), pages 6517–

6525.

Redmon, J. and Farhadi, A. (2018). YOLOv3: An Incre-

mental Improvement. ArXiv, abs/1804.02767.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-Net:

Convolutional Networks for Biomedical Image Seg-

mentation. ArXiv, abs/1505.04597.

Sarmiento, J. (2021). Pavement Distress Detection and Seg-

mentation using YOLOv4 and DeepLabv3 on Pave-

ments in the Philippines. ArXiv, abs/2103.06467.

Sarsam, S. I., Daham, A. M., and Ali, A. M. (2015). Imple-

mentation of Close Range Photogrammetry to Eval-

uate Distresses at Asphalt Pavement Surface. Inter-

national Journal of Transportation Engineering and

Trafﬁc System, 1(1):1–14.

Tan, Y. and Li, Y. (2019). UAV Photogrammetry-Based 3D

Road Distress Detection. ISPRS Int. J. Geo Inf., 8:409.

Wang, C.-Y., Liao, H.-Y. M., Wu, Y.-H., Chen, P.-Y.,

Hsieh, J.-W., and Yeh, I.-H. (2020). CSPNet: A

new backbone that can enhance learning capability of

CNN. In Proceedings of the IEEE/CVF Conference on

Computer Vision and Pattern Recognition Workshops,

pages 390–391.

Wang, K., Liew, J. H., Zou, Y., Zhou, D., and Feng, J.

(2019). Panet: Few-shot image semantic segmenta-

tion with prototype alignment. In Proceedings of the

IEEE/CVF International Conference on Computer Vi-

sion, pages 9197–9206.

Xu, R., Lin, H., Lu, K., Cao, L., and Liu, Y. (2021). A For-

est Fire Detection System Based on Ensemble Learn-

ing. Forests, 12(2):217.

Zhang, A., Wang, K., Fei, Y., Liu, Y., Chen, C., Yang,

G., Li, J., Yang, E., and Qiu, S. (2019). Automated

Pixel-Level Pavement Crack Detection on 3D Asphalt

Surfaces with a Recurrent Neural Network. Comput.

Aided Civ. Infrastructure Eng., 34:213–229.

Zhang, C. and Elaksher, A. (2012). An Unmanned Aerial

Vehicle-Based Imaging System for 3D Measurement

of Unpaved Road Surface Distresses. Comput. Aided

Civ. Infrastructure Eng., 27:118–129.

Zhang, L., Yang, F., Zhang, Y., and Zhu, Y. (2016). Road

crack detection using deep convolutional neural net-

work. 2016 IEEE International Conference on Image

Processing (ICIP), pages 3708–3712.

Zou, Q., Zhang, Z., Li, Q., Qi, X., Wang, Q., and Wang,

S. (2019). DeepCrack: Learning Hierarchical Convo-

lutional Features for Crack Detection. IEEE Transac-

tions on Image Processing, 28:1498–1512.

Towards a Low-cost Vision System for Real-time Pavement Condition Assessment

533