الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 131 |  |  |
| | | from | 131 | | |
Abstract
The most popular form of road surface is asphalt pavement. The
serviceability of this structure has a direct impact on today’s society and is
significantly linked to regional economic growth. As a result, one of the
most important responsibilities of the transportation authorities is to
guarantee the condition of the asphalt road surface during its serviceability
period by proactively detecting road distresses and undertaking timely
repair. This time factor is also important as early works have shown that
inadequate maintenance may dramatically affect costs over the pavement’s
lifetime.
For the last few decades, scientists and researchers have been trying
to find a simple and cost-effective method to apprise decision-makers of
the pavement condition of the road network in a timely manner as a part
of the Pavement Maintenance Management System (PMMS). Fortunately,
and with the evolution of computer vision tools and techniques, very good
results have been achieved regarding auto-detection, classification, and
quantification of road distress. Conversely, the traditional pavement
assessment methods, without a doubt, have got many disadvantages in
terms of time, money, safety, and expertise.
In this research, two computer vision models were introduced using
a new mainstream deep learning framework, Google NASNet and
Facebook AI (YOLO) model. This new family of models is the product of
a deep learning prediction model, which is a revolutionary approach that
has been used worldwide over the past few years. TensorFlow deeplearning framework is used to run the model. Promising results are
obtained from the proposed model with a performance of " ~ "400 FPS and
distress detection every " ~ "5 cm for a 40-km/h moving vehicle.
Furthermore, the output of the developed model will be used as an
input for our pavement rating calculation module to determine the
pavement condition. The proposed system focuses on specific types of
distresses where most of them show a substantial effect on the overall pavement condition: Alligator cracks, Longitudinal cracks, Transverse
cracks, Block cracks, Potholes, and Bleeding.
Experimental results show that the model based on Google NASNet
achieves promising prediction performance. Commercial software is used
to extract a manually labeled and annotated data set of approximately
2,700 images extracted from videos captured using a smartphone camera
mounted on the bumper of a personal automobile for some roads in Cairo,
Egypt, which is then processed using deep-learning software specifically
designed for this research.
In conclusion, we can say that our proposed solution has a potential
to lead to a breakthrough in the economics of road inspection methods
because of many reasons; first of all, it is real-time, cost-effective method
for gathering and analyzing data within PMMS replacing high-end
alternatives, moreover, classification model showed a 90% reduction in
model size and approximately 400% increase in inference speed compared
to the current state – of – the – art models with an outstanding precision of
detection more than 97%. Needless to say, that results showed that the
proposed model can work perfectly in Egypt since it was trained on
pavement distress from Egypt.