Software para identificação de defeitos na superfície de pavimentos rodoviários utilizando Deep Learning

Abstract

The road modal is the main Brazilian transport system. However, however, problems in the country's economy, generating direct costs related to accidents, delay in cargo, impact on the economy as impact, between our transport, increase in the price of products other factors. In this case, a software that identifies, quantifies, georeferences and presents the challenges presented in asphalt routing pavements in order to contribute in the area of pavement management. For the elaboration of this software, use a Deep Learning software to learn from the exercises. A test data collection process was carried out to perform in the test on training images and videos and the neural network. In this process, studies of data sets for smartphones and data acquisition attempts were carried out, where a comparative analysis of each was performed. To assist in the validation of the results, an inventory support software was developed. This software runs on the web platform with the aim of presenting inventory information in a graphical interface, which can also serve as a support in decision making by researchers, engineers, analysts and technicians in the area of management of plans for future surveys. And, finally, the classification was carried out based on the learning to make a diagnostic detection order on the pavement. As an object of study, two classes of used objects were used: holes. Among the results generated, the best dataset obtained an accuracy of 99.89% for holes and 95.91% for patches. All results were made available in the inventory support software. Therefore, this work presents a software that detects holes and patches in paved roads through images captured from cameras or smartphones.