Detecting damage in roads using convolutional neural networks

Resumen

Roads are subject to damage such as cracks and potholes, mainly due to overload and weather over time. To ensure the longevity of roads, to prevent economic losses, and to improve safety, damage detection is crucial in pavement conditions monitoring. Damage detection is usually performed in the field by surveying and is a time-consuming and unsafe task. This paper presents an experimental study using machine learning and convolutional neural networks for detecting damage in roads. Transfer learning and data augmentation techniques were used to build classification models. Three models were evaluated, and the accuracy results on average were near 80%. Best accuracy results were achieved on detecting potholes, exudation, raveling and patches. The models did not perform well when distinguishing among subtypes alligator, transversal, and longitudinal of crack damage. For these types of damage, the models achieved an average accuracy bellow to 70%.