Evaluating road pavement condition is one of the significant factors in managing road structures. In particular, the pavement roughness impacts the safety of the driver, energy efficiency, and public safety as well as the degradation of the road. The ...
Evaluating road pavement condition is one of the significant factors in managing road structures. In particular, the pavement roughness impacts the safety of the driver, energy efficiency, and public safety as well as the degradation of the road. The International Roughness Index (IRI) indicates the extent of pavement roughness and has been adopted World widely. Laser-sensor-based equipment such as road-profiler is used to measure the IRI in field experiments, but its surveying frequency is low due to requirement of unique equipment. With advance the computer techniques, various methods have been suggested for evaluating the IRI: For example, dynamic responses of a vehicle or other road condition information are incorporated with machine learning model for predicting IRIs. Also, the computer vision technologies are used in the road management while most of those applications are for determining the existence of cracks. To dates quantitative evaluation, such as IRI estimation, has been understudied. Thus, in this study, two Convolutional Neural Network (CNN) models are developed for road grade classification and IRI prediction, by using the photo realistic images. The image dataset that trains CNN models are developed in the follow procedures: (i) 3-dimensional road profiles are generated using the measured IRI from an in-service road; (ii) the profile, which contains a single IRI within the unit road segment are rendered using a 3Ds max with V-Ray; and (iii) the photo realistic images were then captured using a physical camera considering its orientation and the sunlight direction variation. In total 9,600 synthetic image are constructed. Two CNN models were developed, which yield accurate model performance when compared with models with different layer sizes and numbers. As a result, the proposed model achieved 99% accuracy for road grade classification and R-squared over 0.99 for IRI prediction. Then to further examine the efficacy of the proposed model, images collected from in-service road are used in the model. Note that because various light directions are considered in the data base, the model is less susceptible to the driving directions. As a results, the proposed framework that utilizes the photo realistic images has shown potentials for facilitating the automated and flexible evaluation of pavement condition.