Training and Testing of Smartphone-Based Pavement Condition Estimation Models Using 3D Pavement Data

Abstract

Three-dimensional (3D) laser scanners have become a mainstream technology for the automatic assessment of pavement condition. The objective of this study is to leverage highly accurate 3D pavement data to train supervised machine learning models for pavement condition estimation using low-cost vehicle-mounted smartphone sensor data. First, the smartphone sensor data and 3D pavement data were registered on a common geographic information system (GIS) model of the road network. Second, recurrent neural networks (RNNs) with long short-term memory (LSTM) units were trained for the estimation of various distresses using smartphone sensor data as the input and 3D pavement data to provide the labels. Finally, the output of the models was accordingly postprocessed to provide distress values generally used for engineering decisions. The methodology was designed such that extensive calibration would not be required. When the Georgia Department of Transportation’s PAvement Condition Evaluation System (PACES) protocol was used as reference, the results presented here show that the proposed methodology can be used for estimating the IRI with a median absolute error (MAE) of 0.61 m/km (38.65 in./mi) and can estimate the average rut depth with a MAE of 4.19 mm (0.16 in.). The performance on cracking, raveling, and potholes was deemed unsatisfactory for engineering purposes.