We present a cost-effective approach to evaluating road surface quality and roughness based on low-cost smartphones by leveraging accelerometer and gyroscope data sampled at 10 Hz. We extracted features from the vertical accelerations and rolling motions, in cloud to standard deviation and interquartile range, and trained a support vector machine (SVM) classifier to identify Good and Poor roughness levels based on IRI thresholds. We selected SVM specifically because it affords consistent performance with small datasets, reliably handles low-dimensional statistical features, and provides a stronger generalization than more complex machine-learning approaches. We provide experimental results from four vehicles on a thrice 50-m segment based dataset, demonstrating that the proposed method can discriminate (i.e., classify) roadway roughness and quality levels 80-100% of the time based on smartphone IMU data, which implies that even low-frequency smartphone IMU signals can provide useful roughness screening for planning and maintenance.
Yiliguoqi et al. (Thu,) studied this question.