Improving pavement anomaly detection using backward feature elimination

Abstract

Early approaches for pavement anomaly detection used simple heuristic on a small number of features and their associated thresholds. Such methods may not be suitable for data collected from heterogeneous sources (e.g., different vehicles, pavements, inertial sensors, etc.). Instead of manually selecting a set of features and their thresholds, we propose using backward feature elimination on a large set of features such that the optimal set of features can be determined. Our experimental results show that the features selected by backward feature elimination yield the best performance, compared to using all features from the sampled data of the accelerometer and gyrometer.