Real-Time Road Anomaly Detection, Using an On-Board Data Logger

This paper attempts to improve real-time detection of Roadway Surface Disruptions (RSDs) while pointing out the limitations imposed on this application owing to hardware and data acquisition constraints. We have developed and publicly shared a large amount of heterogeneous data set collected over a continuous drive for four different kinds of road anomalies -- Cat eyes, Potholes, Manholes and Speed Bumps, along with regular road, using a custom-made, dedicated device for data logging, recording vehicle acceleration and GPS location. The data set collected from four out of eleven vehicles was used to train six common Machine Learning classifiers. Sequential forward selection was performed on three groups of features i.e. FFT, DWT and peak accelerations. Performance metrics including FPR, FNR and TPR were calculated and compared and the best results for each of the four vehicles have been reported. In addition to this, we also present annotated maps showing regular drivers which route has the least number of anomalies and would hence be the smoothest. Finally, our paper also addresses the RSD detection problem as a multiclass classification problem and implements one versus all technique to distinguish two different sets of road anomalies for road repairing authorities as well as drivers and passengers.