A Machine Learning Pipeline to Automatically Identify and Classify Roadway Surface Disruptions

Abstract

Smartphone-based applications for Intelligent Transportation Systems (ITS) have become a real possibility because of the sensing and computing capabilities of these devices. In this work we employ smartphones' accelerometers to sense the quality of roads, detecting the perturbations encountered by the vehicle. The ultimate goal of this line of work is to correctly identify, classify and georeference all obstacles so alleviating measures can be taken. Having a continuous series of accelerometer readings, the first problem is to identify when a perturbation was sensed (segmentation). To approach this problem, we propose using a Support Vector Machine (SVM), obtaining an accuracy of about 82%, outperforming other ad-hoc techniques such as Simple Mobile Average (SMA) and four other competitors. After segmentation, the next problem is to classify the event in one out of four different categories. To this end, we apply a Bag of Words representation and a Random Forest (RF), obtaining an accuracy of about 75%. These results were obtained by exhaustively training and testing this classifier over a newly created dataset that comprises signals for 30 different roads. Altogether, the use of a SVM followed by a RF seems to be a viable option to create a pipeline to automatically recognize and identify Roadway Surface Disruptions.