Becoming Smarter at Characterizing Potholes and Speed Bumps from Smartphone Data — Introducing a Second-Generation Inference Problem

Much has been said regarding the automatic identification of roadway obstacles by analyzing data collected from inertial sensors either fixed to the vehicle or embedded into the drivers' smartphones. Literature is vast in models to, given a record of sensor readings, determine if the sample corresponds to a pothole or speed bump, even with scores beyond 90% in performance. Acknowledging this advance, this article considers the next-generation version of this problem. Specifically, we investigate questions such as: what physical properties of roadway obstacles could be inferred from the same sensor readings? or, what are the best schemes to model this profile problem? To approach these questions we built the first obstacle-detailed data set that is composed of accelerometer and gyroscope readings of 163 potholes and 101 speed bumps. This data set is the first of its kind, since it specifies ground truth labels that correspond to potholes' depths and also, functional status (OK - Not OK) for speed bumps. We approach this fine-grained characterization using three different learning schemes, as a Regression, Classification and Learning to Rank tasks. Results are encouraging, reporting a RMSE for pothole's depth prediction of up to 1.68 cm and classification performance of 0.89 in AUC score. In summary, after more than 10 years of analysis, struggles and achievements, it is time for the community to become smarter and start profiling roads with real detail.