Sensing and Classifying Roadway Obstacles in Smart Cities: The Street Bump System

We develop an infrastructure-free approach for anomaly detection and identification based on data collected through a smartphone application (Street Bump). The approach is capable of effectively classifying roadway obstacles into predefined categories using machine learning algorithms, as well as prioritizing actionable ones in need of immediate attention based on a proposed anomaly index. We explore some novel variants of classification algorithms that combine clustering with classification and introduce appropriate regularization in order to concentrate on a sparse set of most relevant features, which has the effect of reducing overfitting. Furthermore, the anomaly index we introduce combines novel metrics of obstacle irregularity computed based on the data captured by the Street Bump smartphone application. Results on an actual data set provided by the City of Boston illustrate the feasibility and the effectiveness of our system in practice.