Anomaly Detection of Roads from Driving Data Using a Statistical Discrepancy Measure

Kazushi Ikeda, Takatomi Kubo, Hiroaki Sasaki, Masataka Mori, Kentarou Hitomi, Kazuhito Takenaka, Kenji Muto

Abstract: Anomalies in roads such as obstacles and accidents often cause fatal accidents and should be detected as soon as possible. We formulated the problem of detecting obstacles from driving behaviors as a change-point detection, not an anomaly detection problem, because driving behaviors are too distributed to make a `standard driving' model. To model the distribution of driving states at each location at each time, we employed a nonparametric method due to a wide variety of the distribution and applied a change-point detection method called the maximum mean discrepancy to our synthetic data made using a traffic simulator. As a result, we successfully detected the existence of obstacles with high accuracy.