Deep unsupervised learning of a topological map of vehicle maneuvers for characterizing driving styles

In this paper we propose a deep learning approach to analyzing driving styles from multivariate timeseries of driving behaviors, a deep topological map of vehicle maneuvers. Our neural network learns a topological map that embeds and clusters short subsequences of driving behavior data in a latent space. Our approach is tested on a highway-driving dataset that contains driving behavior data recorded with 59 drivers using a driving simulator. The experiments demonstrated that our proposed method embedded the driving behavior data into a two-dimensional, topological map where extracted elemental driving behaviors are represented as clusters. Individual driving episodes can be compactly represented as estimated probabilistic distributions of elemental driving behaviors over the learned map. We further demonstrated that clusters of driving episodes, “driving styles”, and their related features can be efficiently inferred in a data-driven manner by the application of a nonparametric Bayesian method to these estimated probabilistic distributions.