Bayesian model-based sequence segmentation for inferring primitives in driving-behavioral data

Low-level driving-behavioral data often display statistical patterns such as periods of turning, braking and acceleration, as well as different combinations of these. Collectively, these patterns can be regarded as a language of "driving primitives," i.e. building blocks for more complex maneuvers. Such a language allows us to reason about more meaningful driving behaviors, e.g. overtaking and merging, by regarding them as sequences of primitives. In this paper we introduce a new approach for automatically segmenting a sequence of data into primitives. Segmentation, followed by clustering, allows us to build a language of primitives quickly and effectively. Our approach is computationally efficient, completely unsupervised, and requires but minimal preprocessing. We demonstrate its effectiveness with an experiment involving genuine driving-behavioral data from inertial sensors.