Learning the Relevant Percepts of Modular Hierarchical Bayesian Driver Models Using a Bayesian Information Criterion

Modeling drivers' behavior is essential for the rapid prototyping of error-compensating assistance systems. Various authors proposed control-theoretic and production-system models. Based on psychological studies various perceptual measures (angles, distances, time-to-x-measures) have been proposed for such models. These proposals are partly contradictory and depend on special experimental settings. A general computational vision theory of driving behavior is still pending. We propose the selection of drivers' percepts according to their statistical relevance. In this paper we present a new machine-learning method based on a variant of the Bayesian Information Criterion (BIC) using a parent-child-monitor to obtain minimal sets of percepts which are relevant for drivers' actions in arbitrary scenarios or maneuvers.