Discriminative Learning of Relevant Percepts for a Bayesian Autonomous Driver Model

Mark Eilers and Claus Möbus
Proceedings of the Sixth International Conference on Advanced Cognitive Technologies and Applications, COGNITIVE 2014
Models of the human driving behavior are essential for the rapid prototyping of assistance systems. Based on psychological studies, various percepts and measures have been proposed for the lateral and longitudinal control in driver models without demonstrating the generalizability of results to natural settings. In this paper, we present the learning of a probabilistic driver model. It represents and mimics the lateral and longitudinal human driving behavior on virtual highways by performing situation-adequate lane-following, car-following, and lane changing behavior. Because there is considerable uncertainty about the relevant percepts in natural driving behavior, we select hypothetically relevant percepts from the variety of possibilities based on their statistical relevance. This is a new approach to generate hypothesis about the relevant percepts and situation-awareness of drivers in dynamic traffic scenes. The percepts are revealed in a structure-learning procedure using a discriminative scoring criterion based on the Bayesian Information Criterion. Discriminative learning maximizes the conditional likelihood of probabilistic models, whereas the traditional generative learning maximizes the unconditional likelihood. This way, it attempts to find the structure with the best performance for the intended use, which in our application is the best prediction of driving actions given the available percepts.
5 / 2014
Holistic Human Factors and System Design of Adaptive Cooperative Human-Machine Systems