The learning of longitudinal human driving behavior and driver assistance strategies

International audienceModels of the human driving behavior are essential for the rapid prototyping of error-compensating assistance systems. Various authors proposed control-theoretic and production-system models. Here we present machine-learning alternatives to train assistance systems and estimate probabilistic driver models from human behavior traces. We present a partially autonomous driver assistance system based on Markov Decision Processes. Its assistance strategies are trained from human behavior traces using the Least Square Policy Iteration algorithm. The resulting system is able to reduce the number of collisions encountered when following a lead-vehicle. Furthermore, we present a Bayesian Autonomous Driver Mixture-of-Behaviors model for the longitudinal control of human drivers based on the modular and hierarchical composition of Dynamic Bayesian Networks. Their parameters and structures are estimated from human behavior traces using a discriminative scoring criterion based on the Bayesian Information Criterion. This allows the selection of pertinent percepts from the variety of percepts proposed for driver models according to their statistical relevance. The resulting driver model is able to reproduce the longitudinal control behavior of human drivers while driving unassisted or assisted by the presented assistance system