Towards Predictive Driving through Blind Intersections

This paper presents an approach for predictive driving when facing blind intersections based on expert data. Expert drivers anticipate and avoid potential dangerous situations. In most cases these complex behaviors cannot be reproduced by state of the art planning approaches. We present an analysis of expert drivers while passing through blind intersections, we extract useful driving features to model the intersection and propose a cost function based on those features. We use inverse reinforcement learning to extract a feature-based cost function and learn its parameters from driver data. Finally, feature weights are computed based on collected expert driving data (using 211 trajectories). Evaluation was performed in terms of trajectory and speed using modified Hausdorff distance. Experimental results show that the planner is capable of computing trajectories comparable to those ones of the expert drivers when facing blind intersections.