Maneuver prediction at intersections using cost-to-go gradients

According to the analysis of car accidents many casualties occur at intersections. As ongoing research demonstrates, Advanced Driver Assistance Systems that aim at preventing this type of accident, need to reliably predict the turning maneuver of all relevant participants in the scene. In this work an approach is introduced, which models human drivers as the optimizer of an optimal control problem with an unknown terminal state. Tracking the cost-to-go gradient to the terminal state of each driving option leads to the most plausible hypothesis. The optimal control problem itself is formulated with costs that minimize jerk, time and steering effort with good resemblance to typical human driving behavior. In combination with a simplified vehicle model this leads to a nonlinear constrained dynamic optimization problem, which is solved numerically. The performance of the proposed approach is evaluated on data obtained in a field test with promising results.