Interaction-Aware Motion Prediction For Autonomous Driving: A Multiple Model Kalman Filtering Scheme

We consider the problem of predicting the motion of vehicles in the surrounding of an autonomous car, for improved motion planning in lane-based driving scenarios without inter-vehicle communication. First, we address the problem of single-vehicle estimation by designing a filtering scheme based on an Interacting Multiple Model Kalman Filter equipped with novel intention-based models. Second, we augment the proposed scheme with an optimization-based projection that enables the generation of non-colliding predictions. We then extend the approach to the problem of simultaneously estimating multiple vehicles by using a hierarchical approach based on a priority list. The priority list is dynamically adapted in real-time according to a proposed sorting algorithm. Finally, we evaluate the proposed scheme in simulations using real-life vehicle data from the Next Generation Simulation (NGSIM) dataset.