Multiple Model Unscented Kalman Filtering in Dynamic Bayesian Networks for Intention Estimation and Trajectory Prediction

Abstract

Dynamic Bayesian networks (DBNs) are a popular method for driver intention estimation and trajectory prediction. To account for hybrid state spaces and non-linear system dynamics, sequential Monte Carlo (SMC) methods are often the inference method of choice. However, in state estimation problems with high uncertainty, SMC methods typically suffer from either high complexity (using many samples) or low accuracy (using an insufficient number of samples). In this paper, we present a multiple model unscented Kalman filter based DBN inference method for driver intention estimation and multi-agent trajectory prediction. This inference method reduces complexity, while still keeping the benefits of sample-based evaluation of non-linear and non-continuous transition models. Firstly, the state of the DBN is approximated as a mixture of Gaussians and estimated over time by tracking the multi-agent system. Secondly, a probabilistic forward simulation of the belief is performed to generate interaction-aware trajectories for all agents and all intention hypotheses. The proposed method is compared to SMC-based inference methods in terms of accuracy, variance and runtime in both simulations and real-world scenarios.