Deep Particle Filter for Vehicle Trajectory Estimation under Uncertainty

Future intelligent transportation systems depend on accurate and reliable trajectory estimation of road vehicles. However, mainstream researches focus on estimation of trajectories without uncertainty quantification, which is an inevitable problem in a safety-critical system. This paper proposes a fully differentiable deep particle filter (DPF) framework for the trajectory estimation task, which combines the arbitrary posterior distribution representation ability of particle filters and the powerful universal function approximation ability of neural networks. At the core of DPF lies a generative model for simulating state transition uncertainty and a regression model to learn observation uncertainty. Despite its simplicity, the proposed algorithm achieves better accuracy than existing methods when evaluated on the NGSIM dataset. We also show that uncertainty of trajectories can be well approximated from particle distribution. The noise robustness of the proposed algorithm is further demonstrated by injecting dynamic noise into observation.