Interaction-aware Prediction of Occupancy Regions based on a POMDP Framework

Extensive research has been done in the field of automated driving systems to reliably and accurately predict the motion of surrounding vehicles. Predicted occupancy regions are a suitable starting point for trajectory planning, criticality estimation, etc. However, interactions between traffic participants are difficult to predict due to their high complexity and mutual dependencies. This work proposes a novel approach for modeling these traffic interactions with the help of a Partially Observable Markov Decision Process (POMDP). A key component of the presented approach is that the resulting policy is not used for the control of the EGO vehicle. Instead, it is utilized for the generation of continuous predicted occupancy regions. The approach applies a particle ensemble approximation in order to solve the POMDP optimization. Hereby, the reward function is a central aspect to incorporate expert knowledge in the statistical framework. The potential of the proposed method is demonstrated using typical urban junction scenarios.