Interactive Prediction for Multiple, Heterogeneous Traffic Participants with Multi-Agent Hybrid Dynamic Bayesian Network

Abstract

Interactive prediction with multiple traffic participants in highly dynamic scenarios is extremely challenging for autonomous driving, especially when heterogeneous agents such as vehicles and pedestrians are involved. Existing prediction methods encounter problems on interpretability and generalizability to tackle such a complicated task. In this paper, we construct an integrated framework to estimate and predict the behavior of multiple, heterogeneous agents simultaneously. A Multi-agent Hybrid Dynamic Bayesian Network (MHDBN) method is proposed, which can model the state changes of multiple, heterogeneous agents in a variety of scenarios. We incorporate prior knowledge such as map information and traffic rules into the graph structure and use Particle Filter (PF) to track and predict intentions and trajectories of the agents. Motion data with pedestrian-vehicle interactions from a four-way-stop intersection in the real world is used to design the model and verify the effectiveness of the estimation and interactive prediction capability of the proposed framework.