Stochastic Multimodal Interaction Prediction for Urban Driving

Simultaneously predicting two highly interactive agents' future trajectories while considering the multimodal nature of navigation is important for autonomous vehicles to navigate urban scenarios. The interaction prediction is challenging because of the multimodality of driving, complex road structures, and non-obvious interactions among different traffic participants. Aiming to address these challenges, we design a three-channel framework based on the Heterogeneous Edge-enhanced graph ATtention network (HEAT) to model the complex interdependencies among agents and infrastructures, and simultaneously predict pairs of future trajectories of two highly interactive agents with assigned probabilities. Specifically, the method uses sequential models to extract agents' dynamics features from individual past states and applies a HEAT-based graph operator to model inter-agent interactions. Local maps of target agents are represented by top-view images and processed via convolutional blocks. The decoders then consider the combination of this information and produce pairs of trajectories with probabilities as interaction predictions. Evaluation results on the large-scale Waymo open dataset demonstrate that the proposed framework achieves state-of-the-art prediction accuracy. The proposed method has won the first place of the interaction prediction track at Waymo open dataset challenges 2021.