Vehicle Trajectory Prediction With Multimodal and Dynamics-Aware Interaction Neural Networks

Trajectory prediction is essential for improving the reliability and efficiency of vehicle-to-everything (V2X) communications, since link failures and frequent handovers can be alleviated by foreseeing positions of vehicles. Yet, accurate prediction is difficult due to complex interaction and multimodal motions. Deep learning methods are dedicated to addressing these challenging issues and hence surpass traditional physics-based models dramatically. However, current methods mainly model interaction based on spatial relationship while overlooking dynamics-aware interaction, which hinders comprehensive perception. When dealing with multimodality, stochastic latent space without proper goals as regularization is generated to enable one-to-many mapping between historical and future trajectories, which restrains inference accuracy. To tackle these problems, a multimodal and dynamics-aware interaction neural network, MODA, is proposed for vehicle trajectory prediction. To handle dynamics-aware interaction, a graph attention network is conducted. By taking each vehicle as a node and interaction as an edge, self-attention mechanism adaptively calculates edge strength between nodes, which acts as dynamics importance between vehicles for interactive aggregation. To facilitate multimodality, conditional variational autoencoder is incorporated, where supervisory information regularizes the latent space of a recognition network in the training phase for goal orientation. The proposed model is evaluated on two real-world highway datasets: NGSIM I-80 and US-101. The experiments demonstrate that MODA outperforms the state-of-the-art methods by a considerable margin.