Modelling driving behaviours using naturalistic driving data with a hierarchical attention network

Abstract

Listen

Translate article

This paper proposes an integrated deep learning-based trajectory prediction model that jointly models both lateral and longitudinal driving behaviours, addressing the limitations of separately modelling car-following and lane-changing behaviours. The model uses a hierarchical attention framework combining Graph Attention Network (GAT) and Transformer Encoder, which sequentially encodes a vehicle's historical trajectories and interactions with other vehicles, accounting for the continuity of driving scenarios. Additionally, a Gated Recurrent Unit (GRU) predicts multimodal trajectory data, which is processed by a Multilayer Perceptron (MLP) to output probabilities for each potential trajectory. Experiments on the highD dataset demonstrate that the model achieves outstanding accuracy, with an average displacement error (ADE) less than 0.3 m, final displacement error (FDE) less than 0.8 m, and significantly outperforms existing models in ADE and FDE. Furthermore, the multimodal prediction method outperforms the unimodal method, ensuring accurate short-term and long-term predictions, providing a comprehensive solution for driving behaviours modelling.