Vehicle Trajectory Prediction Based on Spatial-aware Transformer

Abstract

In recent years, deep learning-based vehicle trajectory prediction methods have gained significant attention in the field of transportation. However, most existing approaches rely on RNN-based models, which suffer from issues such as vanishing and exploding gradients when dealing with long sequences. These limitations hinder their performance in practical applications. To overcome these challenges, this paper proposes the application of the Transformer model for vehicle trajectory prediction. By incorporating the multi-headed self-attention mechanism and position encoding mechanism, the Transformer model can effectively capture the temporal relationships within vehicle trajectories and handle long sequence data. Moreover, considering that vehicle trajectories are influenced by surrounding neighboring vehicles, this paper introduces a novel attention mechanism-based spatial information awareness method to extract the impact of neighboring vehicles on the target vehicle. Additionally, the vehicle behaviors are categorized into six maneuvers, and multimodal trajectory prediction is achieved by generating corresponding trajectories and their probability distributions based on these six maneuvers. To evaluate the proposed method, experiments are conducted on the real dataset NGSIM and compared against existing approaches. The experimental results demonstrate that the proposed method surpasses the performance of benchmark methods in terms of prediction accuracy, highlighting its potential and practical value in vehicle trajectory prediction.