Leveraging Transformer Model to Predict Vehicle Trajectories in Congested Urban Traffic

Abstract

Accurate vehicle trajectory prediction enables safe, comfortable, and optimal proactive motion planning for connected and autonomous vehicles (CAVs). Because of rapid advances in learning techniques and increasing access to massive amounts of data, deep learning techniques have been applied to predict vehicle trajectories, especially the long short-term memory (LSTM) model. However, the accurate prediction of vehicle trajectories for congested urban traffic remains problematic, as existing LSTM models do not perform well. To address this gap, this paper proposes to leverage an emerging deep learning technique—transformer—and utilizes a recently released dataset (pNEUMA) for predicting vehicle trajectories in congested urban traffic. The proposed transformer model uses the self-attention mechanism, which helps to identify dependencies within the model inputs, to systematically determine the impacts of vehicular interactions on the target vehicle’s future trajectory. The pNEUMA dataset, which provides drone-based large-scale data of congested urban traffic, is processed to fit a typical trajectory prediction scenario, and used to train the transformer model. Numerical studies are conducted to analyze the effectiveness of the proposed modeling approach. A comparison of the proposed model with representative LSTM models highlights the advantages of leveraging the transformer model characteristics for the vehicle trajectory prediction of congested urban traffic. By contrast, existing LSTM models may suffice for the trajectory prediction of freeway traffic. The results also indicate that, unlike for vehicle trajectory prediction for freeway traffic, a longer time window of inputs does not guarantee better prediction performance for congested urban traffic.