Attention Mechanism-Based Maneuver-Aware Pooling Network for Vehicle Trajectory Prediction

Accurate trajectory prediction is particularly important in order to ensure that self-driving vehicles can travel safely and efficiently in complex and changing traffic environments. However, existing trajectory prediction models often do not fully consider the mutual influence between interacting vehicles within a scene when capturing inter-vehicle interactions. To this end, this paper proposes a maneuver-aware pooling network based on the attention mechanism, which aims to effectively capture the interdependencies between interacting vehicles within a scene. This pooling network adopts polar coordinates to represent the vehicle historical trajectory information, combines the vehicle direction and radial velocity, and introduces an attention mechanism to judge the similarity of the historical hidden layer among the surrounding vehicles to comprehensively consider the magnitude of the influence among the interacting vehicles within the scene and improve the local feature information of the maneuver-aware pooling layer. We incorporate this attention mechanism-based maneuver-aware pooling network into the GRU-LSTM codec model architecture and validate it on the NGSIM dataset. The experimental results show that the proposed model is able to accurately predict the future motion trajectories of vehicles, and achieves a significant advantage in prediction accuracy over various existing state-of-the-art algorithms.