Two-stream LSTM Network with Hybrid Attention for Vehicle Trajectory Prediction

Abstract

Trajectory prediction aims to estimate future location by exploring driving behavior and historical trajectory, which is essential for driving decision-making and local motion planning of smart vehicles. However, affected by the multiple complex interaction in the traffic scene, predicting future trajectory precisely is a challenging task. Most existing models simply fuse the inter-vehicle interaction with the vehicle motion state and use the fusion vector for temporal modeling, which affects the extraction of information temporal dependency. Furthermore, the loss of important historical hidden state in recursive loops makes the long-term prediction performance of the sequence model not ideal. To address this issue, this paper proposes the Two-stream LSTM Network with hybrid attention mechanism (TH-Net). Specifically, we construct Two-stream LSTM structure (TS-LSTM) to build independent information transmission links for inter-vehicle interaction and vehicle motion state while maintaining their coupling relationship. In addition, Hybrid Attention Mechanism (H-AM) is proposed to explore the importance of hidden state from the dimensions of time and feature, and guides TH-Net to selectively reuse it. Experiments on the public dataset HighD demonstrate that TH-Net remarkably outperforms the state-of-the-art methods in long-term prediction performance.