IAtraj: Multi-Modal Trajectory Prediction Through Contextual Information Spatio-Temporal Interaction and Awareness

Abstract

Accurately and feasibly predicting the future trajectories of autonomous vehicles is a critically important task. However, this task faces significant challenges due to the variability of driving intentions and the complexity of social interactions. These challenges primarily arise from the need to understand one’s driving behaviors and model the interaction information of the surrounding environment. A substantial amount of research has been focused on integrating interaction information from the surrounding environment, mainly using raster images or High-Definition maps (HD maps). However, the real-time update of environmental maps and the high computational cost associated with processing interaction information using compatible technologies such as vision have become limiting factors. Additionally, ineffective simulation and modeling of real driving scenarios, coupled with inadequate understanding of contextual environmental information, result in lower prediction accuracy. To overcome these challenges, we propose a multi-modal trajectory prediction model based on sequence modeling namely IAtraj, incorporating multiple attention mechanisms, focuses on the three critical elements in real traffic scenarios: the target agent’s historical trajectory, effective interactions with neighboring vehicles, and lane supervision and retention strategies. To better model these elements, we design modules for Temporal Interaction (TI), Spatial Interaction (SI), and Lane Awareness (LA). Through extensive experiments conducted on the publicly available nuScenes dataset, IAtraj exhibits outstanding performance, successfully addressing the challenges of temporal dependencies in trajectory sequences and the representation of scene changes. Finally, comprehensive ablation experiments validate the effectiveness of each significant module, reinforcing the reliability and robustness of IAtraj in dealing with complex traffic scenarios.