STMF-IE: A Spatial-Temporal Multi-Feature Fusion and Intention-Enlightened Decoding Model for Vehicle Trajectory Prediction
Lists
Only you see this
Abstract
Accurate trajectory prediction of surrounding vehicles is crucial for ensuring the safety of autonomous vehicles. However, current methods based on neural networks still have room for improvement by further reducing their long-term error. This challenge stems from extracting temporal feature dependencies and mining spatial interactions in lane change scenarios. Previous research has not adequately established the connection between the past and the future features. To this end, we propose a spatial-temporal multi-feature fusion and intention-enlightened decoding (STMF-IE) model that jointly considers all these aspects. Initially, STMF-IE uses multi-head temporal self-attention to establish temporal dependencies that make explicit the extent of information usage at different time steps. The interplay between the neighboring vehicles and the target vehicle is quantitatively depicted through a devised multi-head spatial cross-attention mechanism. Furthermore, a multi-feature fusion module is proposed such that the extracted temporal dependence and spatial interaction features are integrated to reduce the superfluous features. Additionally, a novel mask matrix and an intention-enlightened decoding module are developed to refine the prediction performance in different lane change scenarios. Experimental results show that STMF-IE outperforms state-of-the-art methods for long-term prediction on NGSIM and highD datasets. We improve the RMSE metric by 13%–36% at a prediction horizon of 3$\sim$5 s. We also analyze the spatial-temporal feature correlation through visual results, promoting more interpretation.
BibTeX
@article{gao2025stmf,
author = {Kai Gao and Xunhao Li and Lin Hu and Xinyu Liu and Jinlai Zhang and Ronghua Du and Yongfu Li},
title = {{STMF-IE: A Spatial-Temporal Multi-Feature Fusion and Intention-Enlightened Decoding Model for Vehicle Trajectory Prediction}},
journal = {IEEE Transactions on Vehicular Technology},
year = {2025},
url = {https://openreview.net/forum?id=KwWI5kWEpM},
}Loading related content...