Multimodal Prediction-Driven Spatiotemporal Risk Field Modeling for Dynamic Interaction Assessment in Autonomous Driving
IEEE Access · 2026 · 14 · 165931-165941
DOI: 10.1109/access.2026.3747316 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Nangui Fan und 5 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Hainan Provincial Key Research and Development Program of China.
Abstract (Originalsprache)
Conventional driving risk metrics have limited ability to represent multimodal future behaviors, while discrete predicted trajectories are difficult to use directly for continuous risk assessment. This paper proposes a multimodal prediction-driven spatiotemporal risk field for autonomous driving. Road feasibility constraints, deterministic dynamic risk, and multimodal predictive risk are represented in a unified local spatiotemporal coordinate system. Predicted trajectories, mode probabilities, motion directions, and temporal evolution are integrated to transform discrete future behaviors into continuous time-varying predictive-risk fields. Position–time alignment synchronizes the predicted ego motion with surrounding-agent risk for dynamic interaction assessment. Experiments on the Argoverse 2 dataset evaluate dynamic interaction identification. The probability-weighted predictive-risk score achieves an area under the receiver operating characteristic curve (AUROC) of 0.8880 and an average precision (AP) of 0.6130. The corresponding values for a baseline combining predicted ego motion with constant-velocity agent extrapolation are 0.8284 and 0.3010, respectively. Removing position–time alignment substantially degrades interaction identification.
Förderung
- Hainan Provincial Key Research and Development Program of China
Zitieren
APA
Fan, N., Li, P., Chen, Z., Zhu, K., Wang, B., & Li, X. (2026). Multimodal Prediction-Driven Spatiotemporal Risk Field Modeling for Dynamic Interaction Assessment in Autonomous Driving. IEEE Access, 14, 165931-165941. https://doi.org/10.1109/access.2026.3747316
Vancouver
Fan N, Li P, Chen Z, Zhu K, Wang B, Li X. Multimodal Prediction-Driven Spatiotemporal Risk Field Modeling for Dynamic Interaction Assessment in Autonomous Driving. IEEE Access. 2026;14:165931-165941. doi:10.1109/access.2026.3747316
BibTeX
@article{Fan2026,
author = {Fan, Nangui and Li, Peixin and Chen, Zhenbin and Zhu, Kanghui and Wang, Bodong and Li, Xinquan},
title = {Multimodal Prediction-Driven Spatiotemporal Risk Field Modeling for Dynamic Interaction Assessment in Autonomous Driving},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {165931--165941},
doi = {10.1109/access.2026.3747316}
}
EUVIMED