Answers are grounded in this paper and cite their sources.
A prediction model demonstrates improved vehicle trajectory accuracy at intersections, suggesting benefits for autonomous driving safety.
Signalised intersections are vital in urban road networks and accurate trajectory prediction for surrounding vehicles is crucial for autonomous driving safety. To overcome limitations in trajectory feature representations and interaction modelling among heterogeneous vehicles, we propose a hierarchical generative adversarial network‐based prediction model that leverages temporal‐spectral feature fusion and multi‐factor interaction modelling. The generator comprises three levels. At the self‐layer, the discrete Fourier transform extracts spectral features: low‐frequency components capture global motion trends, while high‐frequency components encode local dynamics; these are fused with temporal features for a comprehensive motion representation. At the social‐layer, a multi‐factor attention pooling mechanism integrates relative position, velocity vectors and vehicle type to model heterogeneous interactions explicitly. At the scene‐layer, traffic signal phase information is encoded to impose environmental constraints. Adversarial training enables diverse, plausible multi‐vehicle trajectory generation. Evaluated on two real‐world datasets — SinD and Waterloo — the model achieves significant gains. On SinD (6 s/9 s horizons), minADE improves by 21.43% (14.56%) and minFDE by 18.80% (3.27%); on Waterloo (8 s), minADE and minFDE improve by 19.29% and 19.95%, respectively. Comprehensive ablation studies and qualitative‐quantitative analyses confirm the effectiveness and necessity of each component. Code is available at: https://github.com/zzt007/TSF‐GAN .
No takes yet. Share an insight, caveat, or question.
Zhong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: