Multi-Scale Hypergraph Self-Attention for Trajectory Prediction in Autonomous Driving
2026-99-2413
11/20/2026
- Content
- Accurate trajectory prediction is essential for safe decision making and path planning in autonomous driving systems. The core challenge lies in effectively modeling complex multi-agent interactions and structured road constraints in dynamic traffic scenarios. To address these challenges, this paper proposes a structured trajectory prediction framework based on multi-scale hypergraph self-attention, which jointly models local vehicle interactions and global vehicle–lane relationships within a unified representation. A Graph Convolutional Network (GCN) is used to encode lane topology. A hypergraph self-attention mechanism adaptively aggregates critical interactions, while a temporal Transformer captures long-term motion dependencies. Experiments on the Argoverse Motion Forecasting Dataset demonstrate that the proposed method achieves 1.69 minimum Average Displacement Error (minADE), 3.79 minimum Final Displacement Error (minFDE), and 0.60 Miss Rate (MR), outperforming representative methods such as VectorNet by 6.1% and 4.8% in minADE and minFDE. Ablation studies show that the Transformer reduces minADE/minFDE from 1.80/3.98 to 1.75/3.92, while combining it with the weighted hypergraph further improves performance to 1.69/3.79, confirming the effectiveness of jointly modeling high-order interactions and long-term dependencies.
- Citation
- Chen, X. and Lim, K., "Multi-Scale Hypergraph Self-Attention for Trajectory Prediction in Autonomous Driving," 2026 2nd International Conference on Intelligent Transportation and Future Mobility (ITFM2026), Xi'An, China, March 27, 2026, .