Accurately predicting trajectories in complex and uncertain driving scenes is crucial for autonomous driving systems. However, existing trajectory prediction methods often suffer from the presence of redundant information, which leads to increased noise and computational complexity. In this paper, we propose a novel approach for multi-modal trajectory prediction based on sparse heterogeneous interactions. We separate the interactions into two branches: agent-agent interactions and agent-lane interactions, to capture the sparse interactions between traffic elements. We utilize Graph Transformer models to learn local and global interaction features. Additionally, we integrate a vehicle kinematic model to generate trajectory proposals that adhere to physical constraints. Experimental results on the Argoverse Forecasting Dataset demonstrate the effectiveness of our approach, achieving superior performance in multi-modal trajectory prediction tasks.
Sparse Heterogeneous Interactions for Multi-Modal Trajectory Prediction
28.10.2023
1571348 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Europäisches Patentamt | 2024
|Europäisches Patentamt | 2021
|