BROWSE
Detail
Journal Article
효율성과 견고성을 갖춘 궤적 예측을 위한 상호작용 그래프 가지치기
- Issue Date
- 2026-03
- Citation
- 한국자동차공학회 논문집, v.34, no.3, pp.365-373
- ISSN
- 1225-6382
- Publisher
- 한국자동차공학회
- Language
- Korean
- Type
- Journal Article
- Abstract
- Modeling interactions between dynamic and static road participants(e.g., vehicles, pedestrians, lane dividers, road markings) is crucial for accurate and robust trajectory prediction. Transformer-based approaches with self- and cross- attention have become the standard due to their ability to capture complex, long-range dependencies. However, as the number of participants grows, complexity increases quadratically, imposing a significant burden during training and inference. To address this, the researchers propose IGP-Traj, an interaction graph pruning algorithm that eliminates dispensable connections from the fully connected interaction graph. IGP-Traj dynamically prunes irrelevant agent-to-agent(A2A) and agent-to-map(A2M) connections, focusing on critical interactions, such as nearest and forward-looking ones. This reduces GPU memory usage, accelerates training and inference, and enhances model performance. We evaluate IGP-Traj on three state-of-the-art Transformer-based models using the large-scale Argoverse 1 and 2 datasets, demonstrating its effectiveness in achieving efficient and robust trajectory prediction.
- Keyword
- Autonomous driving, Artificial intelligence, Trajectory prediction, Transformer, Interaction graph
- KSP Keywords
- Cross-, Forward-looking, GPU memory usage, Graph pruning, Model performance, Pruning algorithm, artificial intelligence, autonomous driving, interaction graph, large-scale, long-range dependencies
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
