核心概念
Large Language Models (LLMs) are utilized in LG-Traj to enhance pedestrian trajectory prediction by incorporating motion cues and social interactions, demonstrating effectiveness on benchmark datasets.
要約
Accurate pedestrian trajectory prediction is crucial for various applications.
LG-Traj incorporates LLMs to generate motion cues from past trajectories.
Clustering future trajectories using a mixture of Gaussians enhances prediction.
Motion encoder and social decoder capture motion patterns and social interactions.
Ablation studies show the importance of motion cues, position encoding, and trajectory augmentation.
Comparison with state-of-the-art methods highlights the superior performance of LG-Traj.
統計
論文はETH-UCYとSDDの人物軌跡予測ベンチマークで手法の効果を示す。
モデルは過去の軌跡から動きの手がかりを生成するためにLLMを活用している。