[2Win5] Poster session 2

[2Win5-88] Human Trajectory Prediction from First-person 2D Skeleton Sequences

〇Taishu Arashima¹, Hiroshi Kera¹, Kazuhiko Kawamoto¹ (1.Chiba University)

Keywords:Deep Learning, Human Trajectory Prediction, first-person perspective, 2D skeletal sequence

Human trajectory prediction is crucial for autonomous navigation and surveillance systems. However, conventional methods rely on past trajectory data and expensive 3D sensors, limiting practicality. To overcome these challenges, we propose a deep learning framework that predicts trajectories solely from first-person-view 2D skeleton sequences. Our method utilizes 2D skeleton sequences as input during training while leveraging positional data. We employ graph convolutional networks (GCNs) to model spatial dependencies between joints and capture interpersonal interactions. This enables accurate trajectory prediction using only monocular cameras. Experimental results on the JTA dataset demonstrate that our approach predicts trajectories up to 4.8 seconds into the future, proving its effectiveness.