The study focuses on predicting human movement in dynamic environments to enable safe and effective robot navigation. The researchers present a Transformer-based architecture that uses input features such as human positions, head orientations, and 3D skeletal keypoints to predict future human trajectories. The model captures the inherent uncertainty of future human trajectory prediction and shows state-of-the-art performance on common prediction benchmarks. The study also identifies the importance of 3D skeletal poses in reducing prediction error in scenarios with limited historical data.
Publication date: 2 Oct 2023
Project Page: https://human-scene-transformer.github.io/
Paper: https://arxiv.org/pdf/2309.17209
