📄 中文摘要
人类轨迹预测对于社交机器人在拥挤环境中的导航至关重要。大多数现有方法将人类视为点质量，而本研究则通过利用不同的人体骨骼特征来提高多智能体轨迹预测的准确性。系统评估了2D和3D骨骼关键点及其衍生的生物力学线索作为额外输入的预测效用。在JRDB数据集和一个新的360度全景视频社交导航数据集上进行的全面研究表明，关注下肢3D关键点可以使平均位移误差减少13%，而将3D关键点输入与相应的生物力学线索结合使用则进一步提升了预测性能。
腿部优于手臂：基于自我中心机器人感知的下肢姿态对人类轨迹预测的预测价值
📄 English Summary
Legs Over Arms: On the Predictive Value of Lower-Body Pose for Human Trajectory Prediction from Egocentric Robot Perception
Predicting human trajectory is essential for social robot navigation in crowded environments. Most existing methods treat humans as point masses; however, this study presents an approach that leverages various human skeletal features to enhance multi-agent trajectory prediction accuracy. A systematic evaluation of the predictive utility of 2D and 3D skeletal keypoints, along with derived biomechanical cues as additional inputs, was conducted. Through a comprehensive analysis on the JRDB dataset and a new dataset featuring 360-degree panoramic videos for social navigation, it was found that focusing on lower-body 3D keypoints results in a 13% reduction in Average Displacement Error. Furthermore, augmenting 3D keypoint inputs with corresponding biomechanical cues significantly improves prediction performance.