计算机科学
姿势
人工智能
关节式人体姿态估计
计算机视觉
变压器
三维姿态估计
机器人
运动捕捉
图形
人机交互
预处理器
运动估计
运动(物理)
理论计算机科学
物理
量子力学
电压
作者
Ru Huo,Qing Gao,Jing Qi,Zhaojie Ju
标识
DOI:10.1007/978-981-99-6498-7_16
摘要
3D human pose estimation is widely used in motion capture, human-computer interaction, virtual character driving and other fields. The current 3D human pose estimation has been suffering from depth blurring and self-obscuring problems to be solved. This paper proposes a human pose estimation network in video based on a 2D lifting to 3D approach using transformer and graph convolutional network(GCN), which are widely used in natural language processing. We use transformer to obtain sequence features and use graph convolution to extract features between local joints to get more accurate 3D pose coordinates. In addition, we use the proposed 3D pose estimation network for animated character motion generation and robot motion following and design two systems of human-computer/robot interaction (HCI/HRI) applications. The proposed 3D human pose estimation network is tested on the Human3.6M dataset and outperforms the state-of-the-art models. Both HCI/HRI systems are designed to work quickly and accurately by the proposed 3D human pose estimation method.
科研通智能强力驱动
Strongly Powered by AbleSci AI