A Simple Root Location Estimation for 3D Pose Estimation with Root-Relative Coordinate System

There is a strong need for human action recognition in various areas of society. One of the most important technologies for physical action recognition is human pose tracking. In recent years, three-dimensional (3D) pose tracking methods based on deep neural networks (DNNs) have made significant progress. However, most of these methods require a high-end computer equipped with a graphics processing unit (GPU). The purpose of this study is to provide a lightweight pose estimation network for human action recognition tasks. We used Google's MediaPipe Pose, which is optimized to run in real time on the CPU alone, for action recognition on a low-end personal computer (PC). However, MediaPipe Pose cannot track the location of a human body because it estimates poses on a coordinate system with the waist as the origin (i.e., a root-relative coordinate system). Therefore, in this study, we developed a method to obtain the absolute coordinates of the root with a simple calculation. It was also found that moving median filtering on the computed 3D absolute coordinates of the root can reduce jitter in pose tracking. The proposed method can be performed in near real-time on a commercially available PC with a camera, and is expected to have many applications.