Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : May 27, 2020 - May 30, 2020
We study the pose and state estimation of a humanoid robot from a monocular image using deep learning. Until now, soccer humanoid robots have not detected the pose of the other robot. As with human soccer, it is desirable that soccer robots detect the poses of other players, and they estimate other robots’ behaviors such as kicking the ball. We developed the system to estimate the pose of a humanoid robot based on OpenPose and verified its effectiveness through an experiment. We found that the average value of PCK in the developed pose estimation was 0.79. In addition, we proposed a model that estimates the state such as walking and kicking from the pose. We developed the state estimation system and verified its effectiveness though an experiment.
