Visual Perception Method Based on Human Pose Estimation for Humanoid Robot Imitating Human Motions

The goal of the system proposed in this paper is to develop functions for humanoid robots to use vision to perceive human behavior and imitate human motions. With that aim, the position of the human joint points is recorded by a key point detector and the motion data is fed to 3D-baseline to estimate the 3D human skeleton. The accuracy of 3D joint point prediction is directly related to the degree of robot's restoration of human motions. Therefore, data augmentation is applied to improve the accuracy of 3D human pose estimation. The experimental results on Human3.6M demonstrate that our method can yield good performance. In addition, the motion mapping between human and humanoid robot is realized by inverse kinematics. A balance strategy based on ankle joint adjustment is proposed to maintain the stability of the robot. Tested in a humanoid robot, the system is able to imitate human movements naturally after observation, which improves the human-computer interaction of the humanoid robot.