||Humanoid Robot Teleoperation System using a Fast Vision-based Pose Estimation and Refinement Method
||(Jae-Min Sa) ; (Kang-Sun Choi)
|| Pose estimation; Filtering; Teleoperation
||Humanoid robot teleoperation is a complex task that requires sophisticated hardware and software interfaces. Three-dimensional human pose estimation can be effectively applied to the teleoperation of humanoid robots thanks to its intuitive and easy usability. However, when the human pose is wrongly inferred, it can cause critical and dangerous situations for robot operation. In this paper, a reliable humanoid robot teleoperation system is proposed, which involves the operation of a humanoid robot in an isolated location. The status of the humanoid robot is monitored and communicated to the operator in real time. This improves the telepresence of scenes compared to other methods. On the operator’s side, data are retrieved from the humanoid robots and manipulated to control the robot by estimating the operator’s pose visually. To reduce the risk of inaccurate pose estimation, pose candidates obtained from different points of view are fused while removing irrelevant poses. Then, a simple yet effective cascade filter that generates smooth motion changes is used for stable and flexible robot manipulation. Experiments confirmed that the proposed system can functions efficiently in real time and generates more accurate motion compared to conventional systems.