Monocular 3D Hand Pose Estimation for Teleoperating Low-Cost Actuators

摘要

So far, in order to interact with a virtual agent that simulates hands or to teleoperate a robotic hand it is required any kind of device like joysticks, gamepads or VR controllers. These devices map a button press to a predefined action, usually with no fine control of the entity that is being teleoperated, which is unsuitable and unpleasant. In this work, we study different setups to perform 3D hand pose estimation from color images of hands with the aim of controlling a low-cost robotic hand. The experiments we carried out demonstrated that it is possible to do so with a deviation of less than 5 mm per joint.