| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 101 === This thesis mainly implements human following function for home service robot, May, developed in our laboratory. In order to follow the operator accurately, visual tracking is composed by Tracking-Learning-Detection (TLD) and Kinect skeleton, where TLD plays the role as re-detecting the situation that operator is occluded or disappeared, and Kinect skeleton is adopted to track all other situations while TLD is learning how to enlarge operator image patterns in order to enhance recognition rates. For the sake of improving tracking capability, fuzzy head motion control is added in the visual tracking system to compensate the constraints that the mobile platform of May cannot react rapidly. Every instant movement of the operator can be captured by fuzzy head motion control in real time. Q-learning is applied to discover the pose switching of the mobile platform such that May possesses more robust following ability. By Q-learning, states setting are based on three dimensional position, actions are created by the pose of four wheel independent steering and four wheel independent driven (4WIS4WID) platform, and rewards are established on state transitions. Finally, both the experimental results in the laboratory and competition consequents of Follow Me Mission in robot@home league at RoboCup Japan Open 2013 Tokyo demonstrate that our robot May can fluently switch its poses to follow operator by utilizing the proposed schemes.
|
|---|
