| Summary: | 碩士 === 國立嘉義大學 === 資訊工程學系研究所 === 106 === The great fervor of virtual reality (VR) has been aroused from 2016 until now. In order to achieve a better interaction in VR environments, many third-party manufacturers are committed to develop novel VR products. Virtual reality gloves, which can accurately capture user's hand movements, are currently the most important development projects; however, price of the whole set of an equipment is very expensive. To the best of our knowledge, Leap Motion is the most conventional popular hand-tracking device used in virtual reality market. It provides a low-cost and convenient hand interactive interface, but some wrong gesture-tracking results lead to wrong interaction outcomes. Therefore, this thesis improves the tracking accuracy of Leap Motion and provides a lower cost and more stable system in comparison with the original Lead Motion.
This study analyzes Leap Motion's original signals, explores tracking data of Leap Motion's false gestures, and further corrects errors manually to produce a modified dataset for training learning models. We divide machine-learning process into two convolutional neural network models to locate palm area and to detect 2D positions of fingertips. The output of the model is applied the Application Programming Interface provided by Leap Motion to locate 3D fingertips. Finally, the system uses inverse kinematics to restore a full hand gesture and we evaluate experimental results with Leap Motion default's gestures. The hand tracking method proposed in this thesis solves the problem of tracking the wrong number of fingers using Leap Motion, and it implements virtual reality applications with HTC Vive to provide users with more stable and accurate tracking results.
|
|---|
