Finger Detection by Renewed FEMD Method using Kinect Depth Camera Image

• Main Authors: GAO, ZHEN-JIE, 高振傑
• Other Authors: Wang, Kuo-Hua
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/18527072041332052585

碩士 === 輔仁大學 === 資訊工程學系碩士班 === 104 === The production cost reduction of depth sensor makes it cheaper and cheaper than ever before. Such a trend also provides many research hints of using different sensing ways for Human-Computer Interface (HCI) applications. Among the integrated sensors sold in the market, the Kinect senor is relatively cheaper than the others. The Kinect sensor consists of RGB camera, depth sensor, and multi-array microphone. It makes a great progress of using the Kinect sensor for human body tracking, face recognition, and human action recognition. For hand gesture recognition, with respect to the whole body recognition, the gesture analysis region is even smaller, more delicate and complex. Therefore, hand gesture is still an open problem and is under the development stage. In this thesis, we improve the Finger-Earth Mover Distance (FEMD) gesture recognition method used in「Part-Based Hand Gesture Recognition System」. This recognition system uses RGB image as the input image and fixed thresh for the threshing decomposition. The proposed improvements consist of two aspects. Firstly, the hand gesture images are captured with the depth image rather than RGB image, which can effectively reduce the noise on hand image while doing the Background Subtraction and skin color detection failure. Moreover, for the threshing decomposition, we replace the fixed thresh by the relative distance thresh to the circle of the palm. It can immediately compute the corresponding optimal thresh for the recognition of different hand gestures with respect to various types of palms. By our experimental result, our proposed method can increase the recognition rate from 92.9% to 98.2% and the average recognition time is less than 34 ms.