| Summary: | 碩士 === 國立高雄第一科技大學 === 電腦與通訊工程研究所 === 104 === In this thesis, the real-time hand gesture recognition systems are proposed by using Kinect v2, digital image processing and computer vision. There are two kinds of systems to be studied. One system uses the contour of shape as the core feature, the other system considers texture as the core feature.
The hand gestures to be recognized are numbers from zero to nine. The proposed method considers the changes of complex background, high degree of freedom of hand gesture and the differences among palms. The aim of this paper is to solve the scaling, rotating, translating and complex-background problems of hand gestures such that the gestures can be recognized instantly and correctly by proposed method.
In the proposed real-time recognition system, color, depth and human skeleton information of gesture are first gotten from Kinect v2. Then, human skeleton and depth information are used to locate right-hand image and segment the palm region. Next, combining the depth with skin color model, complex background is filtered out. And, the contour or texture features of hand gesture image are extracted by digital image processing method. Finally, the support vector machine (SVM) classifier with multivariate linear classification is used to recognize the numbers from zero to nine.
Experimental results show that the proposed real-time gesture recognition systems are both resistant to complex background, rotation, translation and scaling influence. The average recognition time of two systems are both equal to 29ms. For the recognition system using the features of contour of shape, the average recognition rate is 97.4%. For the recognition system using the texture features, the average recognition rate is 92.5%.
|
|---|
