英文摘要 | With the development of the computer intelligence, human-computer interaction technology has become increasingly important. Gesture interaction technology is a hand-centered intuitive technology. Human hand, considered as one of the most flexible parts, can express various visual messages. Recently, vision-based gesture interaction technology has been widely used in living, entertainment, education, health and industry. The development of depth camera further increases the applications with gesture interaction technology, especially in education and entertainment, because natural interaction is more acceptable. However, there are still some deficiencies. Firstly, color-based segmentation is easy to be influenced by the illumination and complex background. Secondly, the collecting and processing large-scale training data cost a lot of manpower and resources. Thirdly, traditional implementation methods cannot meet the need of real-time applications. So, it is a challenging task to develop vision-based gesture interaction technologies with the features including usability, expansibility, accuracy and real-time performance. On the basis of fully understanding the research status and the development trend of the current hardware and software, this dissertation studies the hand gesture interaction technology based on depth images from the perspective of natural interaction. According to the mode and content of the hand gesture interaction technology, hand gesture recognition, hand posture recognition and 3D hand tracking have been studied deeply. The main work of this dissertation is summarized as follows: (1) As for hand gesture recognition, a 3D gesture recognition method based on one-shot-learning is proposed in consideration of the spatial-temporal features of the gesture. Firstly, a template-based self-adaptive head tracking method combined with a region growing approach is presented for human detection and segmentation. Secondly, gesture is represented by two features: 3 views motion history images and their pyramid histogram of oriented gradient vectors. Thirdly, an action selection method including action segmentation and informative frame selection is proposed for successive gestures. Finally, the proposed action selection and representation methods are employed together for one-shot-learning gesture recognition. The correlations of images and vectors are employed for recognition. This method has a higher recognition rate, needs no training data and... |
修改评论