With the development of computing technology, computers have been widely used in our daily life. As a bridge between human and computer, HCI (Human Computer Interaction) has become a very important component of computer application. The traditional approaches of HCI, such as mouse, keyboard, and pen, are too cumbersome to maximally exploit the power of computers in some emerging applications such as virtual reality, augmented reality and wearable computing etc. Recently, hand gesture interaction, as a promising approach, attracts more and more attention. Especially, vision based hand gesture interaction has become the mainstream due to its simplicity, intuitiveness, and unintrusiveness. In this dissertation, our research mainly focus on some issues on hand posture recognition, hand tracking, classifiers improvement for hand posture and the applications of gesture interaction etc. In the above fields, the major achievements and contributions are summarized as following: (1) A fast scale-space feature detection method is proposed. A set of simple rectangular feature templates are used to approximate Gaussian derivative convolution templates in traditional scale-space feature detection and the fast detectors and descriptors for scale-space geometric shape features are obtained. With the approximation by rectangular feature templates, the computational complexity is greatly reduced. Experiments on standard dataset and the natural scene dataset show that the proposed method significantly reduces time cost of gesture recognition while keeping comparable accuracy with traditional method. (2) The strategy of multiple collaborative kernels is introduced into hand tracking. The motion of hand gesture is decomposed into motion of sub-targets such as palm and fingers. Kernel tracking with SSD and background weighted histogram is employed to the tracking of each sub-target. (3) A hand posture recognition method with co-training strategy is proposed. Semi-supervised and multi-view learning are introduced into hand posture recognition. Two classifiers with Haar and HoG are respectively initiated on a small labeled sample data. This method with co-training utilize the complementary action between different features and improve the performance of classifiers in a semi-supervised framework. (4) A real-time vision-based gesture interaction interface for image browsing is designed and implemented, which combines hand detection, tracking and gesture recognition. Then gesture recognition is performed with fast scale-space feature detection. Finally，the recognition results drive the operations such as shifting cursor position and open/close image preview in the browsing interface. Under the framework combining detection, tracking and recognition, the demo interface gets satisfactory results.
修改评论