Study On Gesture Recognition In Kinect's Depth Image

In the recent years, Human-Computer Interaction(HCI) has become one of the most important part in our works and lives. With the rapid development of computer technology, the HCI is more and more required to be easy and natural. Gestures are easy, natural and straight forward interactions and have become the most important research direction of visual based HCI.By taking actions and poses of human's hands as input, users can define own visual based gesture interaction to control the devices. It improves the experience of HCI and is supposed to be an important direction of future HCI. However, current visual based gesture interaction technology still faced with several challenges such as the diverse semantics of the gestures and temporal and spatial variance of the interactions. This dissertation introduces the existing work of domestic and overseas researchers, and proposes a novel visual based gesture recognition method. The proposed RGB-Depth image based real-time gesture recognition method generally consists of three parts: hand detection, hand feature extraction and hand classification.For hand detection, depth and color information are both used to accurately locate hands area in the images with complex background. A depth histogram based adaptive thresholding method is adopted for the depth image, while an improved Bayesian skin-color detection is performed on corresponding color images to detect skin area. The final hand regions are detected by fusing the above two results with a region-growing algorithm.For hand feature extraction, the histogram of orient gradient(HOG) feature is adopted for hand images description. This feature has been successfully applied in pedestrian detection and is proven to be qualified to describe the hand images as well. The HOG feature is robust for the geometrical invariability and illumination variation.For hand classification, Extreme Learning Machine(ELM) classifier and Support Vector Machine(SVM) classifier are used to recognize different gestures respectively. The accuracy and real-time performance of hand classification are both examined by off-line and on-line experiments. Experiments show that the proposed gesture recognition method achieves high recognition accuracy in real-time.