실시간 손 포즈 인식을 통한 비접촉식 입력 인터페이스

Abstract

In this paper, we proposes a contact-less input interface in a smart glass environment through real-time recognition of hand poses without using any keyboard or mouse. First, a histogram distribution chart of depth values is created through the first-order differentials of depth values in depth camera input images, and inflection points are calculated to segment the areas of left and right hands. Then the hand rotation angle and hand center point of each hand area are calculated. After that, the centroid point of the hand boundary crossing is determined while expanding the circle at a fixed interval from the center of each hand so as to detect the hand joint points and end points. The hand information obtained before is matched with the hand model of the previous frame to recognize a hand pose and the hand model is updated for the next frame. In addition, virtual keyboard is registered by calculating the position of the palm while expanding the circle from the center of left hand and the keyboard click is detected by comparing the depth values of index fingertip to those of its adjacent pixels. This method enables hand pose recognition through the hand model of the previous frame by using the temporal consistency between continuous frames and allows the virtual keyboard input with no physical input device in the wearable environment. Recognition experiments for various hand poses showed over 95% average recognition rate and over 80% accuracy rate for virtual keyboard clicks. Furthermore, the proposed method showed the processing speed of about 34 frame per second, thus enabling real-time processing.