동적 베이스망 프레임워크 기반의 양손 제스쳐 인식

Abstract

It is natural to use hand gestures in interacting with computers because hand gestures are freer in movements and much more expressive than any other body parts. In this paper, we define and recognize ten hand gestures including two-hand gestures as well as one-hand gestures. Skin blobs in a frame are segmented by two different skin color models combined. Each skin blob is modeled with a Gaussian model. For the tracking of the skin blobs, we exploit optical flows computed between the blobs in the previous frame and those in the current frame. The new mean of the Gaussian model for each blob in the current frame is predicted using the optical flows which give the motion information of each blob from the previous frame to the current frame. The motion of hands is defined the change of the mean of each Gaussian and the relative position between two hands, each hand and a face. A new gesture recognition model is proposed based on the dynamic Bayesian network framework which is relatively easy to represent the relationship among features and to incorporate new features or information to a model. Experimental results showed high recognition rate up to 99.59% with our small dataset in isolated gesture recognition and 84% of the detection rate, 76.36% of reliability was obtained in continuous gesture recognition. The proposed model and techniques are believed to have a sufficient potential for successful applications to other hand gestures recognition such as sign languages.