Real-Time Hand Pose Estimation from RGB-D Sensor

Hand pose estimation in cluttered environment is always challenging. In this paper, we address the problem of hand pose estimation from RGB-D sensor. To achieve robust real-time usability, we first design a data acquisition strategy, using a color glove to label different hand parts, and collect a new training data set. Then a novel hand pose estimation framework is presented, so that feature fusion drives hand localization and hand parts classification. Moreover, instead of using articulated model, a simplified and efficient 3D contour model is designed to assist real-time implementation, which does not require a large amount of training data. Experiments show that our approach can handle real-time hand interaction in a desktop environments with cluttered background.