Real-Time Articulated Hand Detection and Pose Estimation

We propose a novel method for planar hand detection from a single uncalibrated image, with the purpose of estimating the articulated pose of a generic model, roughly adapted to the current hand shape. The proposed method combines line and point correspondences, associated to finger tips, lines and concavities, extracted from color and intensity edges. The method robustly solves for ambiguous association issues, and refines the pose estimation through nonlinear optimization. The result can be used in order to initialize a contour-based tracking algorithm, as well as a model adaptation procedure.