Real-time 3D Arm Pose Estimation from Monocular Video for Enhanced HCI

In this paper an approach for 3D arm pose estimation from a monocular video is presented. Our proposal has been designed to provide real-time and realistic reconstruction of the user motion, as required by advanced Human Computer Interaction (HCI) applications. Both a 2D arm tracking and a 3D arm pose estimation algorithm are introduced and discussed. Tracking exploits fast and robust segmentation of the arm silhouette together with detection and tracking of skin colored regions. 3D pose estimation relies on a stickfigure arm model and the Analysis-by-Synthesis approach, but achieves real-time performance using geometrical constraints on tracking results to reduce the search space cardinality. Experiments on the animation of 3D avatars using off-the-shelf hardware demonstrate the effectiveness and real-time performance of our proposal.