Finger posture estimation using 3D medial axes

We propose a method for tracking a hand in real-time, using depth information from a single Kinect sensor. The fingers are segmented using medial axes on an image generated by the depth discontinuities. The palm position and orientation is estimated from the position of the wrist as well as the distance to the closest contour points. The detected medial axes are projected in the palm plane for labelling, and then used for direct posture estimation of unoccluded phalanges. Inverse kinematics provide the missing information in case of occlusion. Our approach can be related to an inverse kinematics approach based on fingertip detection, with additional robustness to fingertip occlusion provided by the medial axis of each visible phalanx.