New Techniques for Hand Pose Estimation Based on Kinect Depth Data

Abstract

In this paper we present new techniques for extension and improvement of approaches to markerless 3D hand pose estimation: A new algorithm for finger segmentation in point clouds of hands is presented which makes use of the narrow and flexible shape of the fingers. The kinematic of a 3D hand model gets aligned to the geodesic paths of the fingers which provides a very natural configuration of the model. Therefore a new technique for optimization of those geodesic paths is introduced. Another benefit of this system is, that it's only data resource is the depth stream of a Kinect. So no additional hardware, markers or training data is needed. A first implementation of our approach provides a proof for the new concepts and looks promising for further investigation.