ABSTRACT
This paper proposes a novel method for automatic extraction of human body skeletons from 3D dynamic shapes by taking advantage of geometry properties of static shapes, knowledge of anthropometry (priori knowledge of human body anatomy), and kinematics characteristics of dynamic shapes. The method consists of three steps. In Step I, the human body is divided into five parts including four limbs and the torso based on the Reeb graph created by a Morse function, which is defined on the surface mesh of a reference pose as geodesic distanee. Then, limbs and torso are sub-segmented using statistical anthropometric proportions. In Step II, seed points that exhibit the best rigid transformation are identified in the middle parts of the sub-segments by the mesh edge-length deviation induced by its transformation through time. In Step III, we use these seed points to determine joint locations of the skeleton. An empirical experiment demonstrates that the proposed method is able to extract skeletons from 3D dynamic meshes automatically. The accuracy of the proposed method appears to be supelior to that of the existing methods.
