Two-phase optimized inverse kinematics for motion replication of real human models

This paper proposes a novel two-phase optimization method to solve the joint variables of a real human model using captured motion data. A dual-mode 3D vision system was used to generate a body geometric model. Twenty-three link segments were separated from the body model. Markers were then affixed on each link for motion capturing by the same vision system. A body kinematic model was then constructed from the geometric model by assigning joint constraints between every two adjacent links. The model contains five kinematic chains with 48 joint freedoms. Joint variables were solved by the first-phase optimization to obtain an appropriate initial posture between two adjacent links. For each kinematic chain, redundant joint variables were solved using the second-phase optimization. By setting proper weightings, the resulting postures of the kinematic model closely match up with those of the motion captured data, while the endpoints can trace the original trajectories as well. Furthermore, the kinematic model offers more reasonable motions than the geometric model that replicates the raw motion data since some excessive movements are corrected by the joint constraints.