Data-driven Inverse Dynamics for Human Motion

摘要

Inverse dynamics is an important and challenging problem inrnhuman motion modeling, synthesis and simulation, as well as inrnrobotics and biomechanics. Previous solutions to inverse dynamicsrnare often noisy and ambiguous particularly when double stancesrnoccur. In this paper, we present a novel inverse dynamics methodrnthat accurately reconstructs biomechanically valid contact information,rnincluding center of pressure, contact forces, torsional torquesrnand internal joint torques from input kinematic human motion data.rnOur key idea is to apply statistical modeling techniques to a set ofrnpreprocessed human kinematic and dynamic motion data capturedrnby a combination of an optical motion capture system, pressurerninsoles and force plates. We formulate the data-driven inverserndynamics problem in a maximum a posteriori (MAP) frameworkrnby estimating the most likely contact information and internal jointrntorques that are consistent with input kinematic motion data. Wernconstruct a low-dimensional data-driven prior model for contactrninformation and internal joint torques to reduce ambiguity ofrninverse dynamics for human motion. We demonstrate the accuracyrnof our method on a wide variety of human movements includingrnwalking, jumping, running, turning and hopping and achieve stateof-rnthe-art accuracy in our comparison against alternative methods.rnIn addition, we discuss how to extend the data-driven inverse dynamicsrnframework to motion editing, filtering and motion control.