Research On Motion Retargeting Method For Motion Data Represented By Joint Position

At present,3D character animation has been widely used in many fields.Skeletal animation is one of the principal ways to generate character animation.However,there is a contradiction between the limited collection and unlimited demand of skeletal motion data needed in the production of skeletal animation.In order to resolve the above contradiction,motion retargeting should be applied to retarget the motion data of a certain character to other characters with heterogeneous topologies or different bone length and proportion,so as to realize the reuse of motion data.On the other hand,with the popularization of simple motion capture equipment such as depth cameras and cameras,more and more motion data are expressed in the form of joint position.Under this background,this thesis explores a method of retargeting skeletal motion data to realize the reuse of motion data represented by joint position.This thesis focuses on the following aspects:(1)First,study the method of motion retargeting between characters with heterogeneous topologies.This thesis presents a ring generation network,which consists of two generated network models.The network is trained by motion data represented by joint position and the loss function defined by cyclic consistency error and intermediate result constraint error.And the training data of the network are two unpaired human skeleton motion data sets.After the network training,input the motion data of the source character and the target skeleton into the network,and the network can quickly generate the motion data retargeted to the target skeleton.(2)Secondly,we study the motion retargeting between characters with the same skeleton topologies and different bone length and proportion.We present an all-purpose bidirectional recurrent autoencoder,which can retarget the motion data from a source to any target character.It is trained from the motion data based on joint position representation with the reconstruction error as a loss function.After training,the hidden units and reconstructed motion of the corresponding source motion data are calculated by the autoencoder.Then,we impose the bone length constraints,foot trajectory constraints,the root joint position constraints and bone-to-bone angle constraints on the reconstructed motion,and the cost is projected back into the hidden-unit space and optimize the hidden units iteratively.(3)Based on the above two algorithms,a fast motion retargeting system is designed and developed.The system implements the above two kinds of motion retargeting functions.Users only need to input the source role motion data and target skeleton data,and the system can output the retargeted motion data.The system can simplify the production process of skeletal motion data in the process of character animation.