面向关节坐标运动数据重定向的通用双向循环自编码器

An All-Purpose Bidirectional Recurrent Autoencoder for Retargeting of Motion Data Represented by Joint Position

针对面向关节坐标表示的骨骼运动数据重定向网络缺乏通用性的问题,提出一种能够实现源骨骼到多种骨骼运动重定向的通用双向循环自编码器.该自编码器由基于关节坐标表示的运动数据以重建误差为损失函数训练得到.在完成训练后,首先用自编码器计算源运动数据对应的隐变量和重建运动,然后对重建运动施加骨骼长度约束、足迹约束、根关节位置约束以及骨骼角度约束,并将损失反向传播至隐变量空间中优化隐变量,通过多次迭代得到重定向后运动.在CMU运动数据库上的实验结果表明,提出的自编码器及4种约束能够实现基于关节坐标表示的运动数据的重定向,并且得到的重定向运动在骨骼长度误差、骨骼角度误差、末端效应器轨迹以及平滑性上具有更好的效果.

We present an all-purpose bidirectional recurrent autoencoder aiming at the lack of the generality of the existing retargeting networks of motion data represented by joint position.The autoencoder can retarget the motion data from a source to any target character.The autoencoder is trained by motion data represented by joint position and the loss function defined by reconstruction error.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.The experimental results on CMU motion database show that the proposed autoencoder and four constraints can implement the retargeting of motion data represented by joint position,and the retargeting results have better effects on bone length error,bone-to-bone angle error,and end effector trajectory and smoothness.