一种基于CNN与位姿自适应的运动模型生成方法
A Motion Generation Method Based on CNN and Pose Adaptation
童立靖 1徐光亚 1冯金芝1
作者信息
- 1. 北方工业大学信息学院,北京 100144
- 折叠
摘要
针对基于数据驱动的人体运动模型生成网络存在的模型生成精度不高,生成模型稳定性不好的问题,提出了一种基于CNN与位姿自适应的三维人体运动模型生成方法.首先,为提升人体运动模型生成网络中编码器对输入运动序列的特征提取效果,在变分自编码器中引入CNN卷积网络,便于更好地从运动数据中提取运动特征;其次,在运动生成网络中引入卷积网络,从而完成运动特征、路径参数与生成模型的自适应映射,并最终生成沿设定路径行进的三维人体运动模型.实验结果表明,与CAE算法以及CAE改进算法相比,该方法有效降低了重构损失,能够生成更加准确、自然的三维人体运动模型.
Abstract
A 3D human motion model generation method based on CNN and pose adaptation is proposed to address the issues of low accuracy and poor stability of data-driven human motion model generation network.Firstly,to improve the feature extraction performance of the encoder in the human motion model generation network for input motion sequences,a CNN convolutional network is introduced into the variational autoencoder to better extract motion features from motion data.Secondly,a convolutional network is introduced into the motion generation network to achieve the adaptive mapping of motion features,path parameters,and generation model.Finally,a 3D human motion model is generated along the set path.The experimental results show that compared with CAE algorithm and CAE improved algorithm,the proposed method can effectively reduce the reconstruction losses and generate a more accurate and natural 3D human motion model.
关键词
卷积神经网络/位姿自适应/运动模型生成/特征提取Key words
convolutional neural network/posture adaptation/motion model generation/feature extraction引用本文复制引用
出版年
2024
NETL
NSTL
万方数据