融入双注意力和姿态增强的3D人体姿态估计
3D Human Posture Estimation with Dual Attention and Posture Enhancement
高翔 1刘韦华2
作者信息
- 1. 西北工业大学计算机学院,陕西 西安 710129
- 2. 西北工业大学软件学院,陕西 西安 710129
- 折叠
摘要
尽管3D人体姿态估计已经在高速发展,但现有的3D人体姿态估计模型对特征的判别能力较弱,无法有效地获取多通道以及空间特征信息,仿真效果受到影响.对此,以VideoPose3D和PoseAug作为基础网络进行改进得到融入双重注意力以及姿态增强的高效姿态估计网络CSNet.通过在结合了PoseAug的姿态估计器里融入通道注意力和空间注意力,构造CS-Block模块作为基础模块,提升特征的精准度和全面度,以应对遮挡、深度的不确定性等问题.在公开数据集Human 3.6M进行验证和测试,结果显示,相比于原来的模型,上述方法的平均关节位置误差降低了3.4%,提升了 3D人体姿态估计模型的精确度,并利用改进后的算法识别人体关键点并驱动虚拟人物仿真,获得更好的仿真效果.
Abstract
Although 3D human pose estimation has been developing at a high speed,the existing 3D human pose estimation models have a weak ability to distinguish features,and cannot effectively obtain multi-channel and spatial feature information,and the simulation effect is affected.To this end,this paper uses VideoPose3D and PoseAug as the basic networks to improve and get an efficient attitude estimation network CSNET that integrates dual attention and gesture enhancement.By integrating channel attention and spatial attention into the pose estimator combined with PoseAug,the CS block module is constructed as the basic module to improve the accuracy and comprehensiveness of features,so as to deal with the problems such as occlusion and uncertainty of depth.The verification and test on the public data set Human 3.6m show that compared with the original model,the average joint position error of this meth-od is reduced by 3.4%,and the accuracy of the 3D human pose estimation model is improved.The improved algo-rithm is used to identify the key points of the human body and drive the virtual character simulation to obtain better simulation results.
关键词
人体姿态估计/通道注意力/空间注意力/姿态增强/虚拟人物Key words
Human pose estimation/Channel attention/Spatial attention/Pose augmentation/Virtual characters引用本文复制引用
基金项目
陕西省重点研发计划(2020ZDLSF04-02)
出版年
2024
NETL
NSTL
万方数据