摘要
目前,常见的三维(3D)人体姿态估计算法在表征学习上取得很好的效果,但是在人体骨架关节点处依然存在估计精度不佳等问题,因此,如何从单目RGB图像中利用冗余的二维(2D)姿态序列时空信息来估计人体姿态的有效方式是一个研究的难点。本文提出一种基于时空多特征融合网络的三维人体姿态估计算法,具体是结合一种图像外观信息和运动时序信息时空多特征融合层级方法,该方法利用一种紧凑的卷积神经网络(convolutional neural network, CNN)学习时空信息将二维关节点位置信息建模为三维关节点位置。实验结果表明,本文所提出的方法能实现较为先进的端对端姿态估计精度,而且不需要任何后处理阶段的姿态优化方法,本文得到的姿态估计在平均精度上得到有效的提升,证明本文方法能够有效提高人体姿态估计的准确性。
At present, the common 3 D human pose estimation algorithms have achieved good results in representation learning, but there are still problems such as poor estimation accuracy at the joint points of the human skeleton.Therefore, how to effectively estimate human pose from a monocular RGB image using the redundant 2 D pose sequence sptatio-temporal information is a different point in the research.In this paper, a 3 D human pose estimation algorithm based on spatio-temporal multi-feature fusion network is proposed.The method utilizes a compact convolutional neural network to learn spatio-temporal information to model the 2 D joint position information as 3 D joint position.The experimental results show that the method proposed in this paper can achieve relatively advanced end-to-end attitude estimation accuracy, and does not require any attitude optimization method in the post-processing stage.The pose estimation obtained in this paper can effectively improve the average accuracy, which proves that the method in this paper can effectively improve the accuracy of human pose estimation.
作者
叶俊
张云
YE Jun;ZHANG Yun(Key Laboratory of Applications of Computer Technology of Yunnan Province,Kunming,Yunnan 650500,China;School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming,Yunnan650500,China)
出处
《光电子.激光》
CAS
CSCD
北大核心
2022年第12期1306-1314,共9页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61262043)
云南省科技计划项目(2011FZ029)
重点实验室开放基金项目(2020106)资助项目。
关键词
三维人体姿态估计
时空特征
运动补偿网络
特征融合网络
3D human pose estimation
spatio-temporal features
motion compensation network
feature fusion network