一种基于语义层消息传递的人体姿态估计方法  

A human pose estimation method based on semantic message passing

作　　者：文政颖 冯广丽 WEN Zhengying;FENG Guangli(College of Software,Henan University of Engineering,Zhengzhou 451191,China)

机构地区：[1]河南工程学院软件学院,河南郑州451191

出　　处：《河南工程学院学报（自然科学版）》2024年第2期58-61,共4页Journal of Henan University of Engineering：Natural Science Edition

基　　金：河南省高等学校重点科研项目(23A520052)。

摘　　要：针对基于卷积神经网络的人体姿态估计中不能充分利用相邻关节信息特征的问题,提出了一种基于语义层消息传递的人体姿态估计方法。该方法可在全卷积神经网络得到的特征向量之间进行消息传递,并可以根据关节之间的距离采用带奖惩机制的消息传递策略。在体育姿势数据集和单人人体关键点检测数据集上进行验证,实验结果表明该方法的准确率较高,性能较好。Aiming at the problem that the feature of adjacent joins can not be fully utilized in the research of human pose estimation based on convolutional neural network(CNN),this paper proposes a semantic message passing based human pose estimation method.The messages are passed among feature maps obtained by full CNN.A message passing strategy with reward and penalty mechanism according to the distance among joints is proposed.The method was validated on the leeds sports pose(LSP)and the frames labeled in cinema(FLIC),the experiment results indicated that this method had a good performance.

关 键 词：姿态估计 深度学习 卷积神经网络 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]