改进的ST-GCN单人姿态估计算法研究  

Research on Improved ST-GCN Single Pose Estimation Algorithm

作　　者：史健婷 王印冉 詹怀远 SHI Jian-ting;WANG Yin-ran;ZHAN Huai-yuan(School of Computer and Information Engineering,Heilongjiang University of Science andTechnology,Harbin 150022,China)

机构地区：[1]黑龙江科技大学计算机与信息工程学院,黑龙江哈尔滨150022

出　　处：《计算机技术与发展》2025年第1期61-66,共6页Computer Technology and Development

基　　金：2023年度黑龙江省属高校基本科研业务费项目(2023-KYYWF-0547)。

摘　　要：近年来,单人姿态估计广泛应用在各个领域,降低单人姿态估计算法对标记数据的依赖同时提高其准确率是计算机视觉中一个具有挑战但是十分重要的课题。针对此问题,该文提出一种改进的时空图卷积神经网络(Spatio-Temporal Graph Convolutional Networks,ST-GCN)的方法。在原来的ST-GCN的基础上,融合MoveNet轻量级神经网络,利用MoveNet的关键点识别功能,解决ST-GCN需要预先标注关键点数据的问题。引入SimAM注意力机制,解决原来的ST-GCN不能很好地区分通道中重点信息,将所有的信息一视同仁的问题。增加ReLU6-Sigmoid组合激活函数,解决原有的激活函数训练波动,非线性拟合不足的问题。即:在提高了原时空图卷积神经网络的检测精度的同时,减少了应用过程中对于标记数据的依赖,降低了训练时的损失率精确率的波动。对于改进后的时空图卷积神经网络,在FLORENCE 3D ACTIONS数据集上证明了其有效性。结果表明,改进后的时空图卷积神经网络准确率从0.8695提升到0.956521。F1值由0.887566提高到0.965432。In recent years,single-person pose estimation has been widely used in various fields.Reducing the dependence on labeled data and improving the accuracy of single-person pose estimation algorithm is a challenging but very important topic in computer vision.To solve this problem,we propose an improved Spatio-Temporal Graph Convolutional Networks(ST-GCN)method.On the basis of the original ST-GCN,the MoveNet lightweight neural network is integrated,and the key point recognition function of MoveNet is used to solve the problem that ST-GCN needs to pre-label key point data.The SimAM attention mechanism was introduced to solve the problem that the original ST-GCN could not distinguish the key information in the channel well and treat all the information equally.The ReLU6-Sigmoid combination activation function was added to solve the problem of training fluctuation and insufficient nonlinear fitting of the original activation function.That is,it improves the detection accuracy of the original spatio-temporal graph convolutional neural network,reduces the dependence on labeled data in the application process,and reduces the fluctuation of the loss rate accuracy during training.For the improved spatio-temporal graph convolutional neural network,its effectiveness is proved on the FLORENCE 3D ACTIONS dataset.The results show that the accuracy of the improved spatio-temporal graph convolutional neural network is improved from 0.8695 to 0.956521.F1 value increased from 0.887566 to 0.965432.

关 键 词：计算机视觉 改进的时空图卷积神经网络 模型融合 SimAM ReLU6-Sigmoid 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]