基于高分辨率网络的人体姿态估计  

Human pose estimation algorithm based on high-resolution net

作　　者：朱翠涛[1] 李博 ZHU Cuitao;LI Bo(Hubei Key Laboratory of Intelligent Wireless Communication,South-Central Minzu University,Wuhan 430074,China)

机构地区：[1]中南民族大学智能无线通信湖北省重点实验室,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2023年第2期229-237,共9页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金资助项目(61671483);湖北省自然科学基金资助项目(2016CFA089)。

摘　　要：针对高分辨率网络中存在不同分支特征交叉融合导致参数量大、运算复杂度高等问题,提出了一种基于高分辨率检测网络(HRNet)的人体姿态估计优化网络模型.引入空洞空间卷积池化金字塔替代多分辨率分支网络交叉融合过程,同时引入注意力机制,提高网络输出特征图质量,从而保证改进后网络检测的准确度.在环境配置和网络输入图像分辨率一致的情况下,所提出的模型在COCO数据集上实验结果较HRNet相比参数量下降38.6%,运算复杂度下降35.2%.实验结果表明:改进后网络在检测精度略微下降的情况下,能有效降低参数量、运算复杂度.Aiming at the problems of large amount of parameters and high computational complexity caused by the cross fusion of different branch features in high-resolution network,a human pose estimation optimization network model based on high-resolution detection network(HRNet)is proposed.The atrous spatial pyramid pooling is introduced to replace the cross fusion process of multi-resolution branch networks,and the convolutional block attention module is introduced to improve the quality of network output feature map,so as to ensure the accuracy of improved network detection.When the environment configuration is consistent with the resolution of the network input image,the results show that the parameters of this model on COCO dataset are reduced by 38.6%compared with HNRet,and the computational complexity is reduced by 35.2%.The experimental results show that the improved network can effectively reduce the amount of parameters and computational complexity when the detection accuracy decreases slightly.

关 键 词：姿态估计 高分辨率网络 空洞卷积 人体检测 关键点相似度 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]