文献类型：期刊文献

中文题名：基于轻量级注意力模块的多人姿态估计

英文题名：Multi-person Pose Estimation Based on Light-weight Attention Module

作者：杨竣乔[1];钱锋[1];唐漾[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237

年份：2023

卷号：30

期号：9

起止页码：1743

中文期刊名：控制工程

外文期刊名：Control Engineering of China

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：国家自然科学基金资助项目(61988101);国家杰出青年科学基金资助项目(61925305);重点国际(地区)合作研究项目(61720106008);中国石油科技创新基金资助项目(2021D002-0902)。

语种：中文

中文关键词：人体姿态估计;轻量级;注意力模块;知识蒸馏

外文关键词：Human pose estimation;light-weight;attention module;knowledge distillation

摘要：为了提高人体姿态估计的检测精度,特别是中小尺寸人体关键点的检测精度,同时针对模块集成造成网络参数量和浮点计算量大幅度增加的问题,以HRNet(high-resolution Net)为基本网络架构,提出了一种基于轻量级注意力模块的分阶段网络(Coarse-Refine Net,CRNet),采用知识蒸馏方法,利用同构的高精度网络作为教师网络,利用关键点的真值和教师网络的输出结果共同训练学生网络(CRNet)。通过在2017COCOval集上进行训练和测试,CRNet获得了78.2的得分(%),比HRNet基准模型提高了1.5%。对于中小尺寸人体关键点的检测,CRNet获得了74.8的得分(%),比HRNet基准模型提高了1.6%。在相似的网络结构下,CRNet相比于HRNet,网络参数量仅增加了0.1M(0.3%),浮点计算量仅增加了0.1G(0.6%)。与现阶段的最新方法相比,CRNet获得了更高的检测精度,同时基本没有增加网络参数量和浮点计算量。
A Coarse-Refine Net(CRNet)based on light-weight attention module is proposed to improve the detection accuracy of human pose estimation,especially the detection accuracy of medium-sized and pony-sized human key points.Meanwhile,we also consider the increasing of the network parameters and the floating-point operations per second caused by module integration.We take advantage of the knowledge distillation method that uses the truth values of key points and the output results of the teacher network to train the student network(CRNet).Extensive experiments are designed on COCO data set,CRNet gets a score of 78.2% on the 2017COCO val set,which is 1.5% higher than HRNetssss.For the detection accuracy of medium-sized and pony-sized human key points,CRNet gets a score of 74.8%,which is 1.6% higher than HR Net.Compared with HRNet in similar network structure,the number of network parameters and the amount of floating-point operations per second in CRNet are only increased by 0.1M(0.3%)and 0.1G(0.6%).CRNet gets higher accuracy than the state of the art,and barely increases the network parameters and the floating-point operations per second.