基于注意力细化的实时人体姿态估计网络  

Real-time human pose estimation network based on attention refinement

作　　者：叶永芳 石守东[1] 蓝艇[1] 邱科迪 赵天翔 YE Yongfang;SHI Shoudong;LAN Ting;QIU Kedi;ZHAO Tianxiang(Faculty of Electrical Engineering and Computer Science,Ningbo University,Ningbo 315211,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《传感器与微系统》2025年第2期104-107,共4页Transducer and Microsystem Technologies

基　　金：浙江省公益技术应用研究资助项目(LGF22F020029);中国创新挑战赛(宁波)项目(2022T001)。

摘　　要：由于人体姿态的复杂性和多样性,取得高准确度的网络模型往往具有大量的参数和复杂的结构,给部署和应用带来一定困难。本文提出一种基于注意力细化的实时人体姿态估计网络,使用多分辨率图像金字塔嵌入模块融合图片中多分辨率的空间特征,减少单分辨率下空间特征损失;使用注意力细化网络进行人体姿态估计,对全局注意力特征进行局部细化,降低参数量与计算量,实现实时的人体姿态估计。本文方法达到701 fps的推理速度,与推理速度为174 fps的HRFormer-small相比,提升4倍,但在精度上略低0.34%。Due to the complexity and diversity of human pose,achieving high accuracy in human pose estimation network models often requires a large number of parameters and complex structures which bring a certein diffivulty to deployment and application.A real-time human pose estimation network based on attention refinement is proposed.A multi-resolution image pyramid embedding module is used to fuse spatial features of multiple resolutions,in image and reduce spatial feature loss under a single resolution.Additionally,an attention refinement network is utilizd for human pose estimation,which refines the global attention features locally,thereby reducing the parameter and computational quatities and enabling real-time human pose estimation.The proposed method achieves an inference speed of 701 fps,which is four times faster compared to the HRFormer-small model with an inference speed of 174 fps.However,there is a slight decrease in precision by 0.34%.

关 键 词：人体姿态估计 TRANSFORMER 注意力细化 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]