计算机应用研究2026，Vol.43Issue(3)：948-953,6.DOI:10.19734/j.issn.1001-3695.2025.05.0174
基于改进堆叠沙漏网络的人体姿态估计
Human pose estimation based on improved stacked hourglass network
摘要
Abstract
This paper developed a lightweight and efficient network architecture to address three major limitations of stacked hourglass networks in complex human pose estimation tasks:weak semantic perception due to the lack of attention mechanisms,poor geometric modeling from fixed convolutional sampling,and low inference efficiency caused by structural redundancy.It in-troduced a spatial-channel dual-path coordinated attention module to enhance keypoint perception and suppress background in-terference in the spatial dimension,while selecting high-level semantic features in the channel dimension for multidimensional feature optimization.It introduced a polymorphic linear deformable convolutional bottleneck module to improve geometric modeling of complex poses by leveraging heterogeneous initial sampling shapes.It introduced an ELA-PCCW hourglass module to significantly reduce model complexity while preserving feature integrity.Experiments on MPII and COCO2017 datasets dem-onstrate that the proposed method improves PCKh@0.5 by 2.3 percent points on MPII,with parameter and computation reduc-tions of 9.1M and 6 GFLOPs,respectively,achieving a good balance between accuracy and efficiency.Comparative studies and visualization analysis further verify the superiority of the proposed method in diverse complex human pose estimation tasks.关键词
人体姿态估计/堆叠沙漏网络/轻量化模型/注意力机制/线性可变形卷积/几何建模能力/特征融合/模型推理效率Key words
human pose estimation(HPE)/stacked hourglass network(SHN)/lightweight model/attention mechanism/linear deformable convolution/geometric modeling capability/feature fusion/inference efficiency分类
信息技术与安全科学引用本文复制引用
吕超,马歌谣..基于改进堆叠沙漏网络的人体姿态估计[J].计算机应用研究,2026,43(3):948-953,6.基金项目
国家重点研发计划资助项目(2024YFC2207103) （2024YFC2207103）
吉林省自然科学基金资助项目(20240101345JC) （20240101345JC）
