Scholarlink

<p>Human pose estimation is a core task in computer vision, aiming to accurately locate keypoints of the human body in images. While current deep learning-based networks (such as HRNet) can preserve fine spatial details, their high computational complexity and parameter count hinder their deployment on mobile devices. Existing lightweight methods (such as Lite-HRNet) often sacrifice accuracy, making it difficult to achieve a balance between accuracy and efficiency under resource constraints. To solve this problem, this study proposes a lightweight high-resolution network for efficient human pose estimation, termed LE-HRNet. The study takes HRNet as the baseline model, designs lightweight core modules MPneck and MSblock, embeds depthwise separable convolution and innovative lightweight attention TSEA to systematically reconstruct the lightweight backbone network, significantly reducing the number of model parameters and computational complexity while achieving collaborative enhancement of features in more dimensions; at the same time, the mixed local channel attention (MLCA) is introduced in Stem to enhance the quality of the initial feature map. In addition, by designing the channel dynamic upsampling method SED, channel attention is embedded in the dynamic sampling process to improve the key point positioning accuracy. Experiments show that on the MPII dataset, LE-HRNet reduces the number of parameters and computational complexity by approximately 88 and 86% respectively compared to HRNet-W32, and achieves a detection accuracy of 88.5% PCKh@0.5, which is superior to the existing lightweight models Lite-HRNet (+2.6%) and EL-HRNet (+0.8%). On the COCO val2017 dataset, this model achieved an AP of 70.4%, reducing the number of parameters by approximately 89% and the computational overhead by approximately 90% compared to HRNet-W48, and achieved 70.6% AP on the COCO test-dev2017 dataset. This research provides a high-performance and low-cost solution for real-time pose estimation in resource-constrained scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LE-HRNet: A Lightweight High-Resolution Network for Efficient Human Pose Estimation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月12日

查看全文
  • Weiya Shi,
  • Xinyu Dong,
  • Changjiang Wu,
  • Saiyang Liu,
  • Yue Guo

摘要

Human pose estimation is a core task in computer vision, aiming to accurately locate keypoints of the human body in images. While current deep learning-based networks (such as HRNet) can preserve fine spatial details, their high computational complexity and parameter count hinder their deployment on mobile devices. Existing lightweight methods (such as Lite-HRNet) often sacrifice accuracy, making it difficult to achieve a balance between accuracy and efficiency under resource constraints. To solve this problem, this study proposes a lightweight high-resolution network for efficient human pose estimation, termed LE-HRNet. The study takes HRNet as the baseline model, designs lightweight core modules MPneck and MSblock, embeds depthwise separable convolution and innovative lightweight attention TSEA to systematically reconstruct the lightweight backbone network, significantly reducing the number of model parameters and computational complexity while achieving collaborative enhancement of features in more dimensions; at the same time, the mixed local channel attention (MLCA) is introduced in Stem to enhance the quality of the initial feature map. In addition, by designing the channel dynamic upsampling method SED, channel attention is embedded in the dynamic sampling process to improve the key point positioning accuracy. Experiments show that on the MPII dataset, LE-HRNet reduces the number of parameters and computational complexity by approximately 88 and 86% respectively compared to HRNet-W32, and achieves a detection accuracy of 88.5% PCKh@0.5, which is superior to the existing lightweight models Lite-HRNet (+2.6%) and EL-HRNet (+0.8%). On the COCO val2017 dataset, this model achieved an AP of 70.4%, reducing the number of parameters by approximately 89% and the computational overhead by approximately 90% compared to HRNet-W48, and achieved 70.6% AP on the COCO test-dev2017 dataset. This research provides a high-performance and low-cost solution for real-time pose estimation in resource-constrained scenarios.