通信与信息技术Issue(5)：38-43,80,7.
面向边缘计算设备的多人姿态估计算法改进
Improvements in Multi-Person pose estimation targeting edge computing devices
宋建辉 1姚彰 1刘晓阳 1赵亚威1
作者信息
- 1. 沈阳理工大学 自动化与电气工程学院,辽宁 沈阳 110159
- 折叠
摘要
Abstract
To address the challenge of large computational demands in multi-person pose estimation algorithms,which hinder de-ployment on edge computing devices,a lightweight improved algorithm based on the YoloPose model is proposed.First,the Anchor-Free detection head is replaced with an Anchor-Based detection head,and the SEAM module is added,reducing the model's computational load while enhancing the recognition capability for occluded skeleton points.Second,the neck network continues to use PANet with the addition of a P6 layer,and the UPsample module is replaced with Dysample,strengthening the connection between key skeletal points and the background.Third,LSKA is integrated into SPPF to improve semantic fusion between different feature layers in the backbone net-work.Finally,the DAttention module is added to the C2F module to increase the connectivity between skeletal points of the same individ-ual.Experimental results demonstrate that the improved model reduces computational complexity by 8.3 GFLOPS compared to the origi-nal YoloPose model,while achieving a 7.49%increase in mAP accuracy.The enhanced model performs well in detecting small individu-als and complex backgrounds,and its reduced computational load is more suitable for edge computing devices.关键词
Yolo/骨骼点检测/深度学习/轻量化/边缘设备/注意力机制Key words
YOLO/Skeletal point detection/Deep learning/Lightweight/Edge devices/Attention mechanism分类
交通工程引用本文复制引用
宋建辉,姚彰,刘晓阳,赵亚威..面向边缘计算设备的多人姿态估计算法改进[J].通信与信息技术,2025,(5):38-43,80,7.
