首页|基于沙漏注意力高分辨率网络的人体姿态评估实验

基于沙漏注意力高分辨率网络的人体姿态评估实验

扫码查看
原文链接
  • 万方数据
  • 维普
基于人体关键点检测的方法需要融合由高分辨率到低分辨率子网络生成的表示,以提高关键点检测的准确性.设计了一种基于沙漏注意力高分辨率网络的人体姿态评估方法实验,在深度高分辨率表示学习的基础上构建沙漏注意力特征模块,并设计特征回传模块和多阶段监督算法,用融合中继监督和自蒸馏的方式实现高分辨率网络的监督训练.与经典方法在标准数据集上完成人体姿态评估实验对比,并在硬件设备上进行了真实场景实验,实现了行人姿态评估和危险行为报警.
Experiment of Human Posture Assessment Based on Hourglass Attention High Resolution Network
The method based on human key point detection needs to integrate representations generated by high-resolution to low-resolution sub networks to improve the accuracy of key point detection.However,in the process of feature extraction from low resolution to high resolution,although rich information exchange units are added between different branches of the high-resolution network,the mainstream information on each branch does not undergo the process of information aggregation from high to low.Therefore,a human posture evaluation method experiment based on the hourglass attention high-resolution network is designed in this paper.The hourglass attention feature module is constructed on the basis of the deep high-resolution representation learning,and the feature retrieval module and multi-stage supervision algorithm are designed to achieve the supervision training of the high-resolution network by fusing relay supervision and self-distillation.Compared with the classical method,the human posture estimation experiment is completed on the standard data set,and the real scene experiment is carried out on the hardware equipment to realize the pedestrian posture evaluation and dangerous behavior alarm.

body posture assessmentkey point detectionattention mechanismself distillationmulti stage supervision

云霄、褚菲、张晓光、程小舟

展开 >

中国矿业大学信息与控制工程学院,江苏徐州 221116

人体姿态评估 关键点检测 注意力机制 自蒸馏 多阶段监督

中国矿业大学教学改革项目江苏省高教学会评估委员会项目教育部产学合作协同育人项目教育部产学合作协同育人项目教育部高等学校自动化类专业教指委教学改革项目

2021YB202021-C158202002109018202002177005202123

2024

10.19927/j.cnki.syyt.2024.01.042

实验室研究与探索
上海交通大学

实验室研究与探索

CSTPCD北大核心
影响因子：1.69
ISSN：1006-7167
年,卷(期)：2024.43(1)
  • 17