污染场地修复环境中人体姿态估计算法研究  

HUMAN POSTURE ESTIMATION ALGORITHM IN CONTAMINATED SITE REMEDIATION ENVIRONMENT

作　　者：张宝峰[1] 田宇 朱均超[1] 刘娜[1] Zhang Baofeng;Tian Yu;Zhu Junchao;Liu Na(Engineering Research Center of Optoelectronic Devices&Communication Technology,Ministry of Education,P.R.China,Tianjin Key Laboratory for Control Theory&Application in Complicated System,Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津理工大学天津市复杂系统控制理论及应用重点实验室光电器件与通信技术教育部工程研究中心,天津300384

出　　处：《计算机应用与软件》2023年第2期212-216,235,共6页Computer Applications and Software

基　　金：天津市互联网跨界融合创新科技重大专项(18ZXRHSF00240)。

摘　　要：针对污染场地修复环境中背景复杂、人体遮挡、视点变化等突出问题,提出一种基于沙漏网络的人体姿态估计算法,在有效过滤复杂背景的同时,提高姿态估计的准确性与鲁棒性。该算法利用感受野与注意力机制,对沙漏网络中的传统残差模块与跳级连接结构进行了改进。其通过扩大有效感受野面积,提高了人体关键点之间的关联性;通过对人体区域添加掩模,保留住关键人体信息的同时,过滤掉复杂背景。实验表明,提出的模型在MPII多人数据集上mAP检测精度达到83.1%,在MSCOCO Test-dev数据集上平均精度较Mask R-CNN、RMPE模型分别提升了9.6百分点和0.4百分点。Aiming at the prominent problems of complex background, human occlusion and viewpoint change in the contaminated site remediation environment, we propose a human posture estimation algorithm based on hourglass network, which can effectively filter the complex background and improve the accuracy and robustness of posture estimation. This algorithm used the receptive field and attention mechanism to improve the traditional residual module and skip connection structure in hourglass network. By expanding the area of effective receptive field, it improved the correlation between key points of human body. By adding a mask to the body area, it retained the key body information and filtered out the complex background. The experimental results show that the proposed model has an average accuracy of 83.1% mAP on the MPII multi-person data set. On the MSCOCO Test-dev data set, the average accuracy is improved by 9.6 and 0.4 percentage points respectively compared with Mask R-CNN and RMPE model.

关 键 词：多人姿态估计 沙漏网络 感受野 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]