| 摘要: |
| 为有效利用机器视觉技术实现对电力作业人员穿戴规范进行准确识别，减少安全事故的发生，构建了一种基于姿态感知的穿戴规范识别复合残差网络。该复合网络首先将VGG(Visual Geometry Group)与分裂-转换-聚合(Split-Transfer-Agregation,STA)模块引入残差网络中，构建高性能的ResNeXt50基础网络模块。对ResNeXt50网络不同层次的残差特征图进行聚合与解码处理，实现对人体姿态的估计与关键区域的定位。将卷积块注意力模块（Convolutional Block Attention Module,CBAM)集成到ResNeXt50网络相邻卷积层之间，以提高目标特征的表述能力，从而实现对电力人员穿戴情况进行准确识别。在训练阶段，采用迁移学习实现对预训练网络的顶层参数进行修正，以解决穿戴设备样本图片不足的缺点，从而提高复合网络的识别准确率。通过与SDD、Res-Net50和Inception-v3网络进行对比实验发现，建立的复合网络获得了更高的平均精确率（Mean Average Precision,MAP)值，单帧识别耗时更小，能有效地实现弱小穿戴设备的识别。 |
| 关键词: 目标识别 姿态感知 残差网络 迁移学习 卷积块注意力模块 |
| DOI： |
|
| 基金项目:四川省科技厅项目(2020JDJQ0075)；国网四川省电力公司科技项目(521997190016) |
|
| An OpenPose-based residual network for electric worker’s wearable device recognition |
| CHANG Zhengwei,PU Wei,WU Jie,HUANG Kunchao,XIONG Xingzhong,CHEN Mingju |
| (Electric Power Research Institute of State Grid Sichuan Electric Power Company,Chengdu 610041,China;School of Automation and Information Engineering,Sichuan University of Science and Engineering,Yibin 644002,China) |
| Abstract: |
| In order to realize the accurate recognition of the wearable device of electric workers and reduce the incidence of accidents by machine vision technology,a hybrid recognition residual network based on OpenPose is proposed.Visual Geometry Group(VGG) and Split-Transfer-Agregation(STA) modules are embedded in the residual network to construct a highly efficient basic network ResNext50.The OpenPose network is used to locate critical areas of the wearers by aggregation and pose-encoding of the ResNext50’s residual feature.Convolutional Block Attention Module(CBAM) is integrated between the adjacent convolutional layers of the recognition network to improve feature representation ability and recognition capability,and the accurate identification of the wearable device is increased.Transfer learning is used to optimize pre-training network by amending Softmax layer to overcome the shortcomings of insufficient image database in the network training stage.Comparative experimental results show that the proposed hybrid network obtains higher Mean Average Precision(MAP) values,consumes less recognition time,and achieves better identifying results of small target than SDD,ResNet50 and Inception-v3. |
| Key words: target detection OpenPose residual network transfer learning convolutional block attention module |
