基于深度学习的电力工程现场动作识别研究  

Research on field action recognition of power engineering based on deep learning

作　　者：陈杰 朱力 CHEN Jie;ZHU Li(Nantong Power Supply Branch,State Grid Jiangsu Electric Power Co.,Ltd.,Nantong 226001,Jiangsu Province,China)

机构地区：[1]国网江苏省电力有限公司南通供电分公司,江苏南通226001

出　　处：《信息技术》2020年第12期53-58,共6页Information Technology

基　　金：国网江苏省电力有限公司科技项目(J2019036)。

摘　　要：电力工程现场人员复杂、交叉施工面多,迫切需要借助技术手段提升工程现场安全管控力度,为此文中设计了一种动作识别模型实现对监控图像中危险动作的快速甄别。模型使用卷积神经网络和长短期记忆网络构建了能够在融合多视图基础上实现动作识别的深度学习网络。为克服监控图像常见的遮挡和轮廓模糊问题,模型使用两个摄像机作为图像传感器,并通过图像融合实现图像同步。由于模型的监督训练需要大量图像数据,人工标记非常费时,因此文中还提出了一种可对图像进行自动标记的堆叠卷积自编码模型。实验结果证明了所提出模型的有效性。The power engineering site has complex personnel and many cross-construction areas.It is urgent to use technical means to improve the safety management and control of the engineering site.For this reason,an action recognition model is designed to quickly identify dangerous actions in the surveillance image.The model use convolutional neural networks and Long-ShortTerm Memory to build a deep learning network that can realize action recognition based on fusion of multiple views.In order to overcome the common problems of occlusion and contour blur in surveillance images,the model use two cameras as image sensors,and realizes image synchronization through image fusion.Since the supervised training of the model requires a large amount of image data and manual labeling is very time-consuming,a stacked convolutional autoencoding model is proposede that can automatically label images.The experimental results prove the effectiveness of the proposed model.

关 键 词：动作识别 多视图模型 卷积神经网络 长短期记忆网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]