基于时空双流3D残差网络的服务动作识别  

SERVICE ACTION RECOGNITION BASED ON SPATIOTEMPORAL DUAL-STREAM 3D RESIDUAL NETWORK

作　　者：欧阳黎 林彤尧 程莺 彭冰莉 温和[3] Ouyang Li;Lin Tongyao;Cheng Ying;Peng Bingli;Wen He(State Grid Hunan Electric Power Limited Company Power Supply Service Center(Metrology Center),Changsha 410004,Hunan,China;Hunan Province Key Laboratory of Intelligent Electrical Measurement and Application Technology,Changsha 410004,Hunan,China;College of Electrical and Information Engineering,Hunan University,Changsha 410000,Hunan,China)

机构地区：[1]国网湖南省电力有限公司供电服务中心(计量中心),湖南长沙410004 [2]智能电气量测与应用技术湖南省重点实验室,湖南长沙410004 [3]湖南大学电气与信息工程学院,湖南长沙410000

出　　处：《计算机应用与软件》2023年第6期112-117,共6页Computer Applications and Software

基　　金：国网湖南省电力有限公司科技项目(5216A020002U);国家自然科学基金项目(61771190)。

摘　　要：人工抽查服务录像是目前企事业单位规范员工服务动作的主要途径,但这种方法费时费力。针对当前供电营业厅服务动作自动识别难题,提出基于时空双流3D残差网络的服务动作识别方法,建立供电营业厅服务动作数据集。将时空双流3D残差网络分为两个通道。RGB通道采用3D残差网络提取信息丰富的RGB图像提升对动作幅度较小的识别率;光流通道采用C3D网络提取能消除场景信息的光流图特征。根据训练时两个通道的识别率,分配对应的通道融合权重值。对两个通道的预测结果进行加权融合得到服务动作识别结果。实验结果表明,使用时空双流3D残差网络对供电营业厅服务动作的识别准确率为90.65%。Manual random inspection of service video is the main way for enterprises and institutions to regulate employee service actions,but this method is time-consuming and labor-intensive.Aimed at the current automatic recognition of service actions in power supply business halls,a service action recognition method based on spatiotemporal dual-stream 3D residual network is proposed.A service action data set of power supply business halls was established.The spatiotemporal dual-stream 3D residual network was divided into two channels.RGB channel used the 3D residual network to extract information-rich RGB images to improve the recognition rate for smaller motion ranges;and optical flow channel used the C3D network to extract optical flow graph features that could eliminate scene information.According to the recognition rate of the two channels during training,the corresponding channel fusion weight value was assigned.The prediction results of the two channels were weighted and fused to obtain the service action recognition result.The experimental results show that the accuracy of the spatiotemporal dual-stream 3D residual network on the service action data set of the power supply business hall reaches 90.65%.

关 键 词：动作识别 3D残差网络 光流 供电营业厅 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]