基于人体属性分析的考场行为识别  

Examination Room Behavior Recognition Based on Human Attribute Analysis

作　　者：姚捃[1,2] 郭志林 赵杰 YAO Jun;GUO Zhi-lin;ZHAO Jie(The Engineering&Technical College,Chengdu University of Technology,Leshan 614000,China;Southwest Institute of Physics of Nuclear Industry,Chengdu 610000,China)

机构地区：[1]成都理工大学工程技术学院,乐山614000 [2]核工业西南物理研究院,成都610000

出　　处：《科学技术与工程》2022年第22期9721-9727,共7页Science Technology and Engineering

基　　金：四川省科学技术厅重点项目(2019YJ0705);西物创新行动项目(201901XWCXRC005);成都理工大学工程技术学院青年科学基金(C122019008)。

摘　　要：在实际监控的边缘设备中利用双流网络(temporal segment networks, TSN)或者3D卷积神经网络(3D convolutional neural network, 3DCNN)网络很难实现实时的、相对准确的监控任务。提出了一种结合人体检测和人体属性分析的考场行为识别算法。相对于以提取时空特征作视频分类算法为主流思想的行为识别,利用视频帧以人体检测和人体属性分析结合的行为识别方法更加快速准确。方法借助了多标签学习、注意力机制和特征金字塔等策略来改进任务,同时利用迁移学习对本地采集的数据集进行再训练,实验结果表明达到了主流数据集的良好性能,并在考场环境具有高效性与实用性。In the actual monitoring edge devices, it is difficult to realize real-time and relatively accurate monitoring tasks by using TSN(temporal segment networks, TSN) or 3 DCNN(3 D convolutional neural network, 3 DCNN) network. An examination room behavior recognition algorithm combining human detection and human attribute analysis was proposed. Compared with the behavior recognition which takes extracting spatio-temporal features as the mainstream idea of video classification algorithm, the behavior recognition method based on video frame and the combination of human body detection and human attribute analysis is more rapid and accurate. The proposed method was improved by means of multi label learning, attention mechanism and feature pyramid. At the same time, the locally collected data sets were retrained by transfer learning. The experimental results show that the good performance on the mainstream data sets is achieved, and it is efficient and practical in the examination room environment.

关 键 词：人体检测 人体属性分析 行为识别 多标签学习 特征金字塔 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]