«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

32

期数:

2022年09期

页码:

174-179

栏目:

新型计算应用系统

出版日期:

2022-09-10

文章信息/Info

Title:

An End-to-end Multi-objective Behavior Recognition Algorithm for Examination Room

文章编号:

1673-629X(2022)09-0174-06

作者:

姚 捃; 郭志林

成都理工大学 工程技术学院,四川 乐山 614000

Author(s):

YAO Jun; GUO Zhi-lin

School of Engineering & Technique,Chengdu University of Technology,Leshan 614000,China

关键词:

多目标检测; 端到端; 行为识别; 多标签学习; 特征金字塔

Keywords:

multi target detection; end-to-end; behavior recognition; multi label learning; feature pyramid

分类号:

TP391. 41

DOI:

10. 3969 / j. issn. 1673-629X. 2022. 09. 027

摘要:

人工监考存在监考人员容易疲惫、监考行为缺乏客观的执行准则、违规行为证据无法留存等问题,因此越来越多的高校建设了智能化教室,并在教室开始实施利用行为识别进行自动化的监考任务,以期在监考工作中解放人工劳动的同时提供公平公正客观的监考程序。 在实际考场监控的边缘设备中利用 TSN 双流、3DCNN 等结合时空特征的网络很难实现实时的、相对准确的监控任务。 提出一种端到端的考场多目标行为识别算法。 相对于以提取空间、时序特征并进行融合为主流思想的行为识别算法,利用视频帧以多目标检测和多目标行为识别相结合的行为识别算法在考场环境中更加快速准确。 算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考场行为视频数据集进行再训练,得到最终的考场行为识别模型,结果表明达到了主流数据集中上水平,并在考场环境中具有良好的高效性与准确性。

Abstract:

There are some problems in manual invigilation, such as invigilator fatigue, invigilator behavior lack of objective implementation criteria,violation evidence cannot be retained. There fore,more and more colleges and universities have built intelligent classrooms and began to implement the task of automatic invigilation by using behavior recognition in the class rooms,so as to liberate manual labor in the invigilation work and provide? ? a fair and objective invigilation procedures. In the edge equipment of actual examination room monitoring,it is difficult to realize real - time and relatively accurate monitoring tasks by using TSN dual stream,3DCNN and other networks combined with temporal and spatial characteristics.? ?An end-to-end multi-objective behavior recognition algorithm for examination room is proposed. Compared with the behavior recognition algorithm which takes extracting spatial and temporal features and fusing as the mainstream idea,the behavior recognition algorithm based on? the combination of multi - target detection and multi-target behavior recognition in video frames is faster and more accurate in the examination room environment. The algorithm improves the task with the help of multi label learning,attention mechanism and feature pyramid. At the same time,the locally collected examination room behavior video data set is retrained by transfer learning to obtain the final examination room behavior recognition model. The results show that it has reached the upper level of the mainstream data set,and has high efficiency and accuracy in the examination room environment.

常用功能

更新日期/Last Update: 2022-09-10