面向考场视频中作弊行为的层次式检测方法  

A Hierarchical Detection Method for Cheating Behavior in Examination Room Videos

作　　者：骆祖莹[1] 万桢洪 李玉顺[1] LUO Zuying;WAN Zhenhong;LI Yushun(Beijing Normal University,Beijing 100875,China)

机构地区：[1]北京师范大学,北京100875

出　　处：《中国考试》2023年第5期45-52,共8页journal of China Examinations

基　　金：国家自然基金面上项目“层次行为事件模型启发的课堂教学行为模式挖掘及其关键技术研究”(61977009)。

摘　　要：针对当前考场视频作弊行为检测精度低、难以实际应用的问题,本文提出一种高效的作弊行为层次式检测方法。该方法首先使用帧间差分,快速筛选出包含异常行为的关键帧图像;随后采用ResNet50作为分类器进行基于骨骼图像的行为识别,检测出疑似作弊行为;最后使用改进的三维卷积(3D-CNN,简称C3D)网络作为主干网络,融合时空注意力机制,构建C3D+attention动作识别网络,使用包括疑似作弊行为在内的单人图像序列单元,精确检测出作弊行为。实验结果表明:本方法融合二维卷积(C2D)网络速度快和C3D网络精度高的优点,能够高效地检测出数量稀少的作弊行为视频片段,对光照变化和低分辨率视频具有强鲁棒性,使智能监考成为可能。This paper proposes an efficient hierarchical cheating behavior detection method to address the problem that cheating behavior(CB)detection in examination room videos is inaccurate and hard to use in practice.The method first uses the inter-frame difference to filter out keyframes including abnormal behaviors quickly,and then uses ResNet50 as the classifier for behavior recognition based on skeletal images to detect suspicious CB.Finally,the method employs the improved three-dimensional convolution(3D-CNN,C3D)network as the backbone and integrates the temporal&spatial attention mechanism to construct the C3D+attention action recognition network,which uses a unit of single-person image sequences that includes a suspicious CB to recognize the CB accurately.The results show that the method combines the advantages of fast speed and high accuracy from two-dimensional convolution(C2D)and C3D networks respectively,which means it can effectively detect a small number of cheating video clips and owns strong robustness to illumination changes and low-resolution videos,making intelligent invigilation possible.

关 键 词：作弊行为 智能监考 行为识别 人体姿态估计 卷积神经网络 

分 类 号：G405[文化科学—教育学原理]