PDF(Pubmed)
Abstract：
In recent years, the pandemic situation has forced the education system to shift from traditional teaching to online teaching or blended learning. The ability to monitor remote online examinations efficiently is a limiting factor to the scalability of this stage of online evaluation in the education system. Human Proctoring is the most used common approach by either asking learners to take a test in the examination centers or by monitoring visually asking learners to switch on their camera. However, these methods require huge labor, effort, infrastructure, and hardware. This paper presents an automated AI-based proctoring system- \'Attentive system\' for online evaluation by capturing the live video of the examinee. Our Attentive system includes four components to estimate the malpractices such as face detection, multiple person detection, face spoofing, and head pose estimation. Attentive Net detects the faces and draws bounding boxes along with confidences. Attentive Net also checks the alignment of the face using the rotation matrix of Affine Transformation. The face net algorithm is combined with Attentive-Net to extract landmarks and facial features. The process for identifying spoofed faces is initiated only for aligned faces by using a shallow CNN Liveness net. The head pose of the examiner is estimated by using the SolvePnp equation, to check if he/she is seeking help from others. Crime Investigation and Prevention Lab (CIPL) datasets and customized datasets with various types of malpractices are used to evaluate our proposed system. Extensive Experimental results demonstrate that our method is more accurate, reliable and robust for proctoring system that can be practically implemented in real time environment as Automated proctoring System. An improved accuracy of 0.87 is reported by authors with the combination of Attentive Net, Liveness net and head pose estimation.
摘要：
近年来,疫情迫使教育系统从传统教学转向在线教学或混合学习。有效监视远程在线考试的能力是限制教育系统中这一阶段在线评估的可扩展性的因素。人类学习是最常用的方法，可以要求学习者在考试中心参加考试，或者通过视觉监控要求学习者打开相机。然而,这些方法需要大量的劳动力，努力,基础设施,和硬件。本文通过捕获考生的实时视频，提出了一种基于AI的自动化监考系统-“考勤系统”，用于在线评估。我们的细心系统包括四个组件来估计错误，如人脸检测，多人检测，脸欺骗，和头部姿势估计。注意网络检测面并绘制边界框以及置信度。注意网还使用仿射变换的旋转矩阵检查面的对齐。将人脸网算法与Attentive-Net相结合来提取地标和面部特征。通过使用浅CNN活跃度网仅针对对齐的面部启动用于识别欺骗面部的过程。使用SolvePnp方程估计考官的头部姿势，检查他/她是否在寻求他人的帮助。犯罪调查和预防实验室（CIPL）数据集和具有各种类型的不当行为的定制数据集用于评估我们提出的系统。大量的实验结果表明，我们的方法更准确，可靠和健壮的监督系统，可以在实时环境中实际实现为自动监督系统。作者报告了改进的准确度为0.87，结合了注意网，活体网和头部姿势估计。
