Object Detection in Online Proctoring Through Two Camera Using Faster-RCNN

Abstract

The COVID-19 pandemic has prompted changes in teaching methods from offline to online, including the implementation of exams. But many reports say that the potential for online exam cheating is very high which can compromise the credibility of the exam. The online exam monitoring system using one camera makes it difficult for officers to make decisions because of the lack of evidence and supporting data. In this study, we propose a monitoring approach using two cameras, namely a camera on a laptop to get a front view of the participant and a cellphone camera to get a side view of the examinee but because of the complexity of the problem, at this stage we only focus on the side camera. Implementation begins with the collection of video recording data, custom data sets for training and pretrained datasets from the zoo model. Training is carried out using a custom dataset to detect objects that are not recognized by the pretrained dataset. The evaluation of the training results using the COCO evaluator showed the average of the bbox-AP is 59,169. The fraud detection process is carried out using 6 exam videos with a total of 192,929 frames, producing two outputs, namely object detection videos and csv files. The csv file contains the frame number, time, object detected in each frame. The next process is to analyze the csv file and mark frames that have the potential to be fraudulent. The evaluation results show an accuracy of 0.884615385 and a recall of 0.821428571