Analysis Of Examinee’s Abnormal Behavior Based On Joint Keypoint Detection

In order to ensure the fairness of the educational examination,the examination institution judges whether there are abnormal examinees during the examination process by manually browsing the invigilation video.However,as the number of examinees increases year by year,the number of invigilation videos increases sharply.Using manual processing to analyze videos will consume a lot of human resources and time costs,and the quality of video review cannot be guaranteed.Therefore,an abnormal behavior analysis method of examinees based on joint keypoint detection is proposed to automatically analyze the invigilation videos.The specific work is as follows:(1)A dataset containing 30,000 images was produced for the problems of incomplete detection of examinee positions,missed detection of examinees with bowed heads,and false detection of invigilators.Determine the location of examinees by uniting them with their desks to eliminate the influence of invigilators during the examination.Adjust the anchor box size of the object detection model according to the clustering results,and use the multi-scale feature map to generate the corresponding candidate box,which improves the accuracy of examinee position detection.(2)For the problem that the original image features contain a lot of invalid information,the examinee’s state is represented by 13 keypoints of the examinee and 4 keypoints of the desk,so that the model’s attention is focused on the examinee.The Self-Attention Graph Neural Networks(GNSA)model is designed by combining the graph neural network and self-attention mechanism,and the Temporal Segment Networks(TSN)framework is used for training,which improves the accuracy of examinee’s abnormal behavior analysis.The abnormal behavior analysis method of examinees based on joint keypoint detection uses graph neural network to extract the relative position relationship between the keypoints of examinees and the keypoints of desks,and uses self-attention mechanism for feature fusion,which can be applied to most examination scenarios and has good universality.The recall rate of the GNSA model is 94.23%,and the precision rate is 96.08%.The accuracy rate is about 20% higher than that of the inflated 3D convolutional network model and TSN model based on video frame sequences,reaching 97.56%.