Research On Anomaly Detection Method Of Examination Surveillance Video Based On Contrastive Learning And Pretext Tasks

In recent years,in the face of the increasingly mature monitoring technology,monitoring systems have been seen everywhere,generating a large amount of monitoring video data.In scenarios where automatic monitoring technology is not implemented,human resources are heavily used to detect anomalies in surveillance videos.Such a monitoring method not only consumes resources,but also loses the timeliness of monitoring in some important monitoring scenarios.For example,in an examination room,as invigilators cannot monitor every student all the time,some cheating behaviors may be neglected.Therefore,in order to save human resources and monitor more efficiently,video anomaly detection technology has attracted much attention.The author collects the monitoring video data of examination and uses related technologies in the field of video anomaly detection to detect the cheating behaviors of examinees,which helps invigilators to find and locate the examinees who have cheating behaviors,so as to ensure the fairness of examination.The video data used in this thesis are all from the real monitoring video data of examination.According to different examination rooms,there are differences in lighting,monitoring position and angle,resulting in different scales of examinees.Examinees who are far away from the surveillance camera are small.In addition,there are occlusion problems between examinees.All of these factors increase the difficulty of detection.Since the target of the detection is examinees,invigilators in an examination room are distractions and should be distinguished from examinees.In addition,due to the huge amount of monitoring data,labeling all the data is a time-consuming and labor-intensive task.In the field of deep learning,having a large amount of data is a strong support for the model to achieve good performance.In general,the cheating behaviors of examinees vary widely,so it is almost impossible to collect all cheating behaviors.Therefore,how to define these cheating behaviors and use limited data to efficiently extract the features of examinees is another major challenge faced by this task.In view of the above problems,the author proposes an anomaly detection method of examination surveillance video based on contrastive learning and pretext tasks.The main contributions are as follows:(1)In order to make the model ignore irrelevant background information and focus on analyzing the behavior of examinees,the author proposes to add object detection and object tracking modules to the abnormal behavior detection method.The examinee and his/her desk as a whole are identified in the object detection and object tracking modules,thereby forming an action video clip for each examinee.This approach not only distinguishes between invigilators and examinees,but also facilitates the positioning of cheating examinees.(2)The author proposes an anomaly detection method of examination surveillance video based on contrastive learning and pretext tasks,which uses both normal and abnormal data for training.By contrastive learning,the model learns features that are more discriminative for distinguishing normal and abnormal behaviors,but is insensitive to pixel-level features and resists interference caused by factors such as lighting.At the same time,the model’s understanding of spatial and temporal features is enhanced by using pretext tasks,which effectively provides supervision signals and improves the interpretability of the model.(3)The author collects and labels the monitoring video data of examination.There are two datasets,one is used for the object detection module and the other is used for anomaly detection,these data can provide strong support for the model.Meanwhile,the author considers the cheating behaviors of examinees as abnormal behavior and defines the boundary of normal behavior.All behaviors that are not normal are defined as abnormal behavior,which enables the model to have good detection performance when encountering various unknown abnormal behaviors.Therefore,it is not necessary to collect all kinds of abnormal behaviors to effectively complete abnormal behavior detection.In this way,the challenge of data collection is alleviated to some extent.(4)A large number of experiments are carried out on the anomaly detection dataset,the experimental results show that the proposed method in this thesis has good performance on detecting cheating behaviors in examination monitoring videos and its performance is better than some current mainstream methods.Meanwhile,ablation studies demonstrate the effectiveness of each module in the proposed method.