| Recently,network technology has developed rapidly,monitoring systems are broadly used in different fields,such as communication,transportation and security.The skynet system has entered the streets.With the improvement of public security awareness,surveillance cameras are widely used in the field of public security,like subway stations,high-speed railway stations and other occasions.For the field of public security,identifying abnormal events and abnormal human behavior in surveillance videos is an important task.Nowadays,surveillance cameras are widely used in examination occasions.In important exams such as the middle school entrance examination and the college entrance examination,there are surveillance cameras to assist the invigilator to invigilate the exam,which ensures a more fair and just exam.In the face of these massive surveillance videos,manual identification consumes a lot of manpower and material resources.Due to the particularity of the exam room,it is unrealistic to invest a lot of manpower for monitoring and identification.Therefore,in order to detect the abnormal behavior of examinees,this thesis designs a model for detecting abnormal behavior of examinees in exam rooms with the surveillance videos of exam rooms and deep learning methods.This thesis also develops a surveillance system using the model to indentify the abnormal behavior of examinees in exam rooms,which helps to timely find the possible abnormal behavior of examinees in exam rooms.The surveillance system could ensure exams more fair and impartial.The thesis proposes an exam room dataset consisting of real exam room surveillance videos.The dataset differs significantly from other publicly available behavior recognition datasets because the exam room has a complex scene and many examinees.The differences are mainly in the following aspects.There are a large number of examinees in a exam room.Many examinees are heavily shielded from each other.Therefore,what is difficult is that identifying the behavior of each examinee.In addition,the number of abnormal behavior of examinees is less than that of normal behavior.The dataset has a serious data imbalance problem.Data annotation needs a large amount of human and material resources.Due to the angle of the camera,examinees who are close to the camera are big and those who are away from the camera are small.In the field of behavior recognition,small target behavior recognition has been neglected.In the exam room dataset,there are many behavior of small targets,which increases the difficulty of identification.To address the above issues,the main work of this thesis is as follows:(1)In this thesis,we propose an algorithm for detecting abnormal behavior of examinees.We construct an eaxm room dataset consisting of real exam room videos.We spend a lot of effort on annotating the dataset.The dataset includes precise spatiotemporal annotations for each examinee.To identify the behavior of examinees,we propose a two-stage approach,object detection and behavior recognition.In object detection,the model uses EMD Loss,which improves the accuracy of detecting examinees.In behavior recognition,the model uses different data augmentation methods and feature pyramid structure,which improve the accuracy of the model in detecting the behavior of examinees.To solve the data imbalance,the weighted loss function and the weighted sampling strategy are adopted to improve the ability of recognize abnormal behavior.The model achieves a high accuracy in the exam room dataset.(2)This thesis designs a system to detect behavior of examinees.The system process is as follows.Firstly,we get the video stream data through web interfaces of cameras.Secondly,the model deals with the video stream data and gets the category of behavior and location information for each examinee.Thirdly,the system visualizes the behavioural category and bounding box of each examinee.Finally,the staff detect the behavior of examinees through visual information,which makes exams more fair and impartial. |
PDF Full Text Request