Research On Examination Events Detection For Intelligent Invigilation System

The intelligent invigilation system is the key and difficult part of the intelligent examination management system.It is of great significance for the education administration to improve the examination management level,reduce human and financial investment in the examination process,and ensure the fairness of the examination.At present,there are still many key and difficult problems that hasn't been solved in the research of intelligent invigilation system.This paper studies and solves the key problems of examination event detection in the intelligent invigilation system based on the exam-room surveillance video,and proposes the corresponding theoretical models and key algorithms.Specifically,it includes three issues:The first issue is global examination event detection.It refers to the identification and classification of overall states of every segment videos based on the time dimension,combined with the predicted text parameters such as examination time and rule,to judge the empty exam-room,candidates entering the exam-room,candidates leaving the exam-room,test preparing and closing,exam ongoing,whether the order is normal and on time.In addition,confirmation of the stage of exam ongoing is the premise of local examination event detection.Aiming at this problem,this paper proposes a two-channel 3D convolutional neural network model D3DCNN.The experimental proves the validity and advancement of the D3DCNN model on the exam-room surveillance video data set EMV-1 and the benchmark data set UCF101.The second issue is recognition of exam-room surveillance video digital clock.In exam-room surveillance video,there are 3 elements about examination events,namely the exam-room video semantics,clock information,and test rules.Examination events are closely related to time,and the same type of examination events happening at different times have completely different semantics and importance.The video clock is also the key of massive video surveillance event retrieval and multi exam-room surveillance synchronization.Therefore,the exam-room surveillance video digital clock recognition is the premise for the global and local examination events.This paper studies the problem of video clock recognition in exam-room surveillance,and proposes some related methods to this problem.The experimental part of this paper verifies the effectiveness of these methods on exam-room surveillance video dataset.The third issue is local examination events detection.It refers to the events such as invigilator's tour and abnormal walking of candidates during the stage of exam ongoing.The core problem is to detect and distinguish the target of local standing person objects in the video frame during the stage of exam ongoing.This paper proposes a lightweight and efficient ISSD target detection neural network model,which realizes the detection of local standing person objects in exam ongoing scene.Combined with the exam-room video digital clock information and test rules,the feature fusion and target distinguish algorithm are used to realize the detection of local examination events.The experimental part of this paper proves the validity and advancement of the ISSD model and its migration training strategy in the exam-room dataset EMV-2.There are 3 main points on contribution and innovation in this paper:1.The research on surveillance video of exam-room is mostly focused on the identification of the behavior of individual candidates' violations.This paper creatively studies all the personnel in exam-room as a group object,and distinguishes the classification and recognition of the group object motion state scene.The global status of different test stage,combined with test rules and video clock of exam-room,realizes the global examination events detection and has important use in examination management.2.Aiming at the problem of video scene classification in the global state of exam-room,this paper proposes a novel two-channel 3D convolutional neural network D3DCNN,which extends 2D convolutional neural network to 3D space.The two channels of the model can learns the spatio-temporal features of the original frame sequence and the optical flow frame sequence as the same time,and extracts and fuses the spatial and motion information contained:in the two channels,improves the ability of the network to express the spatio-temporal features of video frame sequence of exam-room surveillance.In terms of network structure design,D3DCNN model reduces the network depth and widens the network width.Reducing the model parameter and training dificulty,and achieves efficient training under small data sets.3.Aiming at the local standing person objects detection problem,this paper proposes a lightweight and efficient ISSD target detection neural network,which aims to improve the classic SSD network and adopts the lightweight MobileNet network as the benchmark.The network uses depthwise separable convolution layer as a new detection layer to maintain network consistency while greatly reducing network parameters and training difficulty.The ISSD model moves the detection layer forward and merges to improve the expression and detection capabilities of the network for small and medium-sized targets,and removes disproportionate candidate frames to improve detection efficiency.The ISSD model also applies the training strategy of migration learning to the small data set of the exam-room,which improves the expression and detection ability of standing person objects in video of exam-room surveillance.