Research On The Identification Of Abnormal Behavior Of Candidates Based On Machine Learning

At present,hardware and software technologies are becoming increasingly mature and those based on them have been developed.However,passive monitoring is still the mainstream form of examination monitoring,which only records the entire examination video and still requires time to retrieve any abnormal behaviours.Therefore,if abnormal behaviours in the examination video could be automatically retrieved through artificial intelligence technology,more human costs could be saved.To address this problem,this paper proposes a candidate cheating model with improved Spatial-Temporal Graph Convolutional Networks(ST-GCN)for cheating behaviours detection and YOLOv3 contraband detection,which can perform skeletal key point detection on multiple candidates in the examination room at the same time,and based on the key point The model can be trained to extract useful feature vectors based on the key point information,and can better identify the three abnormal behaviours of head reaching,hand reaching and standing.It also enables the detection of three types of prohibited items present in the examination room,namely mobile phones,school bags and books.The main research work and innovation points of the thesis are as follows.(1)An improved ST-GCN model is proposed.The basic ST-GCN model has two problems:firstly,when extracting the spatio-temporal features of human action through convolution operation in the process of action recognition,it is easy to ignore the different effects of each joint point of the human body on the motion,and the basic ST-GCN assigns the same weight to different joint points,which cannot achieve the dynamic adjustment of the weight of the joint points in the process of recognition,easily leading to the problem of inadequate acquisition of spatio-temporal features of the joint points.Secondly,the use of single-layer temporal convolution in the basic ST-GCN temporal dimension feature extraction process easily leads to the problem of insufficient extraction of temporal features due to the poor performance of inter-action dependencies.To solve the problems of the basic ST-GCN,this paper incorporates an attention mechanism based on the basic ST-GCN to achieve dynamic adjustment of nodal weights;at the same time,the temporal convolution layer is extended from a single layer to a two-layer network structure,and the residual network is fused,so as to achieve the effect that feature data can be fully extracted.The experiments are conducted on the NTU RGB+D public dataset,and show that the improved ST-GCN model improves 6.9% over the unimproved STGCN in terms of accuracy;and 4.4% and 5.4% over the single fused attention mechanism and single extended network structure respectively.(2)Proposed improved ST-GCN with YOLOv3 for cheating detection model.Through the experimental analysis of the improved ST-GCN on the dataset in the real environment of the examination hall,the improved ST-GCN model only recognizes the action behaviors appearing in the candidates,but cannot achieve the detection of prohibited items.Therefore,based on the action recognition,the lightweight YOLOv3 algorithm is added as the basis to build the target detection framework.The model is trained on the real data set,and after experiments,it is shown that adding the detection of items to the action recognition,the correct rate is improved in the detection of candidate cheating.(3)Proposed improved ST-GCN with YOLOv3 for cheating detection model.Through the experimental analysis of the improved ST-GCN on the dataset in the real environment of the examination hall,the improved ST-GCN model only recognizes the action behaviour appearing in the candidates,but cannot achieve the detection of prohibited items.Therefore,the lightweight YOLOv3 algorithm is added to the action recognition as the basis for building a target detection framework.Experiments show that adding the detection of items to action recognition can improve the correct rate of detecting candidates’ cheating behaviour.(4)Finally,a validation algorithm simulation system is designed to visualize the examination data and help the invigilator to analysis and achieve the inference results of the desired model.