| In recent years,the construction of standardized examination venues has received wide attention.To some extent,it can improve the management level of examinations and ensure the fairness and impartiality of examinations.At present,the traditional manual supervisory examination method not only requires a huge cost of human resources,but also may have problems that supervisors miss the examination.With the development of computer vision technology,it has become a trend to apply artificial intelligence technology to the examination room environment.Therefore,this paper puts forward an algorithm for examinee behavior detection based on computer vision technology in the examination field environment.It identifies the abnormal behavior of examinee by combining multiple network models,such as target detection network,Alphapose human posture estimation network and space-time convolution neural network,and designs an examination behavior recognition system.The specific research contents are as follows:(1)Select the lightweight target detection network Yolov4-Tiny to complete the examinee target positioning task.In order to solve the problem of occlusion or small target in the monitoring video for candidates in the examination room environment,CBAM attention mechanism is embedded in the main feature extraction part of Yolov4-Tiny network,which can effectively extract fine-grained information of features.Adding PPM pyramid pooling structure after feature extraction can improve the network’s ability to obtain global information.The experimental results show that the improved Yolov4-Tiny network is 6.8%higher than the original network model mAP,and it can effectively solve the problem of small candidates’targets and obscuring targets in the examination room.(2)Construct the candidates’ posture skeleton model.Taking into account the candidates’sitting status under video monitoring,this paper extracts 12 joint point skeleton coordinate information from the candidates’ upper body through the Alphapose human posture estimation model,and establishes the candidates’human posture skeleton diagram model by renumbering the skeleton joints.(3)Using the joint point data of the examinee skeleton as the input vector,the examinee behavior recognition task is completed through the space-time convolution network.Because the spatial-temporal graph network tends to ignore the structural relationship between joint points when extracting spatial-temporal features from skeleton data,and cannot make full use of the time dependence between discontinuous frames and different sequence lengths.Therefore,by incorporating a channel-space cascade attention mechanism into the network layer of the space-time map and a multi-scale time convolution kernel into the time convolution,this paper can better extract fine-grained information from the skeleton data and effectively deal with the different complex time-span categories of candidates’action behaviors.The attention-focused and multi-scale space-time map network is analyzed experimentally on NTU RGB+D,Kinetics-skeleton public skeleton datasets.The experimental results show that the improved space-time map network model has better performance in human motion behavior.The overall fusion algorithm in this paper is validated on the self-made data set of the candidates’skeleton behavior.The results show that the average accuracy of the four types of behavior recognition is 94.6%,and the accuracy is much higher than other models,which can effectively complete the candidates’ behavior detection tasks in the examination room.(4)To further verify the effectiveness of the algorithm for examinee behavior recognition,this paper designs an examinee behavior recognition system,which incorporates the algorithm model in this system.This system can analyze the information of the examinee’s behavior category more intuitively,and capture the examinee with abnormal behavior in real time.To some extent,the design of this system can help supervisors to complete their decision-making quickly. |
PDF Full Text Request