| With the development of deep learning technology,education has gradually stepped into the era of intelligence.How to realize the modern management of classroom teaching has become the focus of discussion in the field of education.In the traditional classroom,teachers get feedback on the teaching situation by observing students’ performance.However,this one-to-many method has some problems,such as incomplete observation,low credibility,poor real-time performance,time-consuming and labor-intensive.Therefore,Applying the object detection algorithm based on deep learning to the detection of students’ behavior in classroom scenes can help teachers to fully grasp the students’ classroom state,timely and reasonably adjust the teaching progress and strategies,and improve the teaching efficiency.Up to now,although the target detection algorithm has been applied in many fields and achieved good detection results,it is difficult to use it directly in the actual classroom.So this paper puts forward an improved model of students’ classroom behavior detection based on YOLOv4.The specific work is as follows:(1)Construct a dataset of student classroom behavior.Since there is currently no public student behavior database,this paper collects classroom surveillance videos of the school,screened video frames to obtain a total of 1855 images,and selects three types of abnormal behaviors of classroom students with high frequency of playing mobile phones,sleeping,and talking to each other as detections category.(2)Aiming at the problem that the detection target of students’ behavior is small and occluded,a U-shaped feature extraction network is proposed.By adding up-sampling convolution layer and jumping connection,the fusion of features at different levels is enhanced,and the network’s ability to capture high-level detail information and bottom-level contour information is improved.In addition,the CBAM attention module is introduced into the skip connection to suppress the interference of background information,thereby improving the feature extraction ability and feature utilization efficiency of the model.Secondly,a densely connected module is added in the small-scale prediction stage to increase the dimension of the feature map,strengthen the information transmission of small receptive fields,so that small targets can be detected more easily.Compared with the YOLOv4 algorithm,the YOLOv4-ST model proposed in this paper effectively improves the detection accuracy of students’ behavior.(3)Aiming at the problems of fewer samples of student data set and weak generalization ability of the trained model,a data augmentation method based on adversarial attack is proposed.The iterative fast gradient notation method is used to expand the original samples,and the original samples and the expanded samples are composed together.The data set composed of the original sample and the extended sample is used to train the network,which increases the network robustness,effectively improve the network performancet.Even in the case of small disturbance noise,the model can have a good detection effect.The experimental results show that the speed and accuracy of the improved YOLOv4-ST model for the detection of abnormal behavior of students in the classroom are better than other mainstream detection models,which can better meet the needs of practical applications.It has certain applicable value and research significance for students’ classroom behavior detection. |
