A Research On Detection And Analysis Of Classroom Students Behaviors Based On Deep Learning

The gradual maturity of computer vision technology and deep learning algorithms has accelerated the transformation of intelligence in various fields,and the application of deep learning algorithms in school classroom teaching scenarios is an area of ?concern.In traditional classroom teaching,students’ classroom behaviors can only be observed manually,which is time-consuming,labor-intensive and inefficient.Therefore,it is of great significance to explore how to use deep learning to analyze students’ classroom behaviors.Aiming at the common behaviors of students in classroom,this dissertation adopts deep convolutional neural networks to detect students’ classroom behaviors,adopts the positional relationship and feature similarity of the same individual student between neighboring video frames to correlate individual student,and adopts statistical analysis to quantify the behaviors of students in the classroom.The main work of this dissertation includes:1.Construct a database of students’ behaviors in the classroom.Since there is currently no public database of students’ classroom behaviors,this dissertation collected a large amount of data in undergraduate classes,and a training dataset is created with the annotation tool,which divides the students’ behaviors into five categories: raising head,bowing head,standing,lying on table,and turning head.At the same time,aiming at the problems of unbalanced target categories and one target size in the dataset,a data augmentation method is proposed that allocates images according to category ratio and scale-diverse combinations,which effectively improves the mean average precision of the model.2.A technique suitable for detecting the behaviors of students in the classroom is presented.This dissertation selects the YOLOv5 deep convolutional neural network as the detection network for students’ behaviors.By improving the feature fusion structure of the network,the low-level features are fused with the high-level features,so that the output features contain more detailed information.At the same time,in order to alleviate the high computational cost of the network model,the ordinary convolution module of the network is replaced with a ghost module,thereby reducing the parameters and computational cost of the network,and reducing the computing power requirements of the model on the deployment equipment.After feature fusion and model lightweight improvement,the mean average precision of students’ behaviors detection reaches 84.9%.3.According to the particularity of the class scene of this dissertation,a method is proposed to correlate students’ behaviors in continuous time by using the positional relationship and feature similarity of students between neighboring video frames.This method can accurately correlate the individual behaviors of small target students in continuous time in the intensive classroom scene,and the correlation accuracy reaches more than 90%.Through the scheme proposed in this dissertation,the state of students’ behaviors can be accurately detected and correlated in the intensive classroom scene.By statistics of behaviors changes of all students,estimating the degree of students’ concentration on study,forming effective concentration feedback,provide teachers with objective and accurate teaching assistance,and helping the construction of smart classrooms.