Research On Dynamic Identification Technology Of Student Behavior Based On Deep Learning

Education has always been the focus of national and social attention.In recent years,with the rapid development of artificial intelligence,the country has put forward "AI +education",pointing out that evaluation tools should be innovated and the whole process evaluation method of students’ learning should be explored.In the traditional classroom teaching,students ’classroom behavior can reflect the teacher’s teaching quality and students’ listening status from the side.Therefore,in the research related to intelligent education,students’ classroom behavior identification has always been the key content of the research.At present,the methods of students ’classroom behavior identification have problems such as low model identification accuracy,poor robustness,insufficient extraction of effective features of students’ behavior,and failure to meet the real-time detection of students’ classroom behavior.For these problems,the main research contents of this paper are as follows:(1)The students in class data,combined with the students classroom action and learning status,the student behavior is divided into up listening to lectures,take notes,reading,sleep,head phone,whispering,hands,using python third-party tools labelimg to mark students classroom behavior,make students’ classroom behavior data set.(2)According to the problem that the YOLOV5 algorithm does not comprehensively extract the effective features of students’ classroom behavior,the YOLOV5 algorithm based on attention mechanism is proposed.Three attention mechanisms,namely,SE(Squeeze-and-Excitation),CBAM(Convolutional Block Attention Module),and CA(Coordinate attention),were integrated into the YOLOV5.Compared based on SE,CBM,CA three attention mechanism YOLOV5 model in students’ classroom behavior data set performance,YOLOV5-CA model m AP value is the highest,2% compared with YOLOV5,the model in considering the characteristics of the channel information into the location information,solves the traditional model on the student classroom behavior recognition task feature extraction,poor robustness.(3)Aiming at the problem of YOLOV5-CA’s insufficient effect of effective feature fusion for students’ classroom behavior and the inability to distinguish the importance of feature information,the YOLOV5-CA-BIFPN model based on bidirectional weighted feature fusion and attention mechanism is proposed.The analysis of different feature fusion networks,replace PANet network structure with BIFPN network structure,and effectively enhance the ability of Neck to fuse important feature information.Compared with the detection effect of several classical deep learning models and the improvement model in the self-built student classroom behavior data set,the YOLOV5-CA-BIFPN model with bi-directional weighted feature fusion and attention mechanism had the highest prediction accuracy,and its m AP value increased by 3%.Further testing of the model using the VOC data set verified that YOLOV5-CA-BIFPN is expandability.This paper studies the dynamic identification of students ’classroom behavior based on deep learning,solves the problem of insufficient extraction and fusion of students’ classroom behavior features by classical algorithm,and improves the recognition accuracy of the algorithm when the algorithm meets the real-time detection.