Student Classroom Behavior Recognition And Analysis Based On Deep Learnin

In order to better understand the overall learning status of students,evaluate the effectiveness of classroom teaching,and promote high quality teaching in universities,it is very important to analyze students’ classroom behavior.Existing research on student behavior recognition only focuses on students’ own recognition,while not enough attention is paid to students’ interactions with surrounding objects.To address this phenomenon,this thesis research adopts human-object interaction detection technology to identify and analyze students’ classroom behaviors from two major directions based on multi-stream branching and based on graph convolution,and the main work is as follows:(1)In order to detect the desired targets in the classroom more accurately,this thesis proposes a multi-target detection method in the classroom based on improved YOLOv5 s.Firstly,for the problem that the scale of targets such as cell phones and pens in classroom scenes is small and the effective features that can be extracted are small,this thesis takes measures to optimize the network structure.Secondly,in view of the interference of irrelevant information such as classroom background and different dress of students in the real classroom environment,it is difficult for the network to extract effective features,so the triplet attention mechanism is introduced to enhance the ability of the network to extract features.Finally,experiments were carried out on self-made datasets and public datasets.The experimental results show that the improved network m AP values are increased by 4.5% and 3.2% respectively,which verifies the effectiveness of the improved work.(2)This thesis proposes a research method of classroom behavior recognition based on multi-stream branch.In view of the large interaction distance between some students and surrounding objects in the classroom,and the small range of attention of the network,resulting in insufficient extraction of context features,this thesis designs a context-based residual module by means of the method of large kernel separable convolution combined with small kernel residual structure,to capture richer context features with a larger field of sensitivity.Finally,the experimental verification was carried out on the self-made student behavior datasets and the public datasets,and the experimental results showed that the improved network m AP values increased by 5.74%and 1.66%,respectively,indicating that the improved network can achieve better recognition effect in the classification task of students’ classroom behavior.(3)Based on the research based on multi-stream branches,this thesis further constructs the interaction relationship graph between students and items,and proposes and implements a student classroom behavior recognition network based on graph convolution.Firstly,this thesis adopts the strategy of optimizing the item nodes in the graph convolution module to reduce the interference of redundant connections to model the effective relationship between students and items.Secondly,for the feature that no interactive items are required for listening behavior,this thesis designs and implements a student behavior recognition network without interactive items.Finally,the recognition accuracy of the optimized network is experimentally verified to be improved by 3.85% and 2.49% on the self-made datasets and the public datasets,respectively,which demonstrates the feasibility of the improved scheme.