Research On Teacher Behavior Recognition Based On Video

The teacher professional development depends on teaching instruction.Lack of classroom instructional data to provide effective reflective material cannot support professional development.However,it is difficult to obtain regular and accurate data continuously in teaching practice,and it is also difficult to provide long-term personalized feedback data for teachers manually.Therefore,it is a valuable subject about the use of artificial intelligence technology to provide intelligent teaching support and service for the teacher professional development.It can help teachers innovate teaching methods and implement personalized teaching.From the perspective of teacher professional development,this paper reviews the relevant literature and finds that using classroom video to improve teaching is a popular method to promote teachers’professional development,but there is a problem of time-consuming and distracting.Therefore,it is an urgent problem to study the use of artificial intelligence technology to support teachers to gain information from classroom video quickly.Although some tools supporting classroom video analysis have been developed,most of them are semi-artificial,which need to be viewed and labeled manually,and then automatically analyzed by the computer.However,the research on more intelligent teaching analysis algorithms is not mature.This paper focuses on the teachers’behavior to study the algorithm of teacher behavior recognition based on human behavior recognition technology,in order to reduce the workload of teachers to gain information from classroom video and provide long-term personalized feedback information for teachers.The main work of this paper is as follows.(1)The paper reviews the literature about analysis of classroom instructional behaviors and formulates a plan for teacher behavior recognition.According to it,we collect and construct the dataset of teachers’ behavior.There are six categories of teachers’behaviors:blackboard,instruction,question,display,description,and non-gesture behavior.4611 video clips of 50 teachers are collected.(2)The algorithm of teacher behavior recognition is studied,which combines the bone features with the spatial-temporal features.In view of the complexity of the real classroom scene,this paper introduces the skeleton feature which is not sensitive to the variation of illumination or background in the space-time characteristics,so as to improve the accuracy of teacher behavior recognition.First,the dense trajectory algorithm is used to detect the temporal and spatial features,and different descriptors are used to explore the effect of teacher behavior recognition.Then,according to the characteristics of the real video in class,the body posture estimation algorithm is adopted to extract the skeleton information.The skeleton data is processed to form three skeletal features,and then explore the effect of teacher behavior recognition through the experiments.Finally,the fusion experiment is carried out to explore the recognition effect of early-fusion and late-fusion.The experimental results show that the method which fuses the spatial-temporal features with skeleton features is significantly better than that of the single feature method in the recognition of teachers’ behavior.(3)The method of teacher behavior recognition based on convolutional neural network which combines attention mechanism is explored.It is easy to be interfered with other behaviors to find the information of teachers’ key behaviors in constant video clips,so we need to reduce the weight of unimportant behavior features to recognize teachers’behaviors better.In this paper,a channel attention module is proposed which integrates spatial-temporal information.It is added into the 3D convolutional decomposition network to improve the accuracy of teacher behavior recognition.In the experiment of the teacher behavior dataset,this paper adopts online data enhancement to prevent the overfitting caused by a small dataset.The experimental results show that the accuracy of this method is higher than that of several classical methods.In the comparison of several attention modules,our attention module has achieved better results in teacher behavior recognition.