| In recent years,high speed development of artificial intelligence,in speech recognition,image processing,automatic translation,automatic driving,and other fields has made breakthrough progress,also has injected new vitality for the education reform,the State Council and the Ministry of Education but also have published many file points out that in the process of promoting education reform the important strategic status of the teaching evaluation.Traditional teaching evaluation methods mainly rely on teaching evaluation experts or observers to complete the evaluation of the classroom and teachers by listening to and evaluating the class.This kind of teaching evaluation method has high requirements on the quality and professional ability of teaching evaluation experts,and needs to consume a lot of energy.With the development of technology and times,in the stage of intelligent teaching evaluation,classroom behaviors of teachers or students directly related to classroom teaching effects can be described and quantified based on classroom teaching videos to assist teaching evaluation experts to carry out teaching evaluation work and promote the innovation of teaching evaluation methods to become a new exploration direction.Therefore,it is of great practical significance to study teachers’ classroom teaching behavior.In this thesis,classroom teaching videos of many primary and secondary schools in Beijing are used as research data,mainly from two aspects:In the first study,human skeleton features were extracted based on OpenPose and perspective transformation,and then classification models of teachers and students in teaching videos were constructed.The effects of different input features and combination of machine learning models on classification effects of teachers and students were studied.The findings: 1.In the combination of four input features(key point feature only,key point feature and location feature,key point feature and artificial feature,full feature)and three machine learning models(regularized Logistic,random forest,XGBoost),when the input feature is full feature,the performance of the three models is optimal.Compared with regularized Logistic and random forest,the number of error samples in XGBoost model decreased by 69% and 65%,and the effect improved significantly.2.In the view of the importance of each feature,left＿Y,LWrist,RWrist,FEATure＿1,FEATure＿2,FEATure＿3have the greatest influence on the performance of teacher-student classifier.In the second study,teachers’ behavior classification rules were formulated based on teachers’ position,speech and blackboard writing,and teachers’ teaching characteristics were classified based on HDclassif clustering method.Based on the principle of exhaustion and mutual exclusion,teachers’ behaviors can be divided into nine types: teaching while writing on the board,asking questions while writing on the board,writing on the board only,lecturing in lectern area,asking questions in lectern area,observing in lectern area,lectern area,asking questions in lectern area,observing in lectern area;Based on the behavior classification rules,the classroom teaching behaviors of all teachers were classified in seconds to obtain the frequency of different behaviors of teachers in each class.The HDclassif clustering method was used to classify teachers’ classroom behaviors into three types: blackboard interaction,classroom teaching and classroom interaction. |
