Research And Application Of Quantitative Computational Methods For Teachers' Non-verbal Behaviors In Intelligent Classroom Environment

In the process of achieving the deep integration of information technology and education,teachers play a very important role.The continuous update of information technology has brought new teaching methods such as blended learning,collaborative learning,and personalized learning.The implementation of Reform of teaching methods has put forward new requirements for teachers'literacy.Teaching activities are a kind of cognition process that takes place in special situations.It is also a process of equal communication and interaction between teachers and students.There are both verbal and non-verbal communication.Due to the characteristics of teaching activities and the particularity of the teaching process,teachers are often the dominant aspect Teachers' teaching behaviors are expressive,infectious,and motivated,which will stimulate students' follow,focus and participation.Most of the current classroom teaching management information systems focus on in-depth research and analysis of students' learning behaviors,while neglecting the leading role of teachers.The research in this thesis is oriented to enhance the effectiveness of education and teaching and promote the professional development of teachers.The goal is to quantify the non-verbal behavior of classroom teaching.The focus is on the intelligent recognition of non-verbal behavior in classroom teaching.In teaching activities,non-verbal behaviors of teachers in the classroom can be used as auxiliary functions of non-intellectual factors to promote the cognitive process of learning and thus improve teaching effectiveness.At the same time,the study of non-verbal behaviors is also conducive to updating teachers' quality concepts,perfecting teaching communication skills,and improving teachers' ability.Therefore,it is an important category of teacher education research to advance the research on quantitative non-verbal behavior.At present,in the traditional teaching practice,the evaluation of teachers emphasises on and rewards?punishments and performance,focusing on results rather than processes.The study of teaching behaviors is generally described macroscopically,lacking micro-analysis,using empirical and qualitative research methods,less empirical,lacking quantification;analysis methods are mostly focused on the preparation of analytical scales;Observers manually code and record teaching behaviors in the classroom based on related scales,and then perform calculation analysis.There are shortcomings such as strong subjectivity,time-consuming and labor-intensive,and small sample size,which is not conducive to discovering universal teaching laws,nor can it guide practice and timely feedback.Nowadays,the development of information technology and the increasing popularity of intelligent teaching environments have brought opportunities to solve these problems.Generally,the research on human behavior is generally carried out by using video sensor devices to obtain video data.However,the smart classroom is an advanced and complex teaching environment,which not only places high requirements on image acquisition and data processing,but also poses higher challenges for the accuracy,robustness and real-time of algorithm recognition.In addition,the existing algorithms often fail to exert ideal performance due to the influence of complex scenarios in practical applications.The key factors affecting performance include:(1)complex scenes;(2)lighting changes;(3)human occlusion;(4)motion blur;(5)video image noise and many other factors;Therefore,in view of the above difficult issues,this paper is' the first to conduct research on behavior recognition methods that take into account robustness,accuracy,and real-time performance in the complex scenarios of smart classrooms,and developed a teacher non-verbal behavior measurement and evaluation system based on behavior recognition.The research work of this paper is mainly reflected in the following aspects:1.Propose a new method for robust human object segmentation based on image and video in complex scenes with multiple light sources and varying lighting.In a smart classroom environment,the algorithm can accurately segment the person's target in a multi-source lighting scenario where the electronic double whiteboard is used as one of the light sources of the individual light emitters.This paper proposes a lightweight and robust end-to-end fully convolutional neural network model;the data input section uses a thin-line read-in method that is compatible with three-channel grayscale images;the network architecture follows the full convolutional neural network's structure,the first half of the down-sampling part performs convolution operations to extract features,and the second half uses improved "content-aware" up-sampling to improve the accuracy of pixel-level inference.Loss calculation part uses weighted cross-entropy(Weighted-Cross-Entropy)function for error calculation.Using "transfer learning" technology,the Backbone model(FCN-8S)pre-trained on the high-resolution image dataset can retain the corresponding functions and parameters in the new model.The new model can predict accurately,quickly and in real time in the semantic segmentation of images in the natural environment.2.Propose segmentation of human targets based on thermal infrared images and videos in complex scenes to achieve live detection.For smart classroom environments,the electronic whiteboard,as one of the teaching content presentation tools,is prone to produce complex and changeable backgrounds.Aiming at the situation that the person image is used as the background to interfere with the real target,this paper proposes an efficient and robust end-to-end full convolutional neural network model to accurately and quickly segment the person target from the infrared heat-sensitive image.So as to avoid the target interference caused by the background.3.This paper proposes a new teaching gesture recognition method based on the fusion of multiple features.For complex scenarios in smart classrooms,a network framework model(An-Net)is proposed,which can accurately recognize human behaviors under light source changes and noise environments.First,the image segmentation algorithm extracts the edge features of human targets;then extracts the skeletal features of human targets;moreover performs feature selection under the guidance of attention mechanism based on motion features;finally,gestures are classified by the spatio-temporal RNN network model.In order to better perform model evaluation and analysis,this paper also establishes a set of teaching gesture behavior dataset in the smart classroom scenario.The experimental results show that our proposed method can objectively and accurately recognize teachers' teaching gestures in the classroom.4.This paper proposes a quantitative calculation and evaluation system for teachers' non-verbalbehavior based on the analysis of teachers' teaching behaviors.The system includes three functional modules:"data acquisition and storage","head momentum recognition","gesture behavior recognition","body distance intimacy behavior recognition" as the core and integrates the existing"teaching cloud platform interaction module".The system can quantitatively calculate and intelligently evaluate the non-verbal behaviors of teachers in the teaching process,in order to improve the teaching effect and promote the professional development of teachers.