Teaching Behavior Recognition Based On Human Pose Estimation

Artificial intelligence has achieved great success in various fields such as speech recognition,image processing,and autonomous driving.Its thriving development trend has injected new vitality into the field of education.In order to promote the development of the educational reform process,the State Council and the Ministry of Education have successively issued multiple documents,emphasizing the important strategic position of conducting teaching evaluations scientifically.Traditional teaching evaluation methods mainly rely on evaluation experts or observers to evaluate the classroom and teachers by listening to and evaluating classes.This evaluation method not only requires high qualifications and professional abilities from the evaluation experts,but also leads to unnecessary waste of human and material resources.Therefore,how to conduct teaching evaluations objectively and scientifically is one of the urgent problems to be solved in the current field of education.Artificial intelligence technology can quantify and analyze the classroom behaviors of teachers or students that are associated with classroom teaching effectiveness based on classroom teaching videos,providing new ideas for industry experts to conduct evaluation work in an orderly manner and promote the innovation of the teaching evaluation system.Therefore,utilizing artificial intelligence technology for intelligent evaluation of teaching is of great significance.The data source of this article is derived from online videos of secondary school classroom teaching,and the research mainly covers the following aspects.First,based on the existing offline classroom teaching action coding system,we constructed a classroom teacher’s body movement analysis and coding system suitable for online video teaching.This paper analyzed the current research on teacher teaching behavior coding systems and constructed a teacher’s body movement coding system suitable for existing classroom environments.The teacher’s body movements are divided into four categories: gestures,standing teaching,walking teaching,and board writing.Second,we constructed a teacher’s body movement dataset.This paper collected secondary school classroom teaching videos from the internet,selected teaching videos that included the entire body of the teacher,and conducted screening,classification,and labeling of the videos to build the required dataset for the current research.The dataset consists of 380 video segments.Third,we performed teacher’s teaching pose recognition based on pose estimation on the dataset video segments.First,this paper used Faster R-CNN and HRNet networks for 2D pose estimation of teacher’s posture.Then,based on the pose estimation results,this paper used three methods for teacher’s posture recognition:Firstly,perform posture recognition using the ST-GCN with the 2D pose estimation results directly.Secondly,to enhance the accuracy of posture recognition,the paper generated 3D heatmaps from the 2D pose estimation results and used the PoseC3D model for posture recognition.Thirdly,considering the limited accuracy of PoseC3D for posture recognition and the characteristics of teacher’s posture,the paper introduced an attention mechanism and proposed the PoseC3D-ECA model.Finally,a comparison was conducted between the proposed pose estimation and posture recognition results.The experimental results demonstrated that the average top-1accuracy of teacher’s teaching posture recognition using the attention-based PoseC3D-ECA model proposed in this paper was the highest,providing more effective data support for the automated analysis of educational teaching and the improvement of new evaluation methods.