In recent years, analyzing teacher behavior in educational videos has emerged as a significant area of research. Conventional approaches depend heavily on manual observation, making them inefficient and labor-intensive when dealing with large-scale video data. Firstly, current behavior recognition work is difficult to apply to teacher behavior due to the complexity of extracting behaviors in teaching scenarios, this paper proposes a teacher region extraction (TRE) algorithm to identify teachers in videos, reducing interference from redundant information. Secondly, considering different visual rhythms and input scales, we propose a 3D-CNN based teacher behavior recognition network (3D-TBR). This network integrates an improved feature-level temporal pyramid module (ITPM) to model visual rhythm, with a bottom-up augment path to better utilize low-level information and enhance the feature pyramid. Additionally, to further improve performance and efficiency of the model, we incorporate the efficient channel attention network (ECA-Net) to better focus on important features. Our comparative and ablation studies demonstrate that the proposed method significantly enhance the accuracy of teacher behavior recognition.