近些年来，基于深度学习的计算机视觉处理技术飞速发展，其研究成果已经开始应用于人们的日常生活。行为识别作为计算机视觉处理技术的具体应用，有着现实的应用前景。随着国家对教育信息化、智能化的日益重视，将人工智能等先进技术引入传统教育、提高传统教育的效率已经成为了未来教育发展的必然趋势。本文旨在将行为识别技术引入教学录播系统，实现在课堂中对学生的行为进行实时监测，并对起立发言的学生进行定位，促进教学录播自动化。

  本文主要研究了从课堂教学过程中的学生图像中识别学生坐和站立两种行为的方法。对于这种目标数目多、行为数目少且都非常简单的行为识别应用场景，本文为了提高算法运行效率，并没有采用目前较为流行的基于深度学习的行为识别算法，而是采用目标检测、人体骨骼检测、最后进行预测分类的融合框架来完成发言学生的定位。实验结果表明，系统在验证集上的平均准确率达到90.93%，基本可以满足录播系统完成对上述两个行为进行识别的需求。

In recent years, computer vision processing technology based on deep learning has developed rapidly, and its research results have begun to be applied to people's daily life. Behavior recognition, as a specific application of computer vision processing technology, has realistic application prospects. With the increasing emphasis on informationization and intelligence in education, the introduction of artificial intelligence and other advanced technologies into traditional education and the improvement of the efficiency of traditional education have become the inevitable trend of future education development. The purpose of this paper is to introduce behavior recognition technology into the teaching recording system to realize real-time monitoring of students' behavior in the classroom, and to locate students who rise to speak, so as to promote the automation of teaching recording.

    This paper focuses on the method of recognizing both sitting and standing student behaviors from the images of students in the classroom teaching process. For this behavior recognition application scenario where the number of targets is large and the number of behaviors is small and both are very simple, this paper does not use the currently popular deep learning-based behavior recognition algorithm in order to improve the algorithm operation efficiency, but adopts a fusion framework of target detection, human bone ship detection, and finally predictive classification to complete the localization of speaking students. The experimental results show that the system achieves an average accuracy of 85.18% on the training set and 80.52% on the validation set. It can basically meet the needs of the recording system to complete the recognition of the above two behaviors.