Video Action Recognition of Students’ Behaviors in Classrooms

Abstract

Computer vision is a very important field in artificial intelligence, and video understanding series tasks are a popular direction among them. Cameras can be seen everywhere in our lives, such as the mobile phone's own lens, the surveillance camera in the classroom, etc. The detection of behavior in videos has become a research hotspot in recent years, video can be seen as a series of stacked frames, but unlike the picture, the video adds a time dimension, and there are factors such as brightness changes and lens shake, which increase the difficulty of video behavior detection. Compared with the traditional video understanding technology, deep learning-based methods can autonomously learn the features in the video, reducing the manual setting steps and improving the performance of the model to a certain extent. In recent years, with the rapid development of China's economy and technology, the level of intelligence in classroom teaching has been continuously improved. Concepts such as smart classrooms and smart campuses are gradually implemented, and devices such as cameras and teaching computers have become the basic configuration of each classroom. Excellent teaching quality not only requires the support of external hardware equipment, but more importantly, the students' attention status. Due to the large amount of students in university classroom, they easily divert attention to other things. In the limited teaching time, teachers can't pay attention to every student at all times, so a certain computer vision algorithm is needed to help teachers understand the students' class status. The scene of student behavior is in the classroom. At present, there is little research on the behavior of such scenes at home and abroad, and a unified data set has not yet been formed. This paper selects the real classroom video of the university classroom as the data source, and makes a representative and challenging classroom behavior data set through data annotation and sorting. Based on the predecessors, this paper improves the existing time series behavior detection network. In this paper, two models have been improved, namely boundary sensitive network and C3D network, making it suitable for detecting behavior in the classroom environment. In this paper, the knowledge related to deep learning is first explained, then the network structure of the two algorithms is introduced in detail, and the details of the optimization algorithm and tuning strategy used for model training are introduced. Finally, experiments have verified the feasibility of the improved algorithm in this paper.