A classroom behavior recognition method based on Temporal Attention Mechanism and Fragment-based Mechanism

Abstract Classroom behavior detection aims to enhance classroom efficiency by collecting information on students' behavior and developing strategies for improved teaching methods. However, existing models suffer from low efficiency, subjectivity, and recognition accuracy. To address these issues, we propose a classroom behavior recognition model utilizing a dataset from monitoring devices, based on the Temporal Attention Mechanism (TAM) and Fragment-based mechanism of a 3D convolutional neural network. Our TAM applies attention distribution in the time dimension, combined with 3D CAM, as the Channel and temporal attention module (CTAM), resulting in a 10\% accuracy improvement. We integrate 3D-CAM and 3D-SAM with TAM, validating the portability and universality of our time attention mechanism. Furthermore, we introduce the Fragment-based mechanism, synthesizing feature outputs of different time scales for improved classification. By incorporating this mechanism, the model achieves a 30\% accuracy improvement on the classroom behavior dataset. Evaluation on a public dataset demonstrates significant improvements compared to the original model.