Classroom Action Recognition Based on Graph Convolutional Neural Networks and Contrast Learning
Yuting Zhou, Shuang Peng, Xinxu Li, Qiancheng Cao, Li Tao · 2024
Student action recognition in the classroom has received extensive attention in recent years. With the remarkable development of deep learning, many models applied to classroom data have emerged. Among them, the method based on human skeleton recognition overcomes the interference of traditional image and video action recognition methods in complex backgrounds. However, human skeleton also has some problems in the field of action recognition, such as the loss of interaction information, and the skeleton graphs of different action categories are more similar. In this study, we propose an action recognition model based on contrastive learning and graph convolutional neural network to recognize student actions. By comparing the skeleton graphs and correcting misclassified samples, the model's feature information is strengthened, and the ability of the model to recognize actions is improved. For evaluation, we create a student classroom action dataset, where the data are represented as sequences of skeletons extracted from videos. Experiments show that the proposed method has achieved significant improvement in student action recognition, with an accuracy of 95.95%, providing a basis for the classroom action recognition system and classroom learning style evaluation.