A domain adaptive and continual learning method for skeleton behavior recognition in classroom environment

Student behavior in the classroom is an important part of the analysis of classroom process and effectiveness. Due to the different physiological and psychological states of the learners in the classroom environment, it is very common that there are differences between the same types of actions, and the diversity of the same types of actions will change over time. To solve the online recognition problem of the same kind of actions with high diversity in the classroom environment, a domain adaptive and continual learning method for skeleton-based behavior recognition is proposed, which achieves the online transfer ability of models between the same kind of actions. Experiments show that this method has better transferability and continuous learning performance on public datasets Florence, Kinetics and real classroom datasets, and the accuracy is improved by 2.3%,1.3% and 7.29% on the three datasets respectively, compared with the current state-of-the-art algorithms.