ABSTRACT To address the challenges of attention distraction and temporal jitter caused by complex backgrounds, human occlusion, and fine‐grained behavioral features in classroom behavior recognition, this paper proposes a recognition framework that integrates a Multi‐branch Spatiotemporal Attention Network (MSTA‐Net) with a Behavior State Kalman Filter (BSKF). At the perceptual level, the framework captures salient student behavior features through decoupled channel, spatial, and short‐term temporal attention branches. At the cognitive level, the recognition task is formulated as a state estimation problem in a high‐dimensional probabilistic space, where temporal smoothing is achieved by leveraging behavioral inertia. Experimental results demonstrate that the proposed method performs excellently on the SCB‐Dataset and real‐world video sequences, achieving an accuracy of 94.7% and a prediction speed of 33 FPS. Compared with pure deep learning‐based models, the proposed framework reduces the Action Category Switching (ACS) rate by 50%, significantly enhancing the robustness for long‐term behavior recognition.
Geng et al. (Tue,) studied this question.