Abstract—As multimedia technology has advanced, surveillance footage has become a vital tool for monitoring student behavior in classrooms, supporting improvements in instructional methods, learning outcomes, and overall discipline. The learning environment might be adversely affected by abnormal activities, including napping during class and using cell phones. Although many researchers have worked on abnormal behavior detection, improvements are still needed to accurately understand how student actions change over time and to detect when the behaviors are subtle or depend on the classroom context. To address this challenge, we propose a Transformer-inspired Attention Model (TAM) designed to classify short classroom videos as normal or abnormal based on student behavior. A novel Context-Aware Attention and Multiscale Detection (CAMD) module was added to extract rich spatial and temporal data at different scales. These frame-level data are subsequently combined into a global temporal representation using a Context-Aware Token (CAT) module, guaranteeing an effective summarization of the full video. To enhance feature integration, the Context-Aware Feature Integration (CAFI) module fuses both local and global contextual insights. Prior to categorization, essential frames were dynamically assigned importance through temporal attention pooling. With a mean Average Precision (mAP) improvement of nearly 4%, the model detection accuracy outperformed current models when tested on classroom videos. These findings show that the modules in the proposed model are effective in detecting classroom conduct from students.

 

Keywords—dual-attention mechanism, student behavior detection, shot-boundary, keyframes, multi-scale fusion, context-aware feature integration

Cite: Vinothina V, Augustine George, Jasmine Gnanadurai, and Vennila J, "A Dual-attention-based Approach for Student Abnormal Behavior Detection in Classroom Videos," Journal of Image and Graphics, Vol. 14, No. 4, pp. 645-654, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).