A Behavior Recognition Network for Complex University Exam Environments Via Feature Enhancement and Attention Mechanism

Behavior recognition is a critical technology that enables computer systems to understand and interpret human activities, with wide-ranging applications in education, humancomputer interaction, medical rehabilitation, autonomous driving, and other domains. However, in examination scenarios, the complexity of the environment, the presence of numerous interference factors, and the scarcity of publicly available datasets significantly hinder the performance of existing mainstream behavior recognition methods. To address these challenges, this paper proposes a novel behavior recognition model for complex university exam environments. Specifically, to alleviate the degradation of behavior feature representation caused by increasing network depth, a Transformer-based global context information module is designed. This module effectively captures long-range dependencies, thereby enhancing the expressive power of behavior features. Moreover, to compensate for the limitations of relying solely on visual information in behavior feature extraction, a behavior pose information extraction module is introduced. This module supplements visual cues by extracting high-level semantic information from deep features. Given the inherent limitations of pose estimation, a behavior attention module is further constructed to enrich visual representations while simultaneously incorporating pose information. Finally, a classification module is designed to integrate global context, pose, and visual attention information, enabling accurate behavior classification. To validate the effectiveness of the proposed model, a university exam behavior dataset is constructed, and extensive experiments are conducted. Experimental results demonstrate that the proposed model performs robustly in complex examination environments, highlighting its strong potential for real-world application and practical deployment.