Concurrency and Computation: Practice and Experience Теоретическая информатика

Enhancing Classroom Engagement Analysis: A Fine‐Grained Behavior Detection Approach With Heterogeneous Adaptive Modeling

Mingyang Liu, Meng Guo, Lisen Wang, Mingxing Yu
Wiley · 2026-11-01 · DOI 10.1002/cpe.71031

Аннотация

ABSTRACT Intelligent classroom behavior detection is crucial for large‐scale, objective analysis of teaching quality and student engagement, yet occlusions, subtle and overlapping actions, viewpoint variations, and cluttered layouts make robust recognition highly challenging. In this paper, we propose Heterogeneous Point Allocation and Adaptive Sampling Mamba Network (HA2M‐Net) for fine‐grained intelligent classroom behavior detection. HA2M‐Net adopts an adaptive encoder–decoder architecture built around three core modules. First, an Adaptive Feature Dimension Estimation (AFDE) module predicts per‐location kernel height and width, providing a continuous estimate of the most informative receptive field under heterogeneous classroom scenes. To convert these continuous estimates into efficient discrete sampling patterns, a Heterogeneous Point Allocation Module (HPAM) maps the predicted kernel dimensions to an odd‐sized sampling grid and adaptively selects the image‐level sampling density, enabling compact spatial coverage of local body parts and classroom interactions. Next, an Adaptive Sampling Map Mamba (ASMM) module rescales a canonical grid into a learned sampling map, captures long‐range spatial dependencies across the entire classroom with favorable computational efficiency, and forms a dense sampling map for subsequent recognition. The experimental results show that HA2M‐Net achieves mean Average Precision (mAP) scores of 92.20% and 93.79% on the SCB‐S and SCB‐U datasets, respectively, with A‐OSE values of 282 and 163.

Читать у издателя →