Scholarlink

<p>Multimodal perception of classroom engagement via behavioral analytics presents a formidable challenge in intelligent education systems. Conventional educational computing frameworks grapple with three core limitations: (i) insufficient discriminative power for behavior patterns with high inter-class similarity, (ii) scarcity of annotated data compounded by cross-cultural labeling inconsistencies, and (iii) poor interpretability of classroom engagement models. We propose a unified multi-task architecture that jointly performs facial landmark localization (29-point regression) and action categorization (7-class recognition). Our methodology introduces three distinct contributions: (i) Switchable Atrous Convolution module that enriches spatial-semantic feature extraction; (ii) Dynamic Gradient Normalization scheme for balancing asymmetric task objectives; and (iii) Non-Parametric Attention Gate to refine spatio-temporal feature fusion. Validated on the Classroom Behavioral Dataset (CBD-7: 41,700 annotated samples spanning 13.04&#xa0;h), our model attains 89.63% mAP0.5 while sustaining 28.7 FPS inference on resource-constrained edge hardware (Jetson Xavier NX). Ablation analyses substantiate the incremental value of each component, notably a 5.2% accuracy gain for late-class period detection attributed to the temporal accumulation mechanism. These results facilitate pragmatic deployment of AI-driven classroom observation tools without sacrificing pedagogical interpretability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal analysis of classroom engagement: a lightweight multi-task learning framework

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Qiongjie Kou,
  • Quanyou Zhang,
  • Yaohui Li,
  • Rongrong Kou,
  • Fangtao Qin,
  • Delan Xiong,
  • Jinling Shi

摘要

Multimodal perception of classroom engagement via behavioral analytics presents a formidable challenge in intelligent education systems. Conventional educational computing frameworks grapple with three core limitations: (i) insufficient discriminative power for behavior patterns with high inter-class similarity, (ii) scarcity of annotated data compounded by cross-cultural labeling inconsistencies, and (iii) poor interpretability of classroom engagement models. We propose a unified multi-task architecture that jointly performs facial landmark localization (29-point regression) and action categorization (7-class recognition). Our methodology introduces three distinct contributions: (i) Switchable Atrous Convolution module that enriches spatial-semantic feature extraction; (ii) Dynamic Gradient Normalization scheme for balancing asymmetric task objectives; and (iii) Non-Parametric Attention Gate to refine spatio-temporal feature fusion. Validated on the Classroom Behavioral Dataset (CBD-7: 41,700 annotated samples spanning 13.04 h), our model attains 89.63% mAP0.5 while sustaining 28.7 FPS inference on resource-constrained edge hardware (Jetson Xavier NX). Ablation analyses substantiate the incremental value of each component, notably a 5.2% accuracy gain for late-class period detection attributed to the temporal accumulation mechanism. These results facilitate pragmatic deployment of AI-driven classroom observation tools without sacrificing pedagogical interpretability.