Scholarlink

<p>Real-time recognition of student behaviors in intelligent classrooms presents significant challenges due to high scene complexity, frequent occlusion, and the need for deployment on resource-constrained edge devices. This paper introduces YOLOv11n-SFS, a lightweight and high-efficiency object detection framework designed for embedded educational monitoring systems. The proposed approach incorporates three key architectural enhancements: (1) StarNet, a parameter-efficient backbone utilizing element-wise multiplication and depthwise separable convolutions; (2) C3k2-Faster, an optimized neck module based on Partial Convolution for improved multi-scale feature fusion; and (3) SCLD (Shared Conv Light Detector), a compact detection head using Group Normalization and shared convolution to ensure stable performance under limited batch sizes. Experimental results on a curated student behavior dataset demonstrate that YOLOv11n-SFS achieves 95.68% mAP@0.5 with only 1.63&#xa0;M parameters and a model size of 3.4&#xa0;MB, while delivering 35.66 FPS on CPU inference. The model demonstrates strong deployment potential in practical classroom systems, offering a practical balance between accuracy, computational cost, and hardware efficiency for embedded AI applications in educational environments.</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

A lightweight YOLOv11n-based framework for student behavior detection in smart classroom environments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月7日

查看全文
  • Yanfei Li,
  • Qianqian Chen

摘要

Real-time recognition of student behaviors in intelligent classrooms presents significant challenges due to high scene complexity, frequent occlusion, and the need for deployment on resource-constrained edge devices. This paper introduces YOLOv11n-SFS, a lightweight and high-efficiency object detection framework designed for embedded educational monitoring systems. The proposed approach incorporates three key architectural enhancements: (1) StarNet, a parameter-efficient backbone utilizing element-wise multiplication and depthwise separable convolutions; (2) C3k2-Faster, an optimized neck module based on Partial Convolution for improved multi-scale feature fusion; and (3) SCLD (Shared Conv Light Detector), a compact detection head using Group Normalization and shared convolution to ensure stable performance under limited batch sizes. Experimental results on a curated student behavior dataset demonstrate that YOLOv11n-SFS achieves 95.68% mAP@0.5 with only 1.63 M parameters and a model size of 3.4 MB, while delivering 35.66 FPS on CPU inference. The model demonstrates strong deployment potential in practical classroom systems, offering a practical balance between accuracy, computational cost, and hardware efficiency for embedded AI applications in educational environments.