基于YOLOv11与空间特征增强的学生课堂行为检测方法研究
Research on Student Classroom Behavior Detection Method Based on YOLOv11 with Spatial Feature Enhancement

针对课堂场景中学生行为目标尺度较小、姿态变化幅度有限及不同行为在外观特征上有较高相似性等问题，本文构建了一种新的YOLOv11-RMS模型。该模型以YOLOv11为基础，在主干网络中引入基于Restormer的特征提取模块，通过建模长程依赖关系增强特征的全局建模能力；其次，在主干网络末端融合多级通道注意力机制(MLCA)，以强化关键语义特征并抑制冗余信息；最后，在检测头阶段引入基于空间自适应特征调制的SAFMP模块，对上采样后的浅层特征进行重构与增强，提升模型在复杂课堂环境下检测稳定性。实验结果表明，YOLOv11-RMS模型在SCB数据集和自建学生课堂行为数据集CLASS上的mAP50分别达到72.6%和84.9%，较主流模型取得了明显提升。
To address the challenges of small target scales, limited posture variations, and high visual similarity among different student behaviors in classroom scenarios, a novel YOLOv11-RMS model is proposed. Based on the YOLOv11 framework, a Restormer-based feature extraction module is incorporated into the backbone to enhance global feature modeling through long-range dependency learning. Subsequently, a multi-level channel attention mechanism (MLCA) is integrated at the end of the backbone to strengthen discriminative semantic features while suppressing redundant information. Furthermore, a spatially adaptive feature modulation module (SAFMP) is introduced in the detection head to reconstruct and enhance upsampled shallow features, thereby improving detection robustness in complex classroom environments. Experimental results demonstrate that the proposed YOLOv11-RMS model achieves mAP50 scores of 72.6% and 84.9% on the SCB dataset and the self-collected classroom behavior dataset (CLASS), respectively, outperforming existing mainstream methods.