Abstract: Aiming at the problem that student targets in classroom scenes are small, densely distributed, and easily obscured, resulting in low detection accuracy and poor recognition effect, an improved classroom behavior detection algorithm based on YOLO11 (you only look once version 11), named MFD-YOLO, is proposed. Through a series of innovative designs, this algorithm significantly improves the accuracy of classroom behavior detection and recognition effects. Firstly, a multi-dimensional feature flow network (MFFN) is designed to enhance the feature representation of small targets by combining the dimension-aware selective fusion module and the multi-dimensional feature diffusion mechanism, significantly improving detection accuracy. Secondly, the feature enhancement aggregation module (FEAM) is constructed in the backbone network, which optimizes the feature extraction process by integrating the information from different scale sensory fields and enhances the network??s enhancement and aggregation capability of multi-scale features, thus improving the detection of dense student groups. Finally, the traditional detection head is improved to a dynamic detection head (DyHead), effectively improving occluded students?? recognition and reducing misdetection and omission by enhancing the multi-scale perception ability. Experimental results show that, on the POCO dataset, MFD-YOLO improves mAP0.50 and mAP0.50:0.95 by 4.2 and 6.0 percentage points, respectively, compared with the base model YOLO11n, which significantly improves the detection accuracy and effectively reduces the false and missed detection rates. On the SCB-Dataset3 dataset, MFD-YOLO improves mAP0.50 and mAP0.50:0.95 by 3.4 and 4.4 percentage points, respectively, which further validates the applicability and robustness of the improved algorithm and proves its potential application in classroom behavior detection.

Key words: classroom behavior detection, high accuracy, YOLO11, multi-dimensional feature flow network (MFFN), feature enhancement aggregation module (FEAM)

摘要： 针对课堂场景中学生目标小、分布密集且易被遮挡，导致检测精度低、识别效果不佳的问题，提出了一种基于YOLO11改进的课堂行为检测算法MFD-YOLO。该算法通过一系列创新设计，显著提升了课堂行为检测的精度和识别效果。设计了多维度特征流动网络（MFFN），通过结合维度感知选择性融合模块和多维特征扩散机制，增强了小目标的特征表示能力，显著提高了检测精度。在主干网络中构建了特征增强聚合模块（FEAM），通过整合不同尺度感受野的信息来优化特征提取过程，增强了网络对多尺度特征的增强与聚合能力，从而提高了对密集学生群体的检测能力。将传统检测头改进为动态检测头（DyHead），通过增强多尺度感知能力，有效提升了对被遮挡学生的识别能力，减少了误检和漏检现象。实验结果表明，与基础模型YOLO11n相比，MFD-YOLO在POCO数据集上的mAP0.50和mAP0.50:0.95分别提高了4.2和6.0个百分点，显著提升了检测精度，并有效降低了误检和漏检率；在SCB-Dataset3数据集上，mAP0.50和mAP0.50:0.95分别提高了3.4和4.4个百分点，进一步验证了改进算法的适用性和鲁棒性，证明了其在课堂行为检测中的应用潜力。

关键词: 课堂行为检测, 高精度, YOLO11, 多维度特征流动网络（MFFN）, 特征增强聚合模块（FEAM）