FC-YOLO：一种课堂行为检测的改进算法

FC-YOLO: An Improved Algorithm for Classroom Behavior Detection

摘要: 针对课堂场景中学生目标密集、行为差异细微、手机和书本等局部目标尺度较小以及复杂背景干扰等问题，提出一种融合前景感知与坐标先验的课堂学生行为检测算法。该方法以YOLO11n为基础模型，在特征融合阶段设计Foreground-Aware Scale Routing Concat模块，简称FASR-Concat，通过前景感知和尺度路由机制增强不同尺度特征的自适应融合能力；在检测头部分设计Coordinate Prior Detection Head模块，简称Coord-Detect，通过注入归一化二维坐标信息增强模型对固定视角课堂图像中目标空间位置的感知能力。实验在公开POCO课堂行为数据集上进行，结果表明，YOLO11n基线模型在验证集上的mAP50为0.812；加入FASR-Concat后，mAP50提升至0.835；采用FASR-Concat与Coord-Detect的完整改进模型后，mAP50达到0.850。与基线模型相比，最终模型mAP50提升3.8个百分点，mAP50:95提升2.6个百分点。类别结果显示，该方法对play_phone、look_book和turn_head等依赖局部区域或姿态细节的类别具有较明显提升。实验结果验证了所提方法在课堂学生行为检测任务中的有效性。

Abstract: To address the problems of dense student targets, subtle behavioral differences, small local objects such as mobile phones and books, and complex background interference in classroom scenes, this paper proposes a classroom student behavior detection algorithm integrating foreground awareness and coordinate priors. Based on YOLO11n, the proposed method designs a Foreground-Aware Scale Routing Concat module, abbreviated as FASR-Concat, in the feature fusion stage to enhance the adaptive fusion capability of multi-scale features through foreground awareness and scale routing. In the detection head, a Coordinate Prior Detection Head module, abbreviated as Coord-Detect, is introduced to inject normalized two-dimensional coordinate information, thereby improving the model’s perception of target spatial positions in fixed-view classroom images. Experiments were conducted on the public POCO classroom behavior dataset. The results show that the YOLO11n baseline achieves an mAP50 of 0.812 on the validation set. After introducing FASR-Concat, the mAP50 increases to 0.835. With the complete improved model using both FASR-Concat and Coord-Detect, the mAP50 reaches 0.850. Compared with the baseline model, the final model improves mAP50 by 3.8 percentage points and mAP50:95 by 2.6 percentage points. Class-wise results show that the proposed method achieves clear improvements for categories relying on local regions or pose details, such as play_phone, look_book, and turn_head. The experimental results verify the effectiveness of the proposed method for classroom student behavior detection.