基于YOLOv11与DeepSeek的课堂专注度智能分析系统设计

Design of an Intelligent Classroom Attentiveness Analysis System Based on YOLOv11 and DeepSeek

摘要: 课堂专注度评估长期以来都是高校教学质量监控中比较薄弱的一个环节，在广西民族地区的高校，这一情况表现得更加明显。人工巡课能够覆盖到的班级比较有限，评价结果主要依靠督导个人的经验，缺少量化的标准；教室里的监控设备每天都在录制视频，但这些视频数据长期没有得到有效的利用；即使引入了行为检测算法，输出的内容也大多停留在行为计数这一层面，与教师真正需要的教学诊断之间还存在一定的差距。针对上述情况，本文设计了一套课堂专注度智能分析系统，检测部分采用YOLOv11模型，诊断部分接入DeepSeek大模型。系统按照视觉感知、量化评估、语义诊断三个层次进行组织：感知层以YOLOv11n作为基线模型，在C3K2特征输出之后加入通道–空间注意力模块，同时在颈部网络增加一条浅层特征通向检测层的融合支路，用来减少教室人员密集遮挡和后排小目标所带来的漏检；评估层采用课堂参与度指数(EPI)对专注度进行量化，行为权重通过层次分析法并结合广西民族院校课堂的实际情况进行标定，输出0到100分的时序评分；诊断层通过三层Prompt调用DeepSeek模型，把时序行为统计转换为教学改进报告。从实验结果来看，基线模型的参数量为2.62M，mAP@0.5为0.842，单帧推理时间约为18.3 ms，五类行为中低头类别的检测效果最好，mAP@0.5达到0.952。系统在校内多个班级进行了试用，EPI评分与教师主观评价的相关系数为0.87。系统可以运行在普通办公电脑上，直接使用教室已有的监控设备，推理过程在本地完成，不需要额外采购硬件，比较适合经费有限的民族地区院校使用。

Abstract: Classroom attentiveness evaluation has long been a relatively weak part of teaching quality monitoring in universities, and this problem is more noticeable at universities in the ethnic areas of Guangxi. Manual patrols can cover only a limited number of classes, and the evaluation mainly depends on the personal experience of supervisors, with no quantitative standard; surveillance cameras in classrooms record videos every day, but these video data have not been effectively used for a long time; even when behavior detection algorithms are applied, their output mostly stays at the level of behavior counting, which still falls short of the teaching diagnosis that teachers need. In view of the above situation, this paper designs an intelligent classroom attentiveness analysis system, in which YOLOv11 is used for behavior detection and the DeepSeek large language model is integrated for diagnosis. The system is organized into three layers, namely visual perception, quantitative evaluation, and semantic diagnosis. The perception layer takes YOLOv11n as the baseline model; a channel-spatial attention module is added after the C3K2 feature output, and a fusion branch from shallow features to the detection layer is added in the neck network, so as to reduce missed detections caused by dense occlusion and small targets in back rows. The evaluation layer uses the Engagement Participation Index (EPI) to quantify attentiveness; the behavior weights are calibrated through the analytic hierarchy process based on the actual conditions of classrooms in Guangxi ethnic universities, and time-series scores from 0 to 100 are produced. The diagnosis layer invokes the DeepSeek model through three-layer prompts, and converts time-series behavior statistics into teaching improvement reports. The experimental results show that the baseline model has 2.62M parameters, an mAP@0.5 of 0.842, and a single-frame inference time of about 18.3 ms; among the five behavior categories, the head-down category achieves the best detection result, with an mAP@0.5 of 0.952. The system has been tried in several classes on campus, and the correlation coefficient between EPI scores and teachers’ subjective evaluation is 0.87. The system can run on ordinary office computers and directly use the existing surveillance equipment in classrooms; the inference is completed locally, and no additional hardware purchase is required, which makes it relatively suitable for ethnic-area universities with limited budgets.