摘要
实时、准确地对学生课堂行为进行检测存在许多挑战,教室场景中存在的大量遮挡和教学图像中后排的小目标是阻碍检测精度提高的重要因素。同时,教室中性能受限的摄像头限制了可部署目标检测模型的计算量和参数量。针对以上问题,提出一种基于改进YOLOv8的学生课堂行为检测模型TAR-YOLO。首先,提出一个结构新颖的任务对齐动态检测头(TAPH),其中共享卷积和任务对齐结构的设计增强了模型识别小目标的能力。其次,在特征融合部分引入层次化特征聚合的部分卷积(MSFAP)来改造YOLOv8特征提取模块,并使用重参数化的空洞残差卷积(R-DWR)来增强模型的多尺度特征能力,有效降低了模型的计算量和参数量。最后,构建一个用于检测7种学生课堂行为和3种教师课堂行为的数据集STCB,包含4242张图像和74571个标注。在该数据集上,TAR-YOLO模型的平均精度(mAP50)达到了79.5%,同时参数量和计算量只有7.0 G和1.8 M。为进一步验证模型在其他场景中的小目标检测能力,在PASCAL VOC数据集上进行实验,验证了该方法的泛化性能。
Real-time and accurate detection of student behavior in classrooms faces many challenges.Significant occlusion in classroom scenes and small targets in the back rows of teaching images are major obstacles to improving detection accuracy.Additionally,resource-constrained cameras in classrooms limit the computational cost and parameter size of deployable object detection models.To address these issues,we propose a student classroom behavior detection model based on an improved YOLOv8,named TAR-YOLO.Firstly,we introduce a novel Task Aligned Prediction Head(TAPH),where the use of shared convolution and task-alignment design enhances the model’s ability to detect small objects.Secondly,we integrate Multi-Scale Feature Aggregation Partial Convolution(MSFAP)into the feature fusion stage to enhance the YOLOv8 feature extraction module,and adopt a Re-parameterized Dilation-Wise Residual Convolution(R-DWR)to improve the model’s multi-scale feature capability,effectively reducing both computation and parameter load.Finally,we construct a dataset named STCB,which includes 4242 images and 74571 annotations for detecting 7 types of student behaviors and 3 types of teacher behaviors.On this dataset,the proposed TAR-YOLO model achieves a detection accuracy of 79.5%mAP,with only 7.0G FLOPs and 1.8M parameters.To further validate the model’s performance in small object detection under different scenarios,we conduct experiments on the PASCAL VOC dataset,demonstrating the generalization capability of our method.
作者
李金龙
郑霖
王浩宇
季伟东
LI Jinlong;ZHENG Lin;WANG Haoyu;JI Weidong(College of Computer Science and Information Engineering,Harbin Normal University;Intelligent Laboratory for Teaching and Development of Future Teachers,Harbin Normal University,Harbin 150025,China)
出处
《软件导刊》
2025年第9期189-198,共10页
Software Guide
基金
黑龙江省自然科学基金项目(PL2024F007)。
关键词
目标检测
学生课堂行为识别
轻量化模型
深度学习
object detection
student classroom behavior recognition
lightweight model
deep learning
