基于RT-DETR-ASF的学生科学探究实验行为检测研究  

Research on the RT-DETR-ASF Based Detection of Students’Experimental Behavior in Scientific Inquiry

作　　者：杨帆 詹泽慧[1] Yang Fan;Zhan Zehui(School of Information Technology in Education,South China Normal University,Guangzhou,Guangdong,China 510631;College of Information Science and Technology,Jinan University,Guangzhou,Guangdong,China 510632)

机构地区：[1]华南师范大学教育信息技术学院,广东广州510631 [2]暨南大学信息科学技术学院,广东广州510632

出　　处：《数字教育》2024年第5期14-23,共10页Digital Education

基　　金：国家自然科学基金面上项目“基于事理图谱的计算思维智能导训模型及可解释性研究”(62277018);教育部人文社科基金项目“基于C-STEAM的粤港澳大湾区教育协同创新机制研究”(22YJC880106);华南师范大学哲学社会科学重大培育项目“面向创新能力培养的跨学科组合策略与应用效果研究”(ZDPY2208)。

摘　　要：深度学习方法在学生科学实验的自动检测和评估方面具有提高效率的潜力。为了解决学生科学实验数据集的缺乏和低准确率的问题,该文提出了一种基于实时的注意力尺度序列融合的目标检测变换器RT-DETR-ASF的学生科学实验检测方法。首先,该文构建了学生科学实验数据集,包括417个视频,18308张视频帧和20331个标注,主要关注5种行为:称重、测高度、丢球、测大小和记录。为提高检测准确率,引入了一种注意力尺度序列融合模块。为解决边界数据问题,提出了行为边界指数,用于识别数据集中的边界样本。为了解决数据不平衡问题,进行了过采样与视频帧扩展的操作。使用科学实验检测模型对数据集进行检测,实验结果表明:行为分类检测的平均准确率达到了71.1%。这证明了该模型的有效性。学生科学实验数据集与RT-DETR-ASF为未来的学生科学实验分析提供了先验基础,有望推动该领域的进一步发展。Deep learning methods have the potential to improve efficiency in automatic detection and evaluation of student science experiments.In order to cope with the lack and low accuracy of student science experiment datasets,this paper proposes a method of student science experiment detection based on realtime attention-scale sequence fusion of target detection converter RT-DETR-ASF.First,this paper constructs a dataset of student science experiments,including 417 videos,18308 video frames and 20331 annotations,which focus on fi ve behaviors:weighing,measuring height,dropping balls,measuring size,and recording.To improve the detection accuracy,an attention-scale sequence fusion module is introduced.To solve the boundary data problem,a behavioral boundary index is proposed for identifying boundary samples in the dataset.To solve the data imbalance problem,oversampling with video frame expansion is performed.A science experiment detection model is used to detect the dataset,and the experimental results show that the average accuracy of behavioral classifi cation detection reaches 71.1%,which proves the effectiveness of the method in this paper.The student science experiment dataset with RT-DETR-ASF provides a priori foundation for future student science experiment analysis,which is expected to promote the further development of this fi eld.

关 键 词：深度学习 学生科学实验 RT-DETR-ASF 数据不平衡 

分 类 号：G4[文化科学—教育学]