Research on the RT-DETR-ASF Based Detection of Students'Experimental Behavior in Scientific Inquiry
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 real-time attention-scale sequence fusion of target detection converter RT-DETR-ASF.First,this paper constructs a dataset of student science experiments,including 417 videos,18 308 video frames and 20 331 annotations,which focus on five 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 classification 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 field.
Deep learningStudent science experimentsRT-DETR-ASFData imbalance
万方数据
维普
