| The academic performance of students is the core issue that the school and parents are concerned about.Studies have shown that there is a close relationship between the academic performance of students and their attention in class,and students who can maintain concentration for a long time tend to have a higher academic performance.Therefore,it is very necessary to improve the level of students’ attention in class.For this reason,it is necessary to detect the attention of students first.The existing attention detection methods include questionnaire survey method,instrument detection method and computer vision algorithm.Although the questionnaire survey method is convenient for statistical analysis,it is usually affected by subjective factors and cannot guarantee the quality and credibility of the experimental results.Although the instrument detection method can avoid subjective influence to a certain extent,these instruments are generally expensive,and an instrument can only detect the attention of one student at a time,which is difficult to meet the needs of attention detection for dozens of students in the class at the same time.Computer vision algorithms generally use third-person image or video data.However,due to the large number of students in the classroom,it is difficult for the algorithm to extract small objects and occluded objects,so it cannot accurately detect the attention of students.In order to achieve the goal of obtaining students’ classroom attention data in batches and accurately,this thesis proposes a new method: students’ classroom attention detection based on first-person video.Compared with the third-person video,the first-person video has the advantages of one-to-one correspondence between the video and the student,and the video content is consistent with the gaze behavior,so it can make the students’ attention detection more accurate.The attention detection model proposed in this thesis includes three modules:gaze point estimation,gaze object recognition and attention level analysis.The gaze point estimation module uses a deep learning algorithm that combines saliency detection and attention shift.In order to improve the accuracy of the algorithm,this thesis tried three attention mechanisms,namely spatial attention,channel attention and hybrid attention.The specific method is to embed the attention block behind the convolution block of the Convolutional Neural Network(CNN).The gaze object recognition module compares the effects of the Single Shot Multi Box Detector(SSD)and the Fast Semantic Segmentation Network(Fast-SCNN).The attention level analysis module comprehensively analyzes and evaluates students’ attention through eye movement measurement index calculation and process evaluation.In order to verify the effect of the attention mechanism on the gaze point estimation algorithm,this thesis conducted experiments on the public first-person video dataset GTEA Gaze Plus,and the experimental results show that all three attention mechanisms have an improvement effect on the gaze point estimation algorithm.Among them,the hybrid attention has the greatest effect,followed by the channel attention,and the spatial attention has a relatively small effect.Since the public first-person video dataset of students in class has not been obtained at present,this thesis also recorded a dataset of First-Person Video of Students in Class(FPVSC)and conducted experiments on this dataset.The experimental results show that the Fast-SCNN algorithm is more accurate than the SSD algorithm in the gaze object recognition module;the method proposed in this thesis can effectively evaluate the classroom attention of individual students and all students. |
