Research On Classroom Behaviors Detection Of Primary School Students Based On Faster R-CNN

With the development of education,the importance of classroom teaching is increasingly emphasized in the field of primary school education.people pay more and more attention to the reactions and their changes of students in the classroom.The new curriculum reform in basic education puts forward higher request for teaching evaluation.To evaluate teaching according to students' classroom behaviors.It is an effective and commonly used way.But the traditional artificial observation has the problem of time-consuming and inefficient.In recent years,object detection technique based on big data and deep convolutional neural network has developed rapidly and is applied to more and more fields.With the popularity of information classrooms equipped with cameras,a large number of classroom teaching videos are produced.This makes it possible to apply object detection method based on image and video data in the field of education.Certainly,a few researchers have adapted object detection technique to analyze education and teaching videos.Although certain achievements have been made,there is still great potentials for making further progress.Aiming at the classroom teaching and learning videos of primary school,this thesis uses the detection technology based on the deep convolutional neural network.The classroom behaviors of primary school students are detected and realize the association of student identities in the video in this thesis.The contributions of this thesis are as follows:(1)The Pupils' Classroom Behaviors Dataset(PCB Dataset)is constructed.75 classroom videos of primary school classroom teaching from different provinces,different schools,different grades,and different courses are collected.The common classroom learning behaviors of primary school students are divided into seven categories: head-up,head-down,head-tilt,hand-raising,standing,turning,and other behaviors.46,888 frames of images and 1286,169 instances are marked in fatal.(2)A method for detecting students' behaviors in class is proposed in this paper.Based on the Faster R-CNN detection framework,the feature pyramid and the prime sample attention mechanism are used to solve the problems of various object scales and the imbalance of data categories respectively.The image-based classroom behaviors detection of primary school students was realized,and 70.9% Average Precision was obtained in the PCBD image test set.(3)A student identity association technique based on classroom video is proposed in this paper.Inspired by two-stage detector structure,based on the improved Faster R-CNN detector,will sign a frame of video detection results as input of the current frame.change the object detection process.It achieves a more accurate student identity Association.The test results and identity correlation can be used to evaluate the classroom learning performance of each individual student or all students overall classroom.The results can also be used as one component of teaching evaluation.