Research And Implementation Of University Classroom Raising Rate Detection System

In recent years,with the population of mobile phone and tablet computer among our daily life,many college student's attention has been attracted to these intelligent terminals when they are taking classes.This leads to bad lecture in the classroom.Improving the quality of classroom has become a problem of great concern to the university.Therefore,a system that can monitor the state of the students in the classroom will be very help to the teacher and university,which can be a big help to evaluate and improve the quality of teaching.For a large number of classrooms in the existing colleges and universities,we use highdefinition cameras to shoot classroom video and transmits the data to the back-end system for the intelligent analysis.The scheme proposed in this paper is mainly based on the face information of students in class.We use the face detection algorithm to get the position information of students,and using convolution neural network to recognize the head state of student.At the same time,based on the fusion of multi-frame face position information,we dynamically calculating the number of students in the classroom.We build HeadState data set for recognition of student's head state,the data is used to train and evaluate the accuracy of algorithm.Meanwhile,we build ClassHead data set for detecting student who raise up his head in classroom,the data is used to evaluate the accuracy of algorithm.Face detection algorithm implemented in our system achieve a recall of 92.9% on FDDB dataset.Head raising recognition model using convolutional neural network achieve accuracy of 92.3% on HeadState data set.Student counting algorithm based on multi-frame face location gives error under 10% compared to real number of student in the classroom.The system is tested on ClassHead data set.The performance gives an accuracy of 94.6% when the recall is 92.3%.To deal with multi-channel video processing,we optimize the memory usage of Caffe framework when deploying our models.Memory usage reduce 67% when testing on a Faster-RCNN network.This helps the system to handle more videos in parallel with a single GPU device,which reduce the cost of system deployment.