Intelligent Classroom Visual Inspection System Based On Deep Learning

In the Internet age,"Internet+education" has gradually become a popular phenomenon.Meanwhile,the use of synchronized classrooms for online distance education is gaining increasing momentum among more and more schools and educational institutions.The application of computer vision-based deep learning technologies renders synchronized classrooms more interactive and intelligent,exerting a fairly positive impact on the form and structure of an Internet-based educational curriculum.At present,recording classroom lectures via traditional manual video-shooting methods often puts restrictions on large-scale lecture capture.Therefore,it is of both necessity and urgency to capture lectures in a way that allows for the automatic rotation of the camera in real-time classroom setting,so that the activities of the teacher and students can be detected and tracked with ease.This issue is further addressed in a three-part research endeavor devised as follows.The first part probes into the algorithmic principles that underlie the system of intelligent classroom detection and tracking,especially the algorithms of detecting&tracking various moving objects and of human presence.It is hoped that,through comparing the effects of different methods,a certain theoretical framework can be formulated to inform the realization and optimization of a deep learning-based visual detection and tracking system in intelligent classroom settings.The second part addresses the algorithms behind the detection and tracking of moving human targets,including teachers in motion and students in standing respectively.With regard to the detection and tracking of teachers in motion,the Faster R-CNN and Gaussian mixture models are used for dynamic human objects and the images thus obtained are then scaled accordingly.In respect of the detection and tracking of students in standing,the Faster RCNN network is designed to operate on the students'standing behaviors.Besides,the action classification algorithm and the pixel position regression method are then applied respectively.Considering the real-time nature of video recording in intelligent classrooms,the faster Single-stage SSD detection method is adopted,and the acceleration method based on model distillation is proposed to improve the real-time calculation efficiency of the system.Where necessary,FocalLoss is added to improve detection and tracking efficacy.The third part presents the actualization of a visual detection and tracking system for intelligent classrooms.On the basis of a comprehensive analysis of the performances of different algorithm models,coupled with certain logical analysis of both panoramic and close-up cameras,we attempt a solution to a deep learning-based visual detection and tracking system in classroom settings.This paper devises a visual detection and tracking system that can be applied to real-time classroom video recording,and it also designs a network that can be trained end-to-end.The experimental results verify the effectiveness of the system.