Research On Statistical Methods Of Classroom People Based On Deep Learning

With the development of science and technology and the continuous improvement of the social and economic level,the use of video surveillance equipment has increased dramatically.A large number of monitoring systems have been deployed in public places such as stations,shopping malls,scenic spots and roads9 and a large number of siarveillance videos have been generated.In recent years,the number of people based on surveillance video has played an increasingly important role in crowd behavior analysis,resource optimization configuration,and modern security.People have more and more demand for population statistics,so how to achieve high accuracy in surveillance video.The statistics of the number of people have become the research hotspots and difficulties in this research field.This paper is based on solving the accurate number of people in classrooms,laboratories and other indoor scenes.At the same time,I will develop a population statistics system based on surveillance images to provide necessary statistics for students1 attendance and self-study classroom occupancy.Since the surveillance cameras are generally installed on the upper side of the classroom,the students sit in the seat in front and rear,the students and students will block each other,and because the surveillance camera has to take a full range of shots of the entire classroom,it is impossible to the students make close-ups,which causes the students to be very small in the surveillance video,unable to fully obtain the students body contour information and clear facial features,and can not adopt the common face detection method,so the student's head information becomes Breakthrough in the classroom statistics.Inspired by the face detection algorithm,in the case that the scales of the objects to be detected are small and overlap each other,I will use a population counting method based on human head detection.The head detection uses the current R-FCN detection algorithm with excellent performance indicators.Because the students are very small in the surveillance video,and the students are close to each other,the human head detector trained directly by the R-FCN object detection algorithm has a poor detection effect.Therefore,I will make a series of improvements to the traditional R-FCN object detection algorithm,so that it has the ability to detect small objects.The comparison experiments show that the improved R-FCN object detection algorithm has better head detection than the traditional R-FCN object detection algorithm.The improved R-FCN object detection algorithm finally achieved 86.53%in the classroom demographic test set.