Technical Research And System Design Of Intelligent People Counting Based On Deep Learning

People or crowd counting based on surveillance video has high research significance and application value.It is the basis of important tasks,such as crowd behavior analysis,optimizing resource allocation,modern security,business information collection and intelligent management.At the same time,it is also an essential functional part of modern intelligent surveillance video.In recent year,with the continuous improvement and rapid development of digital image processing technology,artificial intelligence,computer vision and deep learning,research on people counting based on surveillance video has been greatly promoted.However,there still exist some problems can not be well solved,such as low accuracy of people counting,time-consuming of high-definition image and difficult to generate high quality image of crowd density estimation.This thesis bases on solving the problem of accurate people counting of classrooms,laboratories and other indoor scenes,as well as the problem of people counting and crowd density estimation of squares,stations and other outdoor scenes.Moreover,developing an intelligent people counting service system based on surveillance video,which provides the necessary information of people counting for some tasks,such as student attendance intelligence counting,security management and early warning.The main work is as follows:(1)Aiming at the problem that the accuracy of people counting method based on object detection decreases greatly when the scale of detected object changes greatly,this thesis proposes a people counting method based on the combination algorithm of adaptive overlapping segmentation and deep neural network.The idea of this method comes from attention machanism and makes full use of the information of the scales and numbers of human head object in overlapping segmentation block.The experiment results show that the adaptive overlapping segmentation algorithm can combine with existing object detection model based on neural network.What's more,compared with the direct use of object detection model based on neural network,the combination method of adaptive overlapping segmentation algorithm and deep neural network can greatly improve the accuracy of the people counting.(2)Aiming at the existing people counting method for sigle static images in crowded scene has the problems of low people counting accuracy and difficult to generate high-quality image of crowd density estimation.This thesis simplifies the crowded density estimation problem to the image transform image problem,which is sigle static images in crowded scene to image of crowd density estimation.And proposes a method of crowd density estimation MBcGAN based on multi-scale filter and the framework of conditional adversarial generative nueral networks.Compared with MCNN and MSCNN crowded density estimation methods,the experiment results show that MBcGAN method not only improves the people counting accuracy,but also generates a higher quality image of crowd density estimation.(3)Designed and developed an intelligent people counting service system based on surveillance video,which not only can provide high accuracy people counting functional service for indoor,scenes such as classrooms and laboratories,but also provide crowd density estimation functional service for outdoor scenes,such as square and station.