Research On Crowd Counting And Density Estimationalgorithm Based On Convolutional Neural Network

Due to the rising frequency of large-scale social gatherings and the security threats associated with this environment,the study of crowded scenes has gained popularity in recent years.Pedestrian detection for video surveillance,human behavior analysis and understanding crowd density estimation are just a few of the applications in this field.Despite ongoing advances in computer vision and machine learning techniques,a number of complex issues remain a challenge in nowadays,particularly in the presence of a single view algorithm that cannot fully reflect the global information of crowd movement or in high-density crowd situations,which are the most dangerous and necessitate more research to understand the mechanisms that lead to crushes and stampedes,which can result in the loss of human lives.In this research we propose a Crowd counting and Density Estimation Algorithm Based on Convolution Neural Networks which is multiple deep neural network branches and fusion mechanisms that integrates multi-perspective views in order to overcome the fact that a single perspective cannot fully reflect the global information of crowd movement,to clarify our comments and give more details on our improvement concerning the new model we are proposing,below is a table describing the difference between the existing model and the new model we are proposing.The main works of this work are as follows:1.Most of the existing crowd counting and density estimation algorithm method uses the late fusion(LF)model to merges the heat maps task,and the na(?)ve fusion(NF)model to merges camera view feature maps,but does not dig deep into the weak feature extraction ability of the model and low prediction accuracy.In view of this problem and to improve the prediction and counting ability of the model,we propose a Multi-View Multi-Task fusion model(MVMT).Which is introducing semantic segmentation subtasks of crowd counting and density estimation algorithm based on convolutional neural network,the dependencies of semantic segmentation and feature extraction are mined,and the ability of the network to parse the underlying spatial position features and high-level semantic features of the scene is enhanced.Secondly,a multiple deep neural network branches and fusion mechanisms is designed and Finally,a crowd density map estimation method that integrates Multi-View Multi-Task fusion model(MVMT)is constructed to achieve crowd counting and prediction of movement trends,which can be used to Manage the flow of crowds in large squares,and identify potential public safety accidents that can be caused by crowded crowds as soon as possible.2.Intra-view scale differences are a significant problem in single-view counting because persons appear in the image at varying sizes due to perspective effects.The severity of scale variation is exacerbated by using several views,in addition to intra-view scale variation;multi-view photos contain inter-view scale variations,in which the same individual appears at different scales across multiple viewpoints.Finally,To solve this challenge,we extract feature maps at several sizes and then scale pick the projected features so that they are at consistent scales across all views,Our proposed Multi-View Multi-Task fusion model(MVMT)has four steps as main work:(?)collecting Multi-View feature maps for each camera view by applying a fully-convolutional network(FCN),convolution layers to an image pyramid;(?)up sampling all feature maps to the largest size and then picking scales for each pixel in each camera view based on scene geometry;(?)Using the projection module,project the scale-consistent feature maps to the ground-plane representation;(?)Using the fusion module,fuse the projected features and forecast a scene-level density map.This will improve the prediction and counting performance of the crowd and density estimation algorithm model.3.We have used 3 methods proposed models of fusion framework to test and verify the model proposed in this paper which are the late fusion model that merges the heat maps;the na(?)ve fusion model that merges camera view feature maps and the Multi-View Multi-Task(MVMT)fusion model that favors aligned features on the same ground plane point to have consistent multi scales views.We tested our fusion models on 4 Datasets which is Shanghai Tech,PETS2009,Duke,MTMC,and one newly collected multi-view count data containing a city crowded street.We have designed a multiple deep neural network branches and fusion mechanisms,crowd density map estimation method that integrates multi-view to achieve crowd counting and prediction of movement trends,the experimental results show that the training of our model has relatively stabilized prediction and counting ability.