Research On Crowd Flow Segmentation Method In Complex Scenes And Its Application In Crowd Motion Description

Public safety has been widely concerned by people.At the same time,video surveillance facilities have been installed in a large number of public places,generating a large number of surveillance videos every day.It is difficult to deal with such massive data in a timely manner by relying solely on manual surveillance.Therefore,it is urgent to carry out the research of intelligent monitoring technology,so as to reduce the workload of manual monitoring,and finally completely replace manual monitoring.The environment of public scene is extremely complex,the resolution of surveillance video is often very low,and the movement state of the crowd is complex and changeable,which makes the task of automatic analysis and understanding of crowd behavior extremely challenging.Although crowds in public places generally show disorganized movement,they are actually locally ordered.If we can get the consistent area of each movement,it will be easy to analyze the crowd movement.Therefore,this paper researches the method of crowd flow segmentation,and puts forward a description model of crowd movement state based on crowd flow segmentation method.Most of the existing methods use clustering algorithm to segment the crowd motion,and these clustering algorithms need to specify the number of clustering clusters or clustering threshold.However,it is difficult to select parameters that can adapt to different population scenarios,which greatly limits the application of these methods.In this paper,deep learning is introduced into crowd flow segmentation for the first time,and a deep motion transformation network is proposed to transform crowd flow segmentation into detection of crowd areas with different motion states,which avoids the problem of parameter selection based on clustering method.The model firstly transforms the motion information into a visual triple representation,then uses the pyramid network to extract the features,then uses the area detection network to detect the crowd area with the same motion,and finally obtains the pixel-level crowd flow segmentation results by removing the background network.A large number of crowd segmentation tests have been carried out on the real scene surveillance video.The experimental results show that the proposed deep motion transformation network can avoid the problem of parameter selection and achieve better segmentation effect than the previous methods.The deep motion transformation network mentioned is a supervised model.In order to avoid the heavy marking task,this paper further proposes an unsupervised crowd motion segmentation model based on motion contrast enhancement,which can avoid the problem of parameter selection.The crowd flow segmentation method first enhances the contrast of the motion according to the distribution characteristics of the motion and noise in the sports field,then uses the adaptive threshold segmentation algorithm to obtain the contour of the basically consistent regions of different motion states,and finally obtains the basically consistent regions of each motion state through the marker watershed algorithm,so as to realize the crowd flow segmentation.In a large number of complex crowd scene surveillance videos,the crowd segmentation experiments verify the effectiveness of the method and the accuracy of segmentation.In this paper,an energy model is proposed to describe the crowd movement on the basis of obtaining the consistent region of each movement state.By describing the energy that affects the stability of crowd movement state,the proposed energy model deduces the state change of the whole crowd movement,so as to realize the abnormal early warning.In a large number of real scene experiments,the prediction results of the proposed model are basically consistent with the actual trend of crowd movement.Finally,we design and implement a crowd movement state analysis system,which includes friendly user interface,selection of crowd flow segmentation method,generation of broken line graph of crowd movement state energy change according to segmentation results,so as to realize the prediction of crowd movement state in the future and view the local crowd movement state analysis result file.