Estimation Of Crowd Density Based On Video Image

With the development of the times and the acceleration of urbanization,the population density is getting bigger and bigger.A large number of people in public places are crowded,which can easily produce dangerous accidents.So in these large flow of public places,real-time monitoring of crowd flow is very meaningful.Through the video image of the crowd flow real-time monitoring,security personnel in accordance with the situation of the flow of people in a timely manner to schedule human and material resources,to avoid the occurrence of some social security problems.Therefore,it is of great practical significance and research value to estimate the density of people.In this thesis,digital image processing and support vector regression techniques are used to estimate the population density.In order to reduce the error,the method of obtaining crowd density by the number of crowd is adopted.Flow counting and density estimation is divided into direct and indirect method,the direct method try to detect each pedestrian,and indirect method through the study of image features obtained the number and density of people.In this thesis,the indirect estimation method based on SURF(Speeded Up Robust Features)feature points,compared with the direct method,this method has a better effect in the crowd density.The contents of this thesis include: image preprocessing,foreground detection,feature extraction and support vector machine.Image preprocessing includes color image grayscale,image denoising and histogram equalization process,the purpose is to reduce the irrelevant information so that the image can be processed more easily.Secondly,feature extraction includes target detection and constructing feature vector.In the stage of target detection,a new moving target detection algorithm is proposed for pedestrian target extraction based on the combination of the Gauss model and the five frame difference method.In the process of constructing the feature vector,extraction of SURF feature points,clustering and convex hull are studied.The minimum spanning tree DBSCAN clustering algorithm is used to cluster the feature points,all the pedestrians are divided into different clusters,then extract features of each cluster and construct feature vectors.Based on the feature vector and its corresponding number of people,the support vector regression is used to train the mathematical model.The model can be used to predict the number of images,according to the number of people classify the stream density.Finally,the model is validated by the test data,and the results shows as we expected.