The Large-scale Crowd Analysis In Video Surveillance

With the rapid growth of the economy and worldwide urbanization,large-scale crowd activities have become increasingly frequent. Fatalaccidents in different fields, such as entertainment, sports, religion, etc.,take place from time to time. To prevent the happening of such unfortunateaccidents, the monitoring and management of the crowd has become ahotspot issue in the field of intelligent video surveillance.Recently, with the development of computer vision and artificialintelligence, the study of crowd behavior analysis has made great progressand has been widely used in intelligent video surveillance systems. Density,velocity and direction, as the most important characteristics of thelarge-scale crowd, compose the basis of crowd behavior perception andanalysis.Based on the analysis of existing crowd analysis algorithms, the mainwork of this thesis can be divided into the following two parts. On the onehand, this thesis has proposed a new feature extraction and analysismethod which can be applied to the large-scale crowd density estimationand distribution description. On the other hand, based on the crowd densityinformation and social force model, this thesis has put forward a newmethod to detect the abnormal crowd behavior.Considering the drawback of conventional crowd density estimationalgorithms, this paper proposed a novel density estimation method usingthe sparse spatial-temporal local binary pattern (SST-LBP) descriptorwhich performs well on all density levels. Firstly, use spatial-temporal Hessian matrix to detect sparse locations with notable variance in bothspatial domain and temporal domain. Then extract the local feature whichis represented by the result of the spectrum analysis of the SST-LBP code.Thirdly, the crowd feature is obtained by employing the histogram toextract the statistical property of the local feature. Afterwards，SupportVector Machine (SVM) is adopted to build the relationship between thecrowd feature and the density level of the crowd. Finally, the local featureis transformed to the color space to express the density distribution. Apartfrom above four parts, perspective correction is also introduced to thesystem to reduce the effect of projective deformity. Theoretical analysisand experimental study both prove that this method performs well on alldensity levels and can effectively describe the local crowd densitydistribution. Therefore, the system is suitable and practical for large-scalecrowd surveillance application.For crowd anomaly detection, the detection based on individual targetand the detection based on crowd feature are firstly introduced. Apart fromthat, the crowd behavior analysis using Social Force Model is alsoelaborated in detail. Inspired by fluid dynamics, the unequal information oflocal density distribution has been introduced to the Social Force Model,thus the model has been improved to fit the reality more and the noise ofinvalid background force has been reduced largely. Taking into accountthat the occurrence of abnormal crowd behavior is often accompanied bythe change of the local density, this thesis has integrated the local densityvariation and the improved social interaction force, proposing a novelPressure Model to detect the abnormal crowd behavior. The larger the localdensity changes and the larger the social interaction is, the more unstablethe crowd is and the greater the pressure of the crowd is. Finally, theperformance of the proposed method is evaluated on the publicly availabledatasets from UMN and PETS. The experimental results show that theproposed method can achieve a higher accuracy than that of the previous methods on detecting abnormal crowd behavior.