Crowd Analysis On Video Surveillance

With the development of economy and the acceleration of urbanization,mass activities are increasing.The technology of crowd analysis are increasing rapidly and the monitoring and management of crowd scenes have also attracted the focus of intelligent video surveillance system.The algorithms of crowd scenes motion analysis has also made great progress.At present,crowd monitoring is mostly based on human monitoring.Manual monitoring is timeconsuming and laborious.With the increase of working hours,visual fatigue caused by manual monitoring can lead to make mistakes.The information provided by surveillance video can not be fully utilized.Based on computer computing and continuous operation capability,we can analyze the crowd's video scenes for a long time and make full use of the data information in the video to complete the analysis of all aspects of crowd scenes.Crowd analysis can be used in many aspects,such as crowd control,safety forecast,management and so on.Through analyzing the crowd density,clustering information and motion pattern,we can make full use of information provided by video.The existing crowd scenes and crowd behaviors are becoming more and more complex.Whether the algorithm can be applied to different scenarios,including structured and unstructured,low,medium and high density crowd scenes,has gradually become the focus of crowd algorithm research.First of all,the existing crowd algorithms lack the comprehensiveness of the scene description.Most of the algorithms extract the low-level features of the crowd scenes,and describe the high-level crowd motion characteristics,but lack of high-level description.Most crowd analysis algorithms concern the movement of individuals,which only consider the discrete characteristics of individuals in the crowd and ignore the interaction between particles and continuity.Finally how can the intelligent system can be applied to the real scene in the crowd effectively in order to ensure the accuracy and real-time of the algorithm.In this paper,the existing crowd modeling algorithms are divided into various types according to the different research methods,and the advantages and disadvantages are analyzed,and the research direction of this paper is made clear.In this paper,inspired by the fluid mechanics,presents a contains low-level feature and middle semantic features,and characteristics from the local and global motion modeling,space domain and time domain analysis of the behavioral characteristics of people.This paper proposes three kinds of descriptors: advection descriptor,diffusion descriptor and pressure descriptor.Three kinds of descriptors for the description of the behavior of different people with different abilities,advection descriptors and diffusion descriptors mainly describe the change rule of particle velocity and acceleration in space in the neighborhood.Pressure discriptor describes the interaction between the local information and particle density.Meanwhile,based on the analysis and discussion of the three descriptors and relying on three kinds of descriptors to analyze the behavior of the crowd,a crowd behavior recognition algorithm based on Alexnet classification network is proposed,and the effectiveness of the algorithm is proved by comparing the effect of different data sets.Secondly,this paper studies the existing crowd trajectory prediction algorithms.First,we introduce the Social LSTM algorithm based on the social pooling layer in the long and short term memory network to predict the crowd trajectory.The original algorithm only considers the interactions between people in the neighborhood linearly but ignores the difference of interaction between individuals of different pedestrian densities.The higher density the pedestrian locates,the larger interactions between pedestrians have.The proposed method also combined with the crowd motion consistency in the neighborhood.A new pressure collective pooling layer is proposed to construct the new hidden layer of network information,and it is the new input hidden state of the next module to forcast the new pedestrians' position.Finally,in this paper we propose a system that integrates multiple crowd properties,including stationary and dynamic features,local and global characteristics,and historic statistics analysis in a unied framework.This system has accomplished three crowd analysis including crowd density estimation,crowd segmentation and crowd saliency detection.The crowd density estimation extracts stationary feature,sparse spatial-temporal local binary pattern,and provides local density distribution as well as global density level.The crowd segmentation describes the whole crowd grouping with dynamic features including temporal motion grouping and distribution grouping.Meanwhile,this analysis also presents local moving direction of each grouping.The saliency detection detects salient regions,where abnormal behaviors will happen.Abnormal activities can be detected with the density level results and the semantic segmentation,both locally and globally.At the same time,in order to record the results of crowd analysis in a long time,the system stores the historical data in database in order to make statistics for the trend of crowd density,the number of different moving groups and the percentage of groupings in different directions in a certain period.Meanwhile,the system can play,stop,pause video and connect with remote video recorders and upper devices.