Research On Crowd Flowing Rule In Complex Scenes

With the large-scale activities increasing gradually, crowd behavioranalysis becomes more and more important. Also, with more data collected,mining useful information from multi-modal data becomes more popular,such as vector data, text data and video data. Hence, how to analyze crowdbehavior in complex scenes from gigantic, multi-modal data becomes aworthy topic. Crowd behavior analysis in complex scenes can enhance thelevel of urban services, such as detecting human density and easingcongestion; it can have applications in video surveillance system such asdetecting criminal behavior and unexpected incidents; it can have value incrowd management in large activities such as analyzing crowd flow andconfusion degree to prevent stampede, to decrease waiting time and toimprove customer satisfaction. Taking Shanghai World Expo as anexample, the Expo site recorded gigantic multi-modal data related tovarious information including tourist flow, ticketing, transportation,activity. The data stored as the form of vector, text, video reflects thereal-time information in the most straightforward way.Common methods for multi-modal based crowd behavior analysis arenumerical prediction method for tourists flow, feature extraction and eventdetection in computer vision, pattern analysis and graph clustering basedmobile techniques. Difficulties of crowd behavior analysis are multi-modaldata, complex scenes, multi-camera, real-time monitoring, high levelsemantic analysis, etc. To research on crowd behavior analysis in complexscenes while overcoming viewing angle limitations in video surveillancesystem and data collection privacy limitations in mobile techniques, wemainly focus on two problems: one is prediction for tourists flowdistribution; the other one is crowd behavior analysis for different scenes including simple scene like video surveillance system and complex scenelike large activities.First, this paper proposes prediction method for tourists flowdistribution. We predict tourists flow distribution by combining clusteringstrategy and generalized regression neural network. We choose historicaldata as training data, divide training samples as different classes and traincorresponding neural network in each class. Also, we take advantage oftransition probability matrix to predict tourists flow distribution. Throughminimizing mean square prediction error for historical data, we estimatetransition matrix among different geographical sensor locations betweenadjacent sampling intervals, use this matrix and past tourists flowdistribution to predict tourists flow distribution at next sampling interval.The methods have been testified on tourists flow data for zones andpavilions at Shanghai Expo site. The experimental results show that theproposed prediction method combining clustering strategy and generalizedregression neural network can increase accuracy and efficiency and savetime cost by decreasing samples. Also, the proposed transition probabilitymethod can better reflect crowd transition behavior with geographicalfactors and temporal factors.Secondly, we propose a crowd behavior analysis framework for videosurveillance system. First, we sample some frames from the video anddivide the video scene into regions to form different region location nodesof a sensor network. With the help of image and video processingtechniques, we detect human flow and count the number of people in eachlocation node. After establishing a camera sensor network, based on theconstraints of geographical factors, crowd moving speed and observationdata in camera sensor network, we take advantage of integer programmingto convert tracking problem into standard optimization problem, which canbe used to analyze crowd behavior. The proposed method has beentestified on video surveillance data at Shanghai Expo data. The resultsshow that, combining image processing and integer programming based multi-target trajectory identification can identify analyze crowd behaviorin video surveillance system.Thirdly, we propose an integer programming based crowd behavioranalysis method for large activities, which include some steps like datacollection and preprocessing, establishing topic model for sensor network,whether predefining group number and size for crowd, integer programingbased trajectory identification formulation, mixed integer programmingsolver and visiting pattern analysis. We testify our proposed method onzone visiting pattern at Shanghai Expo site. The experimental results showthat, our proposed method can automatically analyze crowd behavior incomplex scenes. We also can conclude different flowing patterns ondifferent times.