Intelligent Crowd Simulation Based On Video Data-driven

Crowd simulation is an important research topic of virtual reality technology.It reveals the potential rules of crowd behavior influenced by external environment and interpersonal interaction through simulating crowd behavior in the virtual world constructed by computer,and provides the technologies of analysis and visualization of crowd behavior to researchers and users.In recent years,crowd simulation technology has been widely applied in many fields such as public safety,military training,major activity,and computer animation.However,due to the limitations of data acquisition and information extraction methods,the existing methods of crowd simulation mainly focus on model-driven mode,which has certain insufficiency in terms of simulation effects and fidelity.Therefore,this dissertation studies the intelligent crowd simulation based on video data-driven.The object detection and tracking algorithms of computer vision are used to automatically extract pedestrian movement information from crowd videos.The crowd simulations of three different data-driven modes,including 3D reproduction,example learning,and model optimization,are implemented by combing3D visualization and machine learning.This dissertation aims to integrate the technologies of computer vision,artificial intelligence and crowd simulation,and promote the development of theory and application of data-driven mode in the field of crowd simulation.The main contributions of this dissertation are summarized as follows.(1)To address the demand of crowd simulation technology for real motion data,a method based on computer vision is proposed to automatically extract crowd movement information from video data.Firstly,the object detection approach using gaussian mixture model background subtraction is presented to realize the automatic detection of pedestrian targets in video without prior knowledge.Then an improved object tracking approach based on particle filter and data association is proposed to accurately extract the trajectories of multiple pedestrians.Finally,the coordinate transformation is used to transform the trajectory from 2-dimensional coordinates of video images into3-dimensional coordinates of real world,and the 3D visualization of crowd movements in corresponding videos is performed on the Digital Earth Science Platform.The experimental results show that the proposed method can automatically extract accurate pedestrian trajectories from the crowd video,and obtain various relevant parameters of crowd motion.In addition,the crowd movements captured in the video can be 3D reproduced by this method.(2)For most model-driven methods can not effectively simulate sparse crowd behavior with higher motion freedom,a method based on example learning is proposed to simulate sparse crowd behavior.The trajectory example is defined to describe a local spatio-temporal scenario of pedestrian present in the original video,and a database of trajectory examples is constructed by hierarchical clustering algorithm.A motion pattern classifier based on BP neural network is established to realize fast and effective query of the trajectory example.The motion behavior learning method using the k-nearest neighbor algorithm and local weighted linear regression algorithm,is proposed to predict the motion behavior of the agent at the next time step.The experimental results show that the proposed method can realistically simulate the crowd movement,and improve the query matching speed of the example.When the number of trajectory examples is 6000,the query time is less than 0.06 second/frame.(3)To solve the modeling problem of pedestrian cognitive behavior with uncertainty and imprecision in dense crowd,a method based on genetic fuzzy system is proposed to simulate dense crowd behavior.A hybrid pedestrian motion model based on agent-based model and social force model is established to realize hierarchical modeling of pedestrian perception,decision and action.To simulate the pedestrian's field of view,a radial-based method of spatial representation of the environment is proposed.The fuzzy logic system is introduced to simulate imprecise perception and decision-making ability of the agent for pedestrian.The parameter learning of pedestrian motion model is achieved by using the genetic algorithm to optimize the parameters of the fuzzy membership function.The experimental results show that the proposed method can effectively improve the accuracy of crowd movement simulation.The mean square error of test experiments is 0.0213 m~2,and the accuracy is improved by 9.75%after optimization.In addition,the simulation results are more consistent with real observations.(4)The crowd flow in typical scenarios is studied by real data analysis and simulation experiment.The characteristic factors of crowd flow are analyzed through collecting crowd videos of four common public places such as subway stations,shopping malls,parks and pedestrian streets.According to crowd density and flow direction,the state of crowd flow is divided into various types.Furthermore,the common phenomena of crowd flow are discussed.The impact of environmental factors on crowd movement is studied through the simulation of crowd flow in many typical scenarios.The results can provide scientific basis for layout planning and building design,and support crowd safety management and emergency plan formulation.