Study On Virtual Crowd Behaviors Based On Multi-agent And Cellular Automata Model

Currently, many problems exist in virtual crowd behavior simulation, that behavior models are not lifelike enough, a single model can’t simulate crowd behavior actually and the adaptability of crowd behavior simulation frame is poor. However, crowd simulation technology has wide application prospect in many fields, such as security science, physical training simulation, computer animation and virtual reality.Aiming at this situation, this paper develops a deep study on some key technology of virtual crowd behavior simulation.The intensive study launches from several aspects, including virtual human’s motion generation, virtual crowd’s basic behaviors generation and integration, the transform mechanism between behaviors, physiological and psychological factor influence on crowd behavior and so on. Thus, corresponding behavior models and rules which control virtual human’s behavior are proposed. The main contents of the dissertation can be summarized as follows:(1) the research status at home and abroad of the virtual crowd behavior modeling technology and application are summarized. The status, merits and drawbacks of separate utilizing cellular automata and multi-agent technology to study crowd behavior simulation are introduced importantly. A new method which combines cellular automata and multi-agent technology to study crowd behavior simulation is put up.In this method, the part of cellulars in cellular space are regarded as agents, in addition, the cellular and its state are packaged up which are deemed to a type and state attribute of the agent. Through this method, more authentic crowd behavior models are established expectably.(2) As to the poor variability of the traditional virtual human motion generation method, the principal component analysis (PCA) is utilized numerously in motion edition and synthesis, because of its ability to decompose high dimensional data by linear mapping, by which the interpolation and extrapolation to motion data property can be manipulated in different spaces. Nevertheless, linear PCA is not able to detect the whole structure of given data, instead, KernelPCA is more suitable to extract complex original data with nonlinear relation. In comparison to PCA, KernelPCA can extract the more useful characteristics of classification. Furthermore, body motion is a much complex nonlinear system with high dimension, then, we can identify virtual human’s motion more effectively using KernelPCA. Therefore, this dissertation proposes a method based on KernelPCA to generate virtual human’s motion. The first step is to work out the optimal kernel function formula by some optimization algorithms and then produce a linear separable characteristic space with high dimension. The second step is to reduce the dimension of the characteristic space. Finally, we carry on linear mapping to eigenvector’s coefficients and motion attribute parameter. After that, virtual human’s motions will be generated according to user’s demand. Consequently, users can obtain virtual human’s motions with different motion attributes and types via a little motion data.(3) Aiming at the low fidelity and computational efficiency of current virtual crowd behavior, this dissertationconsiders analyses sufficiently the feasibility and superiority of combining Multi-agent and Cellular Automata Technology to simulate crowd behavior. In our models, a crowd is divide into herds, groups and individuals, among that, a herd is composed of more than one group, and a group is composed of many individuals. This kind of structure is capable of enlarging crowd scale conveniently and describing the crowd with multi-task-group.In the dissertation, virtual environment is mapped into cellular space, whose grid size is reduced to improve crowd behavior’s continuity. Based on this foundation, agent’s perception model and decision model are designed particularly, meanwhile, the complex individual and crowd behaviors are generated though design of basic motion, motion integration and motion selection mechanism according to perception information and self-state. Every designed basic behavior and complex crowd behavior in virtual environment are verfied by experiment.(4)In the end, we simulate crowd evacuation behavior in a certain sudden event such as fire disaster by the established crowd behavior model. We design the crowd evacuation simulation system, which has high fidelity and can generate vivid human being and crowd behavior animation and verifies the feasibility and exactness of our virtual human’s motion generation algorithm and crowd model. Research on evacuation behavior rules and characteristics by experiment in the system, scientific foundations for planning and designing buildings, facilities and transportation etc are supplied and opinions for safety management are acquired effectively.