Research Of Crowds Simulation Based On Neural Network

Virtual crowd motion simulation has been an important research field of the major virtual reality group. Results of simulation techniques play a pivotal role in the crowd management(parade system, riots situation), public space design(architecture, urban facility planning), disaster prevention and other aspects. However, a single domestic research environment leads to a low understanding of crowd behavior and experience. People eager to be able to study the crowd behaviors, and simulate the disaster situation. Thus potential problems could be found in advance to protect life safety.In this thesis, we analyzed the key technology in crowd behavior simulation in depth. Focusing on different techniques, we proposed three improvement programs mainly in the modeling based on individual characteristics, the global collision avoidance, and crowd control plan. The main contents are:1. Studying the individual characteristics in crowd environment, we improved the modeling method based on individual emotional characteristics. Based on PEN model, we mapped the individual emotional characteristics to the agent, reflecting the diversity of characteristics of crowd behavior. Meanwhile, combining with GAS model, the characteristics of agents had the initiative in emotion, to achieve a dynamic individual goals for emotion changing.2. Studying Collision between relative agents in large-scale crowd simulation, we proposed a new global adaptive collision avoidance method. The method used artificial neural network learning crowd feature information, leading to generate anticipation space to avoid collision in higher probability, thus avoided the overmuch interactive computation in traditional process. On the other hand, we reduced the consumption of resources and the time complexity with optimization in global collision avoidance.3. Studying the control planning method of large-scale crowd, we used a path planning algorithm based navigation force field. And according to crowd density, it could change dynamically travel speed, closing to the real situation. Due to large-scale high complexity caused by traversing, there is still inadequacy in query efficiency among agents. In this thesis, the storage structure optimization was proposed to ensure efficient and reliable traversal.Based on the improved method and research content of this thesis, we achieved a crowd simulation program of large-scale urban environment to simulate realistically crowd behavior. Then we showed the presentation and analysis. Subsequently, we summarized the study results, and looked forward to the direction of work in future.