Research On Data-driven Crowd Motion Simulation Method

In recent years,crowd simulation have become popular research directions due to its wide range of application.In the field of public safety,crowd simulation can repeatedly simulate the evacuation behavior of crowd under emergency events,which becomes an effective tool for crowd emergency evacuation plan research.In the field of animation,the crowd simulation method can automatically generate a large number of agents in the virtual animation scene,which improve the production efficiency and visual realism of virtual crowd animation.Therefore,the research of crowd simulation method is of great significance.Crowd simulation should first ensure the collision avoidance function of crowd motion,and then ensure crowd motion realism.The crowd simulation model is required to restore the behavioral characteristics of real people as much as possible.Among the existing crowd simulation algorithms,the rule-based approach has achieved better results in collision avoidance through its tough rule binding,however,the agent’s motion pattern is single and not realistic enough.In contrast,the data-driven approach achieves advantages in the diversity of crowd motion modeling by reusing or learning the real data,but makes trade-offs in the model’s obstacle avoidance performance.Based on the above background,this dissertation proposes a data-driven,double-layered crowd simulation method that is mainly divided into two parts:the construction of the example database and the construction of the simulation model:The real trajectory information is constructed into the form of examples composed of state features and action features together,and the clustering process is performed according to the similarity between action features.By reorganizing the real data,agents can reuse the data directly during the simulation;The construction of the simulation model is divided into the motion trend prediction layer and the motion detail filtering layer.The first layer trains a siamese neural network to match the future motion trend based on the agent’s state information.The second layer selects the final example in the corresponding cluster as the agent’s motion strategy based on mathematical rules.The method ensuring the collision avoidance capability while generating crowd motion with a high degree of authenticity.The method in this dissertation is compared with the data-driven method Heter-Sim,the method Fusion Model and DeepORCA which jointly driven by data and rules as well as the rule-driven method ORCA through both quantitative experiments and qualitative experiments.In the quantitative experiments,the method in this dissertation guarantees a collision rate of 0 in all five scenarios of the ETH/UCY dataset,which reflects a great collision avoidance performance.In terms of motion realism,the method in this dissertation is closest to the real data in three metrics:average rate,average speed variation and average angle variation.In the average vertical deviation distance index,it performs only a little less effective than Fusion Model method.In the qualitative experiments,the trajectories of individual agent simulated by different methods are compared,and the trajectories simulated by our method are closer to real data in three cases:agent walking alone,agents walking side by side and agents walking in opposite directions.Then we compare the overall trajectories of agents in different scenarios with other methods.We plot the trajectory shapes and velocity distributions of multiple agents.The experimental results show that the trajectory shapes and velocity distributions of our method are closer to the real agents.Finally,the visual fidelity and generalizability of crowd motion is further verified by simulating various motion behaviors of agents in 3D artificial scenes.