Agent-based Crowd Simulation Calibration With Pedestrian Clustering

With increasing diversity of social activities,it is becoming more and more important to control large-scale crowds.Crowd simulation arises to help analyze and predict crowd dynamics by modeling crowd behaviors.In the field of crowd simulation,agent-based modeling approach has gained main popularity since its superiority in both fidelity and flexibility.Pedestrians in a crowd are modeled as intelligent agents that are capable of sensing environments and making decisions by themselves,which can produce diverse behaviors similar to real-life pedestrians.However,a key issue in agent-based crowd simulation is to properly calibrate and verify the crowd model,so that the simulation results match as close as possible with real crowd movements.Because the problem of calibrating agent-based crowd models is usually high-dimensional and complex,most of the existing calibration methods use unified parameters for all agents in the model.Such methods ignore the inherent diversity in reallife crowds.As the scale of crowd and the complexity of scene grow,pedestrian differences expose a greater impact on crowd dynamics,which greatly reduces the calibration progress of traditional methods.The key aspect of this study is to enable crowd diversity in model calibration.We propose an agent-based crowd simulation calibration method based on pedestrian clustering.Our method assigns the same parameters to members of the same group,so that the total parameter dimensionality is reduced.Meanwhile,members from different groups can have different parameters,so that group-level crowd diversity is enabled.For different pedestrian groups produced by our pedestrian clustering method,a coevolutionary multi-group parameter optimization method is designed to automatically find the best group parameter sets.The idea behind this method is two-folded.First,groups can be seen as relatively independent because members from different groups exhibit quite different motion patterns.According to that observation,we apply a divide-and-conquer strategy to the multi-group parameter optimization problem to reduce complexity.Second,members from different groups interact with each other when steering.To maintain such group linkage,a collaborative evaluation mechanism is adopted.In addition,considering that the computational cost of evaluating the crowd simulation model is too high,we further integrate a resource allocation strategy into the coevolutionary optimization process.By dynamically allocating the chance of evolution among groups according to their contribution to the magnitude of simulation refinement,our method can achieve efficient use of computing resources.Finally,in order to apply the calibration parameters for new simulations,we design a parameter matching process that assigns calibrated parameters to new pedestrians according to their similarity of motion patterns.Our proposed method facilitates more credible and realistic agent-based crowd simulation.