A Data-Driven Crowd Simulation Model Based on Clustering and Classification

In this paper, we propose a data-driven crowd behavior model that is constructed by extracting examples from human motion data describing how humans make decisions. We cluster the examples before the simulation to find similar patterns of behavior. During the simulation, at each simulation time step, we first classify the input state perceived by an agent in the simulation into one example cluster using an artificial neural network classifier. We then combine similar examples of that cluster to produce an output, a velocity vector indicating the position of the agent in the next time step. Such a two step matching process enables the selection of the most similar example accurately and efficiently. To verify our approach, we have developed an initial prototype in which we build our model using motion data generated by a RVO2 simulator, attempting to reproduce the behavior of the RVO2 model. By comparing the position of the same agent simulated by the RVO2 mode land our model respectively at the same time steps, we show that our model has the ability to reproduce the behavior of the RVO2 model accurately. As future work, we will use real human motion data as model input, so that our model may perform human-like motion behavior.